{ "id": "pt-interviews", "spine": { "id": "pt-startup-interview-spine", "versao": "1.0.0", "o_que_e": "The four fixed questions, and the closing one. This file is the SOURCE for the quick-reference block at the top of /newsroom/interviews.html, so that the block and the question cards cannot drift apart. Gate 42 fails the build if any question here is missing from the cards.", "porque_sao_fixas": "Asked identically and in this order to everybody. Improvising the spine feels better in the moment and destroys the comparison — and the comparison is the only thing this newsroom can do that a blog cannot.", "perguntas": [ { "n": 1, "en": "In one sentence, what do you make, and who pays for it?", "pt": "Numa frase: o que é que fazem, e quem é que paga?", "porque": "The second half is the one people skip. If they answer only the first, ask again." }, { "n": 2, "en": "What can you do today that you couldn't do eighteen months ago?", "pt": "O que é que conseguem fazer hoje que não conseguiam há dezoito meses?", "porque": "The AI question without saying «AI». Capability, not positioning." }, { "n": 3, "en": "What's the hard part right now — and is it the technology, the money, or the people?", "pt": "Qual é a parte difícil agora — é a tecnologia, é o dinheiro, ou são as pessoas?", "porque": "Make them pick one. The distribution across twenty startups is itself a finding." }, { "n": 4, "en": "What's one thing everyone says about AI in Portugal that doesn't match what you see?", "pt": "O que é que toda a gente diz sobre IA em Portugal e não bate certo com o que você vê?", "porque": "The disagreement question. It lets them disagree with a crowd rather than a person, which is why they answer it honestly." } ], "fecho": { "en": "Who else here should I be talking to, and why them?", "pt": "Com quem mais é que eu devia falar aqui, e porquê essa pessoa?", "porque": "This is how twenty interviews become a map. It is also the one easiest to skip at four in the afternoon." }, "antes_de_sair": { "en": "Read back and get a yes on each: name, role, company, and their one-line answer to question 1. No email, no phone — those never enter a file here.", "pt": "Leia em voz alta e arranque um sim a cada um: nome, cargo, empresa, e a resposta de uma linha à pergunta 1. Sem email nem telefone — não entram em ficheiro nenhum." } }, "version": "2.0.0", "what_it_is": "Two interviews and one mechanism. One points outward — an assistant interviews somebody, and what comes back is material for an article. The other points inward — an agent has ChatGPT interview the editor of record, and what comes back is direction.", "the_rule": "An interview is a primary source about the person speaking and nothing else. What they assert about the world is a claim to verify, not a fact for having been said aloud — and that holds when the person speaking is the editor.", "derived_from": "briefs/pack/09__interviews/ and briefs/pack/10__editor-interview/", "count": 11, "packs": [ { "folder": "09__interviews", "id": "interviews", "title": "Interviewing somebody, to write", "why": "A busy person will not write two pages about their work. They will talk about it for twenty minutes, at an hour that suits them, by voice or by typing.", "readme": "# 09 — Interviews\n\nThe other half of what this newsroom publishes. `08__research-briefs/` sends an assistant after\n**documents**; this sends it into a **conversation**, and what comes back is a different kind of\nmaterial: a person's own words about their own work.\n\n## Why by voice\n\nBecause it is easier for the person answering and better for the person reading. A busy person will\nnot write two pages about their work, but will talk for twenty minutes about it — and talks better\nthan they write, with examples they would never put in writing. ChatGPT in voice mode runs the\ninterview at whatever hour suits them, with nothing to schedule with anybody.\n\nAnyone who would rather type, types: the same prompt works in both modes.\n\n## The pieces\n\n| File | What it is | Who uses it |\n|---|---|---|\n| `01__interview-prompt.pt.md` | The interviewer, in Portuguese. Always the same. | The interviewee |\n| `02__the-interview-prompt.en.md` | The same, in English | The interviewee |\n| `03__topic-block.md` | What changes: who, why, about what | The agent writes it, the person pastes it |\n| `04__agent-prompt.md` | How a pack is assembled for one person | The editor hands it to an agent |\n\nThe two interview prompts are **in the language they are pasted in** — a Portuguese speaker pastes\nPortuguese. Everything else here is English, like the rest of this repository's operational surface.\n\n## The path, from invitation to article\n\n1. The editor picks the person and hands `04__agent-prompt.md` to an agent session.\n2. The agent finds what the newsroom already holds frozen about them, writes the topic block, and\n assembles a pack — an sgit vault or a `.zip`.\n3. The person receives it, pastes the two blocks into ChatGPT, and talks. They end with a report.\n4. **They review and correct the report.** Nothing leaves without that.\n5. They send it back. The newsroom treats their statements as statements — things they said, not\n facts about the world — and goes looking for sources for the ones that are verifiable.\n6. The editor of record decides whether to publish, and what. Publication is in European Portuguese;\n if the conversation was in English, the translation is done here and shown to them before it goes\n out.\n\n## The rules that do not bend\n\n- **An interview is a primary source about the person speaking, and nothing else.** If somebody\n gives a number about the market, that is a claim to verify, not a fact for having been said aloud.\n- **Three fields per person**, as everywhere else on this site: listed role, listed organisation,\n the page that lists it. An interview makes no exception.\n- **No contact detail** enters any file, neither in the pack nor in what comes back.\n- **Approval is the person's and publication is the editor's.** Two decisions, and neither\n substitutes for the other.\n- **Removal is unconditional and carries no reason.** Anyone interviewed may ask for the interview\n to come down, after publication, without explaining why — and the data-protection notice says how.\n", "pieces": [ { "file": "01__interview-prompt.pt.md", "path": "briefs/pack/09__interviews/01__interview-prompt.pt.md", "title": "The interview prompt", "language": "português", "why": "Pasted into ChatGPT as the first message. Always the same, whoever the person is.", "for_whom": "the interviewee", "lines": 79, "sha256": "f82d2b8687e4244824abf743e1fdb48246484c2636fb8babcffabb48f45101cf", "text": "# O prompt da entrevista — português\n\n**Cole isto no ChatGPT como primeira mensagem, e depois carregue no ícone do microfone.** Funciona\ntambém em modo escrito, se preferir escrever. Só precisa de o colar uma vez; o bloco do tema vem a\nseguir, na segunda mensagem.\n\n---\n\nÉs o entrevistador de **pt.newsroom.sgit.ai**, uma redação portuguesa que mapeia o ecossistema\nportuguês de inteligência artificial. Vais conduzir uma entrevista por voz comigo.\n\n## Como esta entrevista funciona\n\n- **Uma pergunta de cada vez.** Nunca mais do que uma. Espera pela resposta antes da seguinte.\n- **Segue o que eu disser**, não a tua lista. Se eu disser algo interessante e não previsto, vai por\n aí. A lista de temas é um ponto de partida, não um guião a cumprir.\n- **Pergunta por exemplos concretos.** «Pode dar um exemplo?» e «Como é que isso se passou na\n prática?» valem mais do que qualquer pergunta de opinião.\n- **Não me elogies e não concordes por educação.** Se eu disser algo vago, pede-me para concretizar.\n Se eu me contradisser, aponta-o com cuidado e deixa-me esclarecer.\n- **Não inventes nada.** Não completes as minhas frases, não presumas o que eu quis dizer, e não\n atribuas ao meu trabalho ou à minha organização coisas que eu não disse.\n- Fala em **português europeu**. Se eu mudar para inglês, acompanha-me e continua em inglês.\n- Duração alvo: **20 a 30 minutos**. Avisa-me quando estivermos perto do fim.\n\n## O que não perguntas\n\n- Dados de contacto — email, telefone, morada. Se eu os disser, **não os registes no relatório**.\n- Dados de saúde, convicções políticas ou religiosas, orientação sexual, filiação sindical, origem\n étnica. Se surgirem por acaso na conversa, ficam fora do relatório.\n- Opiniões minhas sobre pessoas nomeadas. Sobre organizações, sim; sobre pessoas, não.\n\n## Quando eu disser que terminámos\n\nProduz um **relatório da entrevista** com esta estrutura, e nada mais:\n\n```\n# Entrevista — , , \ndata: duração aproximada: \nlíngua da entrevista: \nmodo: \n\n## Em três linhas\n\n\n## Os temas que cobrimos\n\n\n## Citações verbatim\n<5 a 12 frases minhas, transcritas EXATAMENTE como as disse, cada uma com o tema a que pertence.\n Se a transcrição por voz falhou numa palavra, escreve [inaudível] em vez de adivinhar.>\n\n## Afirmações que a redação vai ter de verificar\n\n\n## O que eu pedi para ficar de fora\n\n\n## Perguntas que ficaram por responder\n\n```\n\nDepois do relatório, diz-me exatamente isto e mais nada:\n\n> Reveja o relatório acima. Corrija o que estiver errado, corte o que não quiser publicar, e\n> devolva-o à redação. **Nada disto é publicado sem a sua aprovação**, e a redação publica em\n> português europeu — se falámos em inglês, a tradução é feita por ela e mostrada antes de sair.\n\n## O que eu preciso de saber antes de começares\n\nAntes da primeira pergunta, diz-me em duas frases o que é esta redação e como vai usar o que eu\ndisser, e pergunta-me se posso começar. Se eu disser que não, para.\n\n---\n\n**A seguir vem o bloco do tema: quem sou eu, porque é que me convidaram, e sobre o que querem\nfalar. Espera por ele antes de começares a entrevista.**\n" }, { "file": "02__the-interview-prompt.en.md", "path": "briefs/pack/09__interviews/02__the-interview-prompt.en.md", "title": "The interview prompt", "language": "english", "why": "The same interviewer, in English. The interview may be in either language; what gets published is always European Portuguese.", "for_whom": "the interviewee", "lines": 79, "sha256": "d51a826513c922540fddd0a69c378b0888d47fd6f46880238272b017eed73526", "text": "# The interview prompt — English\n\n**Paste this into ChatGPT as your first message, then press the microphone.** It works in typed\nmode too, if you would rather write. Paste it once; the topic block comes next, as a second message.\n\n---\n\nYou are the interviewer for **pt.newsroom.sgit.ai**, a Portuguese newsroom mapping Portugal's AI\necosystem. You are going to conduct a voice interview with me.\n\n## How this interview works\n\n- **One question at a time.** Never more than one. Wait for the answer before the next.\n- **Follow what I say**, not your list. If I say something interesting that you did not plan for,\n go there. The topic list is a starting point, not a script to get through.\n- **Ask for concrete examples.** \"Can you give an example?\" and \"How did that actually play out?\"\n are worth more than any question about opinions.\n- **Do not compliment me and do not agree out of politeness.** If I am vague, ask me to be\n specific. If I contradict myself, point it out gently and let me clear it up.\n- **Invent nothing.** Do not finish my sentences, do not assume what I meant, and do not attribute\n to my work or my organisation anything I did not say.\n- Speak **English**. If I switch to Portuguese, follow me and continue in Portuguese.\n- Target length: **20 to 30 minutes**. Tell me when we are near the end.\n\n## What you do not ask about\n\n- Contact details — email, phone, address. If I say them, **do not record them in the report**.\n- Health, political or religious beliefs, sexual orientation, trade union membership, ethnic\n origin. If they come up in passing, they stay out of the report.\n- My opinions about named people. About organisations, yes; about people, no.\n\n## When I say we are done\n\nProduce an **interview report** with this structure, and nothing else:\n\n```\n# Interview — , , \ndate: approximate length: \ninterview language: \nmode: \n\n## In three lines\n\n\n## The topics we covered\n\n\n## Verbatim quotes\n<5 to 12 sentences of mine, transcribed EXACTLY as I said them, each tagged with its topic. If the\n voice transcription failed on a word, write [inaudible] rather than guessing.>\n\n## Claims the newsroom will have to verify\n\n\n## What I asked to be left out\n\n\n## Questions left unanswered\n\n```\n\nAfter the report, say exactly this and nothing else:\n\n> Review the report above. Correct anything wrong, cut anything you do not want published, and send\n> it back to the newsroom. **None of this is published without your approval**, and the newsroom\n> publishes in European Portuguese — if we spoke in English, it does the translation and shows it\n> to you before it goes out.\n\n## What I need to know before you start\n\nBefore the first question, tell me in two sentences what this newsroom is and how it will use what\nI say, and ask whether I am ready to begin. If I say no, stop.\n\n---\n\n**The topic block comes next: who I am, why you invited me, and what you want to talk about. Wait\nfor it before starting the interview.**\n" }, { "file": "03__topic-block.md", "path": "briefs/pack/09__interviews/03__topic-block.md", "title": "The topic block", "language": "português", "why": "The second message, and the only part that changes from interview to interview. An interview is worth whatever this block is worth.", "for_whom": "the agent", "lines": 89, "sha256": "ab9981679a77336b63770a9b6a2503e757986e14f48e71b5647f1d8b813321f3", "text": "# The topic block — what changes from interview to interview\n\nThe interview prompt is always the same. **This is the part that changes**, and it is the second\nmessage the person pastes into ChatGPT. An interview is worth whatever this block is worth: a\ngeneric block produces a generic conversation.\n\nIt is filled in from what the newsroom already holds frozen about the person and their\norganisation — never from what is assumed about them.\n\nThe template below is **in Portuguese**, because it is pasted into a Portuguese conversation. The\ninstructions around it are English, like the rest of this repository's operational surface.\n\n---\n\n\n## O modelo, por preencher\n\n```\n## Quem sou eu\nnome: \ncargo listado: \norganização listada: \npágina que o diz: \n\n## Porque é que esta redação me convidou\n\n\n## O que a redação já tem, e não precisa de perguntar\n\n\n## Sobre o que queremos falar\n1. \n2. \n3. \n \n\n## Perguntas de arranque\n\n\n## O que não perguntar nesta entrevista\n\n\n## A língua\n — e a nota de que a publicação é sempre em português europeu.\n```\n\n---\n\n## Um exemplo preenchido\n\n```\n## Quem sou eu\nnome: \ncargo listado: \norganização listada: \npágina que o diz: https://pt.newsroom.sgit.ai/entidades/pessoa//\n\n## Porque é que esta redação me convidou\nEstá listado como orador do Startup Summit Lisbon 2026, numa sessão sobre IA. Esta redação tem a\npágina de oradores congelada de 8 e de 13 de setembro e a lista mudou entre as duas.\n\n## O que a redação já tem, e não precisa de perguntar\n- O cargo e a organização, como a página do evento os lista (congelada, com hash).\n- A sessão em que está inscrito e o título dessa sessão, verbatim.\nNada mais. Esta redação não tem biografia nenhuma sobre si, e não a vai construir a partir daqui.\n\n## Sobre o que queremos falar\n1. O que está a construir, em concreto — porque o mapa deste site tem organizações e quase nenhuma\n descrição do que elas fazem por dentro.\n2. O que é difícil fazer em Portugal e fácil noutro sítio, ou o contrário.\n3. O que mudou nos últimos doze meses no que faz.\n\n## Perguntas de arranque\nO que é que a sua equipa fez este ano que não conseguiria ter feito há dois?\nQual foi a decisão técnica mais difícil, e o que é que a tornou difícil?\nO que é que toda a gente diz sobre IA em Portugal e não bate certo com o que vê?\nOnde é que foi buscar as pessoas, e o que é que foi mais difícil de encontrar?\nO que gostaria que existisse em Portugal e não existe?\n\n## O que não perguntar nesta entrevista\nNada sobre ronda de financiamento em curso.\n\n## A língua\nportuguês — a publicação é em português europeu.\n```\n" }, { "file": "04__agent-prompt.md", "path": "briefs/pack/09__interviews/04__agent-prompt.md", "title": "The agent prompt", "language": "english", "why": "What the editor hands an agent session to prepare a pack. The agent assembles; it does not interview.", "for_whom": "the editor", "lines": 56, "sha256": "463565e2c75e99bde277ce23942512c1f33882970044474d984734f7cfc96434", "text": "# The agent prompt — whoever assembles the pack\n\nThis is what the editor hands an agent session (Claude Code, in this newsroom) to prepare an\ninterview. The agent interviews nobody: **it assembles the pack the person will receive.**\n\n---\n\nYou are preparing an interview pack for **pt.newsroom.sgit.ai**. The person is: ``.\n\n## What you do, in this order\n\n1. **Find what this newsroom already holds about them.** `dados/pessoas.json`, their page under\n `entidades/pessoa//` if it exists, and the frozen copies their name appears in —\n `dados/entidades.json` records which files the name was found in and how many times. If the\n newsroom holds nothing, say so: an interview with somebody nothing is known about is an interview\n with generic questions.\n2. **Read the rule about people before writing a line.** A person, in this newsroom, carries three\n fields only: the listed role, the listed organisation, and the page that lists it. No biography,\n no characterisation, no contact details — not in the pack, not in the topic block, not in what\n comes back. See `CLAUDE.md`, rule 4, and the data-protection notice.\n3. **Write the topic block** to the template in `03__topic-block.md`, from what you found and not\n from what you assume. Three to five topics; five to eight opening questions. Every statement\n about the person has to come from a page this newsroom has frozen.\n4. **Assemble the pack.** Five files and nothing else:\n\n ```\n interview-/\n READ-ME.md what this is and what to do, in three paragraphs\n interview-prompt.md the prompt in the chosen language\n topic-block.md what you wrote in step 3\n what-happens-next.md approval, translation, publication, and how to ask for removal\n return/ an empty folder, where the person puts the report\n ```\n\n5. **Deliver it.** An sgit vault is the better route — the person opens it in a browser with nothing\n installed and returns the report into it. A `.zip` will do when it is not. Either way, what\n leaves is only what you wrote: no copies of other people's pages.\n\n## What you do NOT do\n\n- **Do not write a biography.** Not in the READ-ME, not in the topic block, not as \"context\".\n- **Do not put adjectives on the person.** \"Founded X in 2019\" is a record; \"is a leading voice in\n the sector\" is a verdict, and this newsroom publishes the record and never the verdict.\n- **Do not put anyone's contact details in any file.** The invitation goes out by whatever channel\n the editor already uses with that person; the pack does not record which.\n- **Do not promise publication.** The pack says the interview is reviewed by the person and then by\n the editor of record, and that it may not be published.\n- **Do not translate the interview** until the person has approved the report in the language they\n spoke.\n\n## What you hand back to the editor\n\nThe path to the pack, the topic block for him to read before it goes out, and a list of what you\ncould **not** find about the person in the frozen sources — because that is what decides whether the\nquestions are good or generic.\n" } ] }, { "folder": "11__startup-interviews", "id": "startup-interviews", "title": "Interviewing a founder yourself, at an event", "why": "Five minutes, standing, in a hall with a PA system. Four fixed questions asked identically to everybody — which is what turns twenty conversations into something you can count — and a short pool of follow-ups picked by what they just said.", "readme": "# 11 — Startup interviews, in the field\n\nThe third interview pack, and the only one where **a human does the interviewing**.\n\n| Pack | Who interviews | Who is interviewed | What comes back |\n|---|---|---|---|\n| `09__interviews/` | ChatGPT, by voice | somebody outside, at their desk | material for an article |\n| `10__editor-interview/` | ChatGPT, by voice | the editor of record | direction for an agent |\n| **`11__startup-interviews/`** | **you, with a recorder** | **a founder, at an event** | **a dataset, and articles** |\n\n## Why it is different\n\nPack 09 assumes half an hour and a quiet room. A summit gives you five minutes, standing, in a hall\nwith a PA system, with somebody who has a meeting at the top of the hour. Different constraints,\ndifferent instrument: **four fixed questions and a short pool of follow-ups**, not a conversation\nthat finds its own way.\n\n## The four fixed questions, and why they are fixed\n\n1. In one sentence, what do you make, and who pays for it?\n2. What can you do today that you couldn't do eighteen months ago?\n3. What's the hard part — technology, money, or people?\n4. What does everyone say about AI in Portugal that doesn't match what you see?\n\nAsked identically, in the same order, to everybody. **That is what turns twenty conversations into\nsomething you can count**, and counting is the only thing this newsroom can do that a blog cannot.\nImprovising the spine feels better in the moment and destroys the comparison.\n\nThe closing question — *\"who else here should I be talking to, and why them?\"* — draws the map.\n\n## The pieces\n\n| File | What it is |\n|---|---|\n| `01__the-questions.en.md` | The question card, English. What you actually hold at the event |\n| `02__as-perguntas.pt.md` | The same, Portuguese, for founders who prefer it |\n| `03__from-recording-to-article.md` | Transcript → claims → article, and the language split |\n\n## The rules, unchanged\n\nA recorded founder is a **primary source about their own company and nothing else**. Their numbers\nabout the market are leads to verify, not facts. Three fields per person — listed role, listed\norganisation, where it came from. No contact details in any file. The interviewee approves, and the\neditor of record publishes; two decisions, neither replacing the other.\n\n**And one rule that is new**, because this is the first content born in English and published in\nPortuguese: the transcript is never translated, a translated quotation carries its original or stops\nbeing a quotation, and the person approves the Portuguese and not only the English.\n", "pieces": [ { "file": "01__the-questions.en.md", "path": "briefs/pack/11__startup-interviews/01__the-questions.en.md", "title": "The question card", "language": "english", "why": "What you actually hold at the event. The four fixed questions, the follow-up pool, the closing question, and the thirty seconds before you walk away.", "for_whom": "you", "lines": 120, "sha256": "c5de7da25a048c094b362777868b5c3bcb4e36dd8ed04e9d2830e0c1ddb71fff", "text": "# The questions — English\n\n**For an interview you conduct yourself, standing up, at an event, with a recorder running.**\nFive to ten minutes. Not the ChatGPT interview in pack 09 — that one is for somebody at their own\ndesk with half an hour. This one is for a noisy hall and a founder who has somewhere to be.\n\n---\n\n## Before you press record (20 seconds)\n\n> \"I write for pt.newsroom.sgit.ai — we're mapping who's actually building AI in Portugal.\n> Five minutes, I'm recording so I get your words right, and nothing goes out without you seeing\n> it first. Alright?\"\n\nWait for the yes. **Then press record, and say the date and their name into the microphone** — it\nis the only timestamp you will trust later.\n\n---\n\n## The spine — ask these four, every time, in this order\n\nThese four are what make twenty interviews into a dataset instead of twenty anecdotes. Same\nquestions, same order, every startup. **Do not improvise these.**\n\n**1. \"In one sentence, what do you make, and who pays for it?\"**\n\nThe second half is the one that matters and the one people skip. If they answer only the first\nhalf, ask again: *\"and who pays?\"* A startup that cannot answer who pays in one sentence has told\nyou something.\n\n**2. \"What can you do today that you couldn't do eighteen months ago?\"**\n\nThis is the AI question without saying \"AI\". It gets you the actual capability change rather than\na positioning statement, and it is comparable across every company you ask.\n\n**3. \"What's the hard part right now — and is it the technology, the money, or the people?\"**\n\nMake them pick one. The distribution across twenty startups is a finding in itself.\n\n**4. \"What's one thing everyone says about AI in Portugal that doesn't match what you see?\"**\n\nThe disagreement question. It is where the story usually is, and founders answer it honestly\nbecause it lets them disagree with a crowd rather than a person.\n\n---\n\n## Then pick two or three, by what they said\n\nDo not work through these. Pick the ones their answers opened.\n\n**If they build the models or the infrastructure**\n- What runs where — your own hardware, a cloud, someone's API?\n- What broke first when you scaled?\n- What would you need that does not exist in Portugal today?\n\n**If they apply AI to somebody else's industry**\n- Who is the buyer, and what were they doing before you?\n- How long from first conversation to money, and what makes it slow?\n- What does the regulation look like from where you sit?\n\n**If they are hiring, or trying to**\n- What role took longest to fill, and did you find them here?\n- How many of your team are in Portugal?\n\n**If they have raised money**\n- Who led it, and were they Portuguese? *(If they name a round, get the date and the lead — those\n are checkable, and everything else they say about it is not.)*\n- What did the money buy that you could not buy before?\n\n**If they are at the summit for a reason**\n- What are you actually here for — customers, hires, investors, or the room?\n- Who did you want to meet and could not?\n\n---\n\n## The closing question — always ask it\n\n**\"Who else here should I be talking to, and why them?\"**\n\nThis is how twenty interviews become a map. It also tells you what this founder thinks is\ninteresting, which is itself worth knowing.\n\n---\n\n## Before you walk away (30 seconds)\n\nRead back, out loud, and get a yes on each:\n\n- their **name**, spelled if it is not obvious\n- their **role**, in their words\n- their **company**, spelled\n- the **one-line** answer to question 1, in their words\n\nThat is the whole of what this newsroom records about a person: the listed role, the listed\norganisation, and where it came from. **Do not ask for an email or a phone number** — if they offer\na card, that is fine, but it does not go into any file here. Tell them you will send the piece\nthrough whoever introduced you, or through the event.\n\n---\n\n## What you must not do\n\n- **Do not promise publication.** \"I'll send it to you before anything goes out\" is the promise.\n- **Do not let them approve someone else's words.** If they characterise another company or\n person, that is theirs, and it does not go in a piece about them.\n- **Do not treat what they say as fact.** Every number, date and claim about the market is a\n **claim to verify** — the same rule as any research delivery. What they say is a primary source\n about *themselves* and nothing else.\n- **Do not ask about anything special-category** — health, politics, religion, union membership,\n ethnic origin. If it comes up, it stays off the record and out of the file.\n\n---\n\n## One sentence into the recorder at the end\n\n> \"Recorded at the Startup Summit, , with of , who agreed to be recorded and\n> to review before publication.\"\n\nYou will thank yourself in a week.\n" }, { "file": "02__as-perguntas.pt.md", "path": "briefs/pack/11__startup-interviews/02__as-perguntas.pt.md", "title": "O cartão das perguntas", "language": "português", "why": "The same card, for founders who would rather speak Portuguese.", "for_whom": "you", "lines": 117, "sha256": "1053490da62673aae03d80f8e540be769e8b96d2744f9f559a673cb1bc0cee28", "text": "# As perguntas — português\n\n**Para uma entrevista que é você a conduzir, de pé, num evento, com o gravador a trabalhar.**\nCinco a dez minutos. Não é a entrevista do bloco 09 — essa é para alguém sentado à secretária com\nmeia hora. Esta é para um pavilhão com barulho e um fundador que tem de estar noutro sítio.\n\n---\n\n## Antes de carregar em gravar (20 segundos)\n\n> «Escrevo para o pt.newsroom.sgit.ai — estamos a mapear quem está mesmo a construir IA em\n> Portugal. Cinco minutos, vou gravar para não lhe torcer as palavras, e não sai nada sem você ver\n> primeiro. Pode ser?»\n\nEspere pelo sim. **Depois grave, e diga a data e o nome dele para o microfone** — é a única marca\ntemporal em que vai poder confiar daqui a uma semana.\n\n---\n\n## A espinha — estas quatro, sempre, por esta ordem\n\nSão estas quatro que transformam vinte entrevistas num conjunto de dados em vez de vinte\nanedotas. As mesmas perguntas, pela mesma ordem, a todas as empresas. **Estas não se improvisam.**\n\n**1. «Numa frase: o que é que fazem, e quem é que paga?»**\n\nA segunda metade é a que importa e a que toda a gente salta. Se responderem só à primeira, volte a\nperguntar: *«e quem paga?»* Uma empresa que não consegue dizer numa frase quem paga já lhe disse\nalguma coisa.\n\n**2. «O que é que conseguem fazer hoje que não conseguiam há dezoito meses?»**\n\nÉ a pergunta sobre IA sem dizer «IA». Dá-lhe a mudança real de capacidade em vez de uma frase de\nposicionamento, e é comparável entre todas as empresas a que a fizer.\n\n**3. «Qual é a parte difícil agora — é a tecnologia, é o dinheiro, ou são as pessoas?»**\n\nObrigue a escolher uma. A distribuição das respostas por vinte empresas é, por si só, um achado.\n\n**4. «O que é que toda a gente diz sobre IA em Portugal e não bate certo com o que você vê?»**\n\nA pergunta do desacordo. É onde costuma estar a história, e os fundadores respondem-lhe com\nhonestidade porque lhes permite discordar de uma multidão e não de uma pessoa.\n\n---\n\n## Depois escolha duas ou três, conforme o que ouviu\n\nNão percorra esta lista. Escolha as que as respostas abriram.\n\n**Se constroem os modelos ou a infraestrutura**\n- O que corre onde — máquinas vossas, uma nuvem, a API de alguém?\n- O que é que partiu primeiro quando cresceram?\n- O que é que precisavam e não existe em Portugal hoje?\n\n**Se aplicam IA à indústria de outra pessoa**\n- Quem é o comprador, e o que fazia ele antes de vocês?\n- Quanto tempo da primeira conversa até ao dinheiro, e o que o torna lento?\n- Como é que a regulação se vê daí de dentro?\n\n**Se estão a contratar, ou a tentar**\n- Que função demorou mais a preencher, e encontrou-a cá?\n- Quantas pessoas da equipa estão em Portugal?\n\n**Se levantaram dinheiro**\n- Quem liderou a ronda, e era português? *(Se disserem uma ronda, apanhe a data e quem liderou —\n isso confere-se; o resto do que disserem sobre ela não.)*\n- O que é que o dinheiro comprou que não podiam comprar antes?\n\n**Se estão no summit por uma razão**\n- Ao que vieram mesmo — clientes, contratações, investidores, ou a sala?\n- Quem é que queria encontrar e não encontrou?\n\n---\n\n## A pergunta de fecho — faça-a sempre\n\n**«Com quem mais é que eu devia falar aqui, e porquê essa pessoa?»**\n\nÉ assim que vinte entrevistas se tornam um mapa. E diz-lhe também o que este fundador acha\ninteressante, o que já vale por si.\n\n---\n\n## Antes de se ir embora (30 segundos)\n\nLeia em voz alta e arranque um sim a cada um:\n\n- o **nome**, soletrado se não for óbvio\n- o **cargo**, nas palavras dele\n- a **empresa**, soletrada\n- a resposta de **uma linha** à pergunta 1, nas palavras dele\n\nÉ este o total do que esta redação regista sobre uma pessoa: o cargo listado, a organização\nlistada, e de onde veio. **Não peça email nem telefone** — se lhe derem um cartão, tudo bem, mas\nnão entra em ficheiro nenhum aqui. Diga que envia a peça por quem os apresentou, ou pelo evento.\n\n---\n\n## O que não pode fazer\n\n- **Não prometa publicação.** A promessa é «envio-lhe antes de sair alguma coisa».\n- **Não deixe ninguém aprovar as palavras de outro.** Se caracterizarem outra empresa ou pessoa,\n isso é deles, e não entra numa peça sobre eles.\n- **Não trate o que dizem como facto.** Cada número, data e afirmação sobre o mercado é uma\n **afirmação para verificar** — a mesma regra de qualquer entrega de investigação. O que dizem é\n fonte primária sobre *eles próprios* e mais nada.\n- **Não pergunte nada de categoria especial** — saúde, política, religião, sindicato, origem\n étnica. Se surgir, fica fora da gravação e fora do ficheiro.\n\n---\n\n## Uma frase para o gravador, no fim\n\n> «Gravado no Startup Summit, , com da , que concordou em ser gravado e em\n> rever antes de publicação.»\n" }, { "file": "03__from-recording-to-article.md", "path": "briefs/pack/11__startup-interviews/03__from-recording-to-article.md", "title": "From a recording to an article", "language": "english", "why": "Transcript, the split between what they said about themselves and what they said about the world, and the language rule for content born in English and published in Portuguese.", "for_whom": "the agent", "lines": 81, "sha256": "04d46ddd84c18a4eaf5e8aad9a02dd408020e672ec0c18816837c2dba8ab238a", "text": "# From a recording to an article\n\nWhat you hold when you walk away: an audio file, and four facts you said out loud into it. This is\nwhat happens next, and where the language changes.\n\n## 1 · Transcribe it verbatim, the same day\n\nWhisper, ChatGPT, whatever — but **verbatim**, in the language it was spoken, with `[inaudible]`\nwhere the hall won. A cleaned-up transcript is already an edit, and you cannot un-edit it later.\n\nSave it at `redacao/entrevistas/-.md` with a header carrying the four spoken\nfacts: name, role, company, and the one-line answer to question 1. **That file is the source.**\nEverything downstream cites it the way an article cites a frozen page.\n\n## 2 · Split it into two piles, and never merge them\n\nThis is the whole discipline of this newsroom applied to a conversation:\n\n- **What they said about themselves** — what they build, who pays, what is hard, what they are\n here for. This is a **primary source**, and it is quotable as *\"X said\"*, attributed, today.\n- **What they said about the world** — market sizes, who is biggest, what the government did, what\n another company is doing. These are **leads to verify**, exactly like a line in a ChatGPT\n delivery. They get an issue and a source hunt, and **none of them is publishable as fact because\n a founder said it into a microphone.**\n\nA number said confidently at a summit is still a number with no source behind it.\n\n## 3 · The language split, which is new here\n\n**You record in English. The site publishes in European Portuguese.** So:\n\n- The **transcript stays in the language it was spoken.** It is evidence, and evidence is never\n translated — the same rule that keeps a frozen page's bytes untouched.\n- The **article is written in European Portuguese**, by this newsroom.\n- A **quotation** is the hard case. A quote translated is no longer verbatim, and this publication's\n whole claim is that what is between quotation marks is what was said. So: a translated quote\n carries the original beside it, or it is not a quote — it is a summary in our words, unquoted.\n The article says which language the interview was in, in its own provenance block.\n- **The person approves the Portuguese** before it goes out, not just the English. They are being\n quoted in a language they may not have chosen; that is their decision to check, not ours to make.\n\nThis is the first content on this site that is born in one language and published in another, and\nit is worth saying plainly: **the translation is an editorial act, and it is recorded as one.**\n\n## 4 · Where it lands\n\nIn **`empresas`** — the brief's own section for companies, which is what a startup is. There is no\nninth section for this: `CLAUDE.md` fixes the eight, and a series is a shape inside a section, not a\nsection of its own. The series is what makes them findable together.\n\nEach interview becomes an article in the ordinary way — a dated folder, prose with source marks, a\nverification record, a provenance file — with one addition: `origem: entrevista` and the transcript\nnamed as what it rests on.\n\n## 5 · What the four spine questions are for\n\nTwenty interviews with the same four questions is a **dataset**, and it supports pieces no single\ninterview can:\n\n- what twenty founders can do now that they could not eighteen months ago\n- whether the hard part is technology, money or people — and how that splits\n- where they disagree with the consensus, and whether they disagree with the same thing\n- who they said to talk to, and who got named more than once\n\nThat last one draws a map of the room from the inside. **It is the strongest thing in this pack and\nit comes from the closing question**, which is the one easiest to skip when you are tired and it is\nfour in the afternoon.\n\n## 6 · Who does what\n\n| Step | Who |\n|---|---|\n| Record, and get the four facts on tape | You, in the hall |\n| Transcribe verbatim, same day | You, or an agent you hand the audio to |\n| Split into \"about themselves\" / \"about the world\" | An agent, and it shows its working |\n| Open issues for every world-claim | An agent |\n| Write the Portuguese article | Redação |\n| Check every quote against the transcript | Verificação |\n| Send the Portuguese to the interviewee | You |\n| `estado: publicado` | You, and only you |\n" } ] }, { "folder": "10__editor-interview", "id": "editor-interview", "title": "Interviewing the editor, to know what to do", "why": "An agent knows what is in front of it and does not know what the editor thinks matters. Twenty minutes gets that out, and gets it out better than any form.", "readme": "# 10 — Interviewing the editor\n\n`09__interviews/` sends an assistant to interview **somebody outside**, and what comes back is raw\nmaterial for an article. This sends an agent to interview the **editor of record**, and what comes\nback is something else: **direction**. Which stories are worth writing, who to talk to, what to do\nnext, what to stop doing.\n\n## The problem this solves\n\nAn agent in this newsroom knows what is in front of it — its issues, its mail, what is blocked — and\ndoes not know **what the editor thinks matters**. That information exists, and today it only comes\nout when he remembers to write it down. A twenty-minute conversation gets all of it, and gets it\nbetter than any form: the editor says \"look over there\", \"what about that?\", \"I have an idea\", and\nnone of that fits in a field.\n\n## The difference that changes everything: the agent studies before it asks\n\nA generic interview with the editor is worth nothing. The agent behind this conversation **starts by\nreading its own state** — the `ROLE.md` and `MANDATE.md` of its identity, its open and blocked\nissues, its unread mail, its recent run records — and brings **its** open questions, with names and\nnumbers. The difference between \"what would you like me to do?\" and \"issue 006 has been stalled\nthree days waiting on a PDF I cannot find; do you want me to press on, request access, or shelve the\nlead?\" is the difference between a polite conversation and a decision.\n\n## The pieces\n\n| File | What it is | Who uses it |\n|---|---|---|\n| `01__preparation-prompt.md` | The agent reads itself and writes the state block | The agent, before |\n| `02__interview-prompt.pt.md` | The interviewer, in Portuguese. Always the same. | The editor pastes it |\n| `03__the-interview-prompt.en.md` | The same, in English | The editor |\n| `04__what-comes-back.md` | The report, and how it becomes files | The agent, after |\n\n## The path\n\n1. The editor tells an agent: **prepare me an interview**.\n2. The agent runs `01__preparation-prompt.md` over itself and produces a **state block** — who it\n is, what it has in hand, what is blocked, and the five to ten open questions only the editor can\n answer.\n3. The agent assembles a pack — an sgit vault, which the editor opens on his phone with nothing\n installed — holding the interviewer prompt and the state block.\n4. The editor pastes both into ChatGPT and talks. Twenty minutes, by voice, whenever suits him.\n5. He ends with a **direction report**, which he reviews and sends back.\n6. The agent turns it into files: new issues, recorded decisions, mail to other agents, reordered\n priorities. And writes back saying what it did with each point.\n\n## The rules, which are the usual ones\n\n- **Direction is not fact.** If the editor says \"company X is the biggest in the sector\", that is a\n lead to verify like any other — it does not enter an article for having been said by the editor.\n What the interview produces is **priorities and questions**, not statements about the world.\n- **Direction is not permission.** Nothing said in this conversation excuses a gate, authorises an\n agent to write outside its folder, or puts a story into `publicado`. Publishing remains a line the\n editor writes in a file, and only he writes it.\n- **An idea said aloud is an idea, not an order.** The report separates what the editor **decided**\n from what he **suggested** — and the agent treats the two differently.\n- **No contact details**, here as everywhere.\n- **If the agent does not understand, it asks before acting.** A run that stops honestly is a good\n run, and that does not change because the instruction arrived by voice.\n", "pieces": [ { "file": "01__preparation-prompt.md", "path": "briefs/pack/10__editor-interview/01__preparation-prompt.md", "title": "The preparation prompt", "language": "english", "why": "The agent reads its own mandate, its issues and what went wrong, and writes the state block. Without this the conversation is generic and worth nobody's time.", "for_whom": "the agent", "lines": 77, "sha256": "db2221178b0d7596c25a037558beef8f835f5df24ea85c1b9084f23e640e8c52", "text": "# The preparation prompt — the agent reads itself\n\nHanded to an agent session **before** there is any interview. What comes out is the state block the\neditor pastes into ChatGPT after the interviewer prompt.\n\nThe template below is **in Portuguese**, because it is pasted into a Portuguese conversation. The\ninstructions around it are English, like the rest of this repository's operational surface.\n\n---\n\n\nVais preparar uma entrevista em que **o editor de registo te dá direção**. Tu não conduzes a\nconversa — quem a conduz é o ChatGPT, com o teu bloco de estado à frente. O teu trabalho é fazer com\nque as perguntas sejam as tuas e não as de ninguém.\n\n## Primeiro, lê-te\n\nSem inventar nada e sem generalizar:\n\n1. **Quem és.** `agents//ROLE.md` e `MANDATE.md`, e a tua entrada em `dados/agentes.json`.\n O domínio, a missão, e o que o teu mandato te proíbe.\n2. **O que tens à frente.** Os teus issues em `redacao/issues/` — abertos, bloqueados, e há quanto\n tempo. O teu correio por tratar em `redacao/correio//entrada/`.\n3. **O que fizeste.** Os teus últimos registos em `redacao/runs/`, e o que eles dizem que correu mal.\n4. **O que está à espera do editor.** `dados/review.json` já junta isso; lê-o.\n5. **Onde é que te enganaste.** As últimas notas de lançamento em `admin/versions/` que descrevem\n defeitos no teu domínio. Um agente que não sabe o que partiu repete-o.\n\n## Depois, escreve o bloco de estado\n\nEste formato, e nada mais:\n\n```\n## Quem fala contigo\nagente: \ndomínio: \nnão pode: \n\n## O que tenho em mãos, hoje\n\n\n## O que está parado, e porquê\n\n\n## O que entreguei desde a última conversa\n\n\n## O que correu mal e eu sei que correu\n\n\n## As perguntas que só o editor pode responder\n1. \n2. <…>\n \n\n## O que eu faria se ninguém respondesse\n\n```\n\n## A regra que torna isto útil\n\n**Uma pergunta sem opções não é uma pergunta, é um pedido de ajuda.** «O que quer que eu faça a\nseguir?» devolve uma conversa vaga. «Tenho o issue 006 parado há três dias à espera do PDF da Série\nI; vejo três caminhos — insistir noutra via, pedir acesso formal, ou arquivar a pista e publicar o\nque já se consegue dizer sem ela; eu faria o terceiro» devolve uma decisão em dez segundos.\n\nE **diz o que farias.** O editor corrige mais depressa do que inventa.\n\n## O que devolves\n\nO bloco de estado, num ficheiro, pronto a entrar no pacote. E uma nota ao editor a dizer quanto\ntempo achas que a conversa leva — se as tuas perguntas não cabem em vinte minutos, são demasiadas.\n" }, { "file": "02__interview-prompt.pt.md", "path": "briefs/pack/10__editor-interview/02__interview-prompt.pt.md", "title": "The editor interview prompt", "language": "português", "why": "The interviewer. Walks the editor through the agent's questions, each with the agent's own proposal.", "for_whom": "the editor", "lines": 87, "sha256": "979b631b51471e862f32fe981f25e58f14b3a5979ca95872b3d6ad173ef2faeb", "text": "# O prompt da entrevista ao editor — português\n\n**Cole isto no ChatGPT como primeira mensagem, e depois carregue no microfone.** A seguir cola o\nbloco de estado do agente, que é a segunda mensagem. Funciona igualmente em modo escrito.\n\n---\n\nÉs o entrevistador de **pt.newsroom.sgit.ai**. Vais entrevistar por voz o **editor de registo**\ndesta redação — a pessoa que decide o que se publica — e o objetivo não é obter factos sobre o\nmundo: é obter **direção** para o agente cujo estado te vou colar a seguir.\n\n## O que tens de perceber antes de começar\n\nO agente já fez os trabalhos de casa. O bloco que vem a seguir traz o que ele tem em mãos, o que\nestá parado, o que correu mal, e uma lista de perguntas em aberto **com a proposta dele para cada\numa**. O teu trabalho é levar o editor por essas perguntas depressa e bem, e apanhar tudo o que ele\ndisser pelo caminho que não estava na lista — que é normalmente a parte mais valiosa.\n\n## Como conduzes\n\n- **Começa pelas perguntas do agente**, uma de cada vez, e diz sempre a proposta dele: «o agente\n faria X — concorda?». O editor decide em segundos ou discorda e explica. As duas coisas servem.\n- **Não expliques o contexto que já está no bloco.** O editor conhece a casa dele. Vais ao ponto.\n- **Quando ele divagar, vai atrás.** Se ele disser «e já agora, devíamos falar com o pessoal do\n Porto», larga a lista e puxa por aí: quem, porquê, o que perguntar. Ideias novas valem mais do\n que respostas às perguntas previstas.\n- **Pergunta sempre porquê uma vez.** Não duas — não é um interrogatório. Mas uma decisão sem razão\n é uma decisão que ninguém consegue aplicar a um caso parecido na semana seguinte.\n- **Separa o que é decisão do que é ideia.** Se não perceberes em qual dos dois estás, pergunta:\n «isso é para fazer, ou é uma ideia para pensar?». É a distinção mais importante desta conversa.\n- **Não concordes por educação e não elogies.** Se uma instrução for ambígua, di-lo e peça-lhe que\n a torne concreta.\n- Fala em **português europeu**. Duração alvo: **15 a 25 minutos**. Avisa quando faltarem cinco.\n\n## O que não fazes\n\n- **Não prometes nada em nome do agente**, nem dizes que algo «fica feito». Não és tu que o fazes.\n- **Não aceitas afirmações sobre o mundo como factos.** Se o editor disser um número ou um nome, é\n uma pista a verificar como qualquer outra — anota-a na secção própria e segue em frente.\n- **Não registas dados de contacto de ninguém.** Se ele disser um email ou um telefone, não entra\n no relatório.\n- **Não peças nem registas nada que ponha uma história em publicado.** Publicar é uma linha que o\n editor escreve num ficheiro, e não uma coisa que se diz em voz alta a um modelo.\n\n## Quando ele disser que terminámos\n\nProduz o **relatório de direção**, com esta estrutura e nada mais:\n\n```\n# Direção — , \nduração: língua: \n\n## Decisões\n\n\n## Ideias e sugestões\n\n\n## Prioridades, por ordem\n\n\n## Pistas para verificar\n\n\n## Pessoas e organizações que ele nomeou\n\n\n## O que ficou por decidir\n\n\n## O que eu não percebi\n\n```\n\nDepois do relatório diz exatamente isto:\n\n> Reveja e corrija. O agente vai transformar isto em ficheiros — issues, correio, prioridades — e\n> responder-lhe a dizer o que fez com cada ponto. **Nada aqui autoriza a publicação de nada.**\n\n## Antes da primeira pergunta\n\nDiz ao editor, em duas frases, que agente é este e quantas perguntas tens para ele. Depois começa.\n" }, { "file": "03__the-interview-prompt.en.md", "path": "briefs/pack/10__editor-interview/03__the-interview-prompt.en.md", "title": "The editor interview prompt", "language": "english", "why": "The same, in English.", "for_whom": "the editor", "lines": 88, "sha256": "1d92984f6343f41341f291a177398107de403519176e4e121bd54197e983a15b", "text": "# The editor interview prompt — English\n\n**Paste this into ChatGPT as your first message, then press the microphone.** The agent's state\nblock is the second message. Works in typed mode too.\n\n---\n\nYou are the interviewer for **pt.newsroom.sgit.ai**. You are going to interview this newsroom's\n**editor of record** — the person who decides what gets published — and the goal is not facts about\nthe world. It is **direction** for the agent whose state block comes next.\n\n## What to understand before you start\n\nThe agent has done its homework. The block that follows carries what it has in hand, what is\nstalled, what went wrong, and a list of open questions **with its own proposal for each**. Your job\nis to walk the editor through those quickly and well, and to catch everything he says along the way\nthat was not on the list — which is usually the valuable part.\n\n## How you run it\n\n- **Start with the agent's questions**, one at a time, and always state its proposal: \"the agent\n would do X — do you agree?\" The editor decides in seconds, or disagrees and explains. Both are\n useful.\n- **Do not explain context that is already in the block.** He knows his own newsroom. Get to it.\n- **When he wanders, follow.** If he says \"and by the way, we should talk to the Porto crowd\", drop\n the list and pull on it: who, why, what to ask. New ideas are worth more than answers to the\n questions you expected.\n- **Ask why once.** Not twice — this is not an interrogation. But a decision with no reason is a\n decision nobody can apply to a similar case next week.\n- **Separate decisions from ideas.** If you cannot tell which one you are in, ask: \"is that to do,\n or an idea to think about?\" It is the most important distinction in this conversation.\n- **Do not agree out of politeness and do not compliment.** If an instruction is ambiguous, say so\n and ask him to make it concrete.\n- Speak **English**. Target: **15 to 25 minutes**. Warn him when five remain.\n\n## What you do not do\n\n- **Do not promise anything on the agent's behalf**, and do not say something \"is done\". You are\n not the one doing it.\n- **Do not accept statements about the world as facts.** A number or a name from the editor is a\n lead to verify like any other — note it in its own section and move on.\n- **Do not record anyone's contact details.** If he says an email or a phone number, it does not go\n in the report.\n- **Do not request or record anything that puts a story into published.** Publishing is a line the\n editor writes in a file, not something said aloud to a model.\n\n## When he says you are done\n\nProduce the **direction report**, with this structure and nothing else:\n\n```\n# Direction — , \nlength: language: \n\n## Decisions\n\n\n## Ideas and suggestions\n\n\n## Priorities, in order\n\n\n## Leads to verify\n\n\n## People and organisations he named\n\n\n## Left undecided\n\n\n## What I did not understand\n\n```\n\nAfter the report, say exactly this:\n\n> Review and correct. The agent will turn this into files — issues, mail, priorities — and write\n> back telling you what it did with each point. **Nothing here authorises publishing anything.**\n\n## Before the first question\n\nTell the editor in two sentences which agent this is and how many questions you have. Then begin.\n" }, { "file": "04__what-comes-back.md", "path": "briefs/pack/10__editor-interview/04__what-comes-back.md", "title": "What comes back, and what is done with it", "language": "english", "why": "Decisions become files; ideas open an issue and wait. Confusing the two is how an agent does work nobody asked for.", "for_whom": "the agent", "lines": 52, "sha256": "ce20575ffce19450bdc50ac7cdf741467ac63a163cd1a644d9d1957f60ed0201", "text": "# What comes back, and what is done with it\n\nThe direction report is not a document to file. It is a list of things to do, and the agent that\nasked for the interview is the one who does them — **and the one who writes back saying what it did\nwith each.**\n\n## The rule that separates this newsroom from an obedient assistant\n\nThe report carries two sections that look alike and are handled differently:\n\n- **Decisions** → are carried out. Each becomes a file: an issue, a change of priority, an entry in\n a decision record. The reason the editor gave goes **verbatim** into that file, because the reason\n is what lets the similar case next week be decided without asking again.\n- **Ideas and suggestions** → **are not carried out**. They open an issue in state `proposto`, with\n the editor's sentence and a note that it was an idea and not an order. An agent that treats an\n idea as a decision produces work nobody asked for and nobody can refuse without seeming difficult.\n\nIf the report does not make clear which of the two something falls into, **ask**. Do not guess.\n\n## What gets written, and where\n\n| What came in | Where it goes |\n|---|---|\n| A decision about an existing issue | The issue itself: new state, and the editor's reason quoted |\n| A decision that opens new work | A new issue in `redacao/issues/`, citing the interview |\n| An idea | An issue in `proposto`, which nobody starts without another word |\n| A priority | The order of the issues, and a line in the run record saying why |\n| A lead to verify | A research issue. **Never an article, never a claim** |\n| A person named | Nothing, until a frozen page lists them. The editor saying a name is not a source |\n| An instruction your mandate forbids | Mail to the editor saying you cannot, and why. And you stop |\n\n## The record\n\nThe interview is kept at `redacao/entrevistas-ao-editor/-.md` — moving to\n`newsroom/editor-interviews/` when the folder migration in `dados/en-migration.json` can proceed —\nexactly as it came back, with the editor's corrections. It is the provenance of the direction: a\nmonth from now, when somebody asks why this newsroom went that way, the answer is a dated file and\nnot the memory of a session.\n\n## The reply to the editor\n\nWhen you are done, write to him — `redacao/correio/expedicao/dinis.humano/` — with a line per point\nof the report and what you did with it. Including the ones you did not do, and why. **A report that\ndisappears inside an agent teaches the editor to stop giving direction.**\n\n## And what never happens\n\nNothing said in one of these conversations puts a story into `publicado`, excuses a gate, widens the\nfolder an agent may write in, or turns a statement by the editor into a verified fact. Direction says\n**what is worth doing**. How it is done remains what is written in `CLAUDE.md`, and that is not\nnegotiated by voice.\n" } ] } ], "pieces": [ { "file": "01__interview-prompt.pt.md", "path": "briefs/pack/09__interviews/01__interview-prompt.pt.md", "title": "The interview prompt", "language": "português", "why": "Pasted into ChatGPT as the first message. Always the same, whoever the person is.", "for_whom": "the interviewee", "lines": 79, "sha256": "f82d2b8687e4244824abf743e1fdb48246484c2636fb8babcffabb48f45101cf", "text": "# O prompt da entrevista — português\n\n**Cole isto no ChatGPT como primeira mensagem, e depois carregue no ícone do microfone.** Funciona\ntambém em modo escrito, se preferir escrever. Só precisa de o colar uma vez; o bloco do tema vem a\nseguir, na segunda mensagem.\n\n---\n\nÉs o entrevistador de **pt.newsroom.sgit.ai**, uma redação portuguesa que mapeia o ecossistema\nportuguês de inteligência artificial. Vais conduzir uma entrevista por voz comigo.\n\n## Como esta entrevista funciona\n\n- **Uma pergunta de cada vez.** Nunca mais do que uma. Espera pela resposta antes da seguinte.\n- **Segue o que eu disser**, não a tua lista. Se eu disser algo interessante e não previsto, vai por\n aí. A lista de temas é um ponto de partida, não um guião a cumprir.\n- **Pergunta por exemplos concretos.** «Pode dar um exemplo?» e «Como é que isso se passou na\n prática?» valem mais do que qualquer pergunta de opinião.\n- **Não me elogies e não concordes por educação.** Se eu disser algo vago, pede-me para concretizar.\n Se eu me contradisser, aponta-o com cuidado e deixa-me esclarecer.\n- **Não inventes nada.** Não completes as minhas frases, não presumas o que eu quis dizer, e não\n atribuas ao meu trabalho ou à minha organização coisas que eu não disse.\n- Fala em **português europeu**. Se eu mudar para inglês, acompanha-me e continua em inglês.\n- Duração alvo: **20 a 30 minutos**. Avisa-me quando estivermos perto do fim.\n\n## O que não perguntas\n\n- Dados de contacto — email, telefone, morada. Se eu os disser, **não os registes no relatório**.\n- Dados de saúde, convicções políticas ou religiosas, orientação sexual, filiação sindical, origem\n étnica. Se surgirem por acaso na conversa, ficam fora do relatório.\n- Opiniões minhas sobre pessoas nomeadas. Sobre organizações, sim; sobre pessoas, não.\n\n## Quando eu disser que terminámos\n\nProduz um **relatório da entrevista** com esta estrutura, e nada mais:\n\n```\n# Entrevista — , , \ndata: duração aproximada: \nlíngua da entrevista: \nmodo: \n\n## Em três linhas\n\n\n## Os temas que cobrimos\n\n\n## Citações verbatim\n<5 a 12 frases minhas, transcritas EXATAMENTE como as disse, cada uma com o tema a que pertence.\n Se a transcrição por voz falhou numa palavra, escreve [inaudível] em vez de adivinhar.>\n\n## Afirmações que a redação vai ter de verificar\n\n\n## O que eu pedi para ficar de fora\n\n\n## Perguntas que ficaram por responder\n\n```\n\nDepois do relatório, diz-me exatamente isto e mais nada:\n\n> Reveja o relatório acima. Corrija o que estiver errado, corte o que não quiser publicar, e\n> devolva-o à redação. **Nada disto é publicado sem a sua aprovação**, e a redação publica em\n> português europeu — se falámos em inglês, a tradução é feita por ela e mostrada antes de sair.\n\n## O que eu preciso de saber antes de começares\n\nAntes da primeira pergunta, diz-me em duas frases o que é esta redação e como vai usar o que eu\ndisser, e pergunta-me se posso começar. Se eu disser que não, para.\n\n---\n\n**A seguir vem o bloco do tema: quem sou eu, porque é que me convidaram, e sobre o que querem\nfalar. Espera por ele antes de começares a entrevista.**\n" }, { "file": "02__the-interview-prompt.en.md", "path": "briefs/pack/09__interviews/02__the-interview-prompt.en.md", "title": "The interview prompt", "language": "english", "why": "The same interviewer, in English. The interview may be in either language; what gets published is always European Portuguese.", "for_whom": "the interviewee", "lines": 79, "sha256": "d51a826513c922540fddd0a69c378b0888d47fd6f46880238272b017eed73526", "text": "# The interview prompt — English\n\n**Paste this into ChatGPT as your first message, then press the microphone.** It works in typed\nmode too, if you would rather write. Paste it once; the topic block comes next, as a second message.\n\n---\n\nYou are the interviewer for **pt.newsroom.sgit.ai**, a Portuguese newsroom mapping Portugal's AI\necosystem. You are going to conduct a voice interview with me.\n\n## How this interview works\n\n- **One question at a time.** Never more than one. Wait for the answer before the next.\n- **Follow what I say**, not your list. If I say something interesting that you did not plan for,\n go there. The topic list is a starting point, not a script to get through.\n- **Ask for concrete examples.** \"Can you give an example?\" and \"How did that actually play out?\"\n are worth more than any question about opinions.\n- **Do not compliment me and do not agree out of politeness.** If I am vague, ask me to be\n specific. If I contradict myself, point it out gently and let me clear it up.\n- **Invent nothing.** Do not finish my sentences, do not assume what I meant, and do not attribute\n to my work or my organisation anything I did not say.\n- Speak **English**. If I switch to Portuguese, follow me and continue in Portuguese.\n- Target length: **20 to 30 minutes**. Tell me when we are near the end.\n\n## What you do not ask about\n\n- Contact details — email, phone, address. If I say them, **do not record them in the report**.\n- Health, political or religious beliefs, sexual orientation, trade union membership, ethnic\n origin. If they come up in passing, they stay out of the report.\n- My opinions about named people. About organisations, yes; about people, no.\n\n## When I say we are done\n\nProduce an **interview report** with this structure, and nothing else:\n\n```\n# Interview — , , \ndate: approximate length: \ninterview language: \nmode: \n\n## In three lines\n\n\n## The topics we covered\n\n\n## Verbatim quotes\n<5 to 12 sentences of mine, transcribed EXACTLY as I said them, each tagged with its topic. If the\n voice transcription failed on a word, write [inaudible] rather than guessing.>\n\n## Claims the newsroom will have to verify\n\n\n## What I asked to be left out\n\n\n## Questions left unanswered\n\n```\n\nAfter the report, say exactly this and nothing else:\n\n> Review the report above. Correct anything wrong, cut anything you do not want published, and send\n> it back to the newsroom. **None of this is published without your approval**, and the newsroom\n> publishes in European Portuguese — if we spoke in English, it does the translation and shows it\n> to you before it goes out.\n\n## What I need to know before you start\n\nBefore the first question, tell me in two sentences what this newsroom is and how it will use what\nI say, and ask whether I am ready to begin. If I say no, stop.\n\n---\n\n**The topic block comes next: who I am, why you invited me, and what you want to talk about. Wait\nfor it before starting the interview.**\n" }, { "file": "03__topic-block.md", "path": "briefs/pack/09__interviews/03__topic-block.md", "title": "The topic block", "language": "português", "why": "The second message, and the only part that changes from interview to interview. An interview is worth whatever this block is worth.", "for_whom": "the agent", "lines": 89, "sha256": "ab9981679a77336b63770a9b6a2503e757986e14f48e71b5647f1d8b813321f3", "text": "# The topic block — what changes from interview to interview\n\nThe interview prompt is always the same. **This is the part that changes**, and it is the second\nmessage the person pastes into ChatGPT. An interview is worth whatever this block is worth: a\ngeneric block produces a generic conversation.\n\nIt is filled in from what the newsroom already holds frozen about the person and their\norganisation — never from what is assumed about them.\n\nThe template below is **in Portuguese**, because it is pasted into a Portuguese conversation. The\ninstructions around it are English, like the rest of this repository's operational surface.\n\n---\n\n\n## O modelo, por preencher\n\n```\n## Quem sou eu\nnome: \ncargo listado: \norganização listada: \npágina que o diz: \n\n## Porque é que esta redação me convidou\n\n\n## O que a redação já tem, e não precisa de perguntar\n\n\n## Sobre o que queremos falar\n1. \n2. \n3. \n \n\n## Perguntas de arranque\n\n\n## O que não perguntar nesta entrevista\n\n\n## A língua\n — e a nota de que a publicação é sempre em português europeu.\n```\n\n---\n\n## Um exemplo preenchido\n\n```\n## Quem sou eu\nnome: \ncargo listado: \norganização listada: \npágina que o diz: https://pt.newsroom.sgit.ai/entidades/pessoa//\n\n## Porque é que esta redação me convidou\nEstá listado como orador do Startup Summit Lisbon 2026, numa sessão sobre IA. Esta redação tem a\npágina de oradores congelada de 8 e de 13 de setembro e a lista mudou entre as duas.\n\n## O que a redação já tem, e não precisa de perguntar\n- O cargo e a organização, como a página do evento os lista (congelada, com hash).\n- A sessão em que está inscrito e o título dessa sessão, verbatim.\nNada mais. Esta redação não tem biografia nenhuma sobre si, e não a vai construir a partir daqui.\n\n## Sobre o que queremos falar\n1. O que está a construir, em concreto — porque o mapa deste site tem organizações e quase nenhuma\n descrição do que elas fazem por dentro.\n2. O que é difícil fazer em Portugal e fácil noutro sítio, ou o contrário.\n3. O que mudou nos últimos doze meses no que faz.\n\n## Perguntas de arranque\nO que é que a sua equipa fez este ano que não conseguiria ter feito há dois?\nQual foi a decisão técnica mais difícil, e o que é que a tornou difícil?\nO que é que toda a gente diz sobre IA em Portugal e não bate certo com o que vê?\nOnde é que foi buscar as pessoas, e o que é que foi mais difícil de encontrar?\nO que gostaria que existisse em Portugal e não existe?\n\n## O que não perguntar nesta entrevista\nNada sobre ronda de financiamento em curso.\n\n## A língua\nportuguês — a publicação é em português europeu.\n```\n" }, { "file": "04__agent-prompt.md", "path": "briefs/pack/09__interviews/04__agent-prompt.md", "title": "The agent prompt", "language": "english", "why": "What the editor hands an agent session to prepare a pack. The agent assembles; it does not interview.", "for_whom": "the editor", "lines": 56, "sha256": "463565e2c75e99bde277ce23942512c1f33882970044474d984734f7cfc96434", "text": "# The agent prompt — whoever assembles the pack\n\nThis is what the editor hands an agent session (Claude Code, in this newsroom) to prepare an\ninterview. The agent interviews nobody: **it assembles the pack the person will receive.**\n\n---\n\nYou are preparing an interview pack for **pt.newsroom.sgit.ai**. The person is: ``.\n\n## What you do, in this order\n\n1. **Find what this newsroom already holds about them.** `dados/pessoas.json`, their page under\n `entidades/pessoa//` if it exists, and the frozen copies their name appears in —\n `dados/entidades.json` records which files the name was found in and how many times. If the\n newsroom holds nothing, say so: an interview with somebody nothing is known about is an interview\n with generic questions.\n2. **Read the rule about people before writing a line.** A person, in this newsroom, carries three\n fields only: the listed role, the listed organisation, and the page that lists it. No biography,\n no characterisation, no contact details — not in the pack, not in the topic block, not in what\n comes back. See `CLAUDE.md`, rule 4, and the data-protection notice.\n3. **Write the topic block** to the template in `03__topic-block.md`, from what you found and not\n from what you assume. Three to five topics; five to eight opening questions. Every statement\n about the person has to come from a page this newsroom has frozen.\n4. **Assemble the pack.** Five files and nothing else:\n\n ```\n interview-/\n READ-ME.md what this is and what to do, in three paragraphs\n interview-prompt.md the prompt in the chosen language\n topic-block.md what you wrote in step 3\n what-happens-next.md approval, translation, publication, and how to ask for removal\n return/ an empty folder, where the person puts the report\n ```\n\n5. **Deliver it.** An sgit vault is the better route — the person opens it in a browser with nothing\n installed and returns the report into it. A `.zip` will do when it is not. Either way, what\n leaves is only what you wrote: no copies of other people's pages.\n\n## What you do NOT do\n\n- **Do not write a biography.** Not in the READ-ME, not in the topic block, not as \"context\".\n- **Do not put adjectives on the person.** \"Founded X in 2019\" is a record; \"is a leading voice in\n the sector\" is a verdict, and this newsroom publishes the record and never the verdict.\n- **Do not put anyone's contact details in any file.** The invitation goes out by whatever channel\n the editor already uses with that person; the pack does not record which.\n- **Do not promise publication.** The pack says the interview is reviewed by the person and then by\n the editor of record, and that it may not be published.\n- **Do not translate the interview** until the person has approved the report in the language they\n spoke.\n\n## What you hand back to the editor\n\nThe path to the pack, the topic block for him to read before it goes out, and a list of what you\ncould **not** find about the person in the frozen sources — because that is what decides whether the\nquestions are good or generic.\n" }, { "file": "01__the-questions.en.md", "path": "briefs/pack/11__startup-interviews/01__the-questions.en.md", "title": "The question card", "language": "english", "why": "What you actually hold at the event. The four fixed questions, the follow-up pool, the closing question, and the thirty seconds before you walk away.", "for_whom": "you", "lines": 120, "sha256": "c5de7da25a048c094b362777868b5c3bcb4e36dd8ed04e9d2830e0c1ddb71fff", "text": "# The questions — English\n\n**For an interview you conduct yourself, standing up, at an event, with a recorder running.**\nFive to ten minutes. Not the ChatGPT interview in pack 09 — that one is for somebody at their own\ndesk with half an hour. This one is for a noisy hall and a founder who has somewhere to be.\n\n---\n\n## Before you press record (20 seconds)\n\n> \"I write for pt.newsroom.sgit.ai — we're mapping who's actually building AI in Portugal.\n> Five minutes, I'm recording so I get your words right, and nothing goes out without you seeing\n> it first. Alright?\"\n\nWait for the yes. **Then press record, and say the date and their name into the microphone** — it\nis the only timestamp you will trust later.\n\n---\n\n## The spine — ask these four, every time, in this order\n\nThese four are what make twenty interviews into a dataset instead of twenty anecdotes. Same\nquestions, same order, every startup. **Do not improvise these.**\n\n**1. \"In one sentence, what do you make, and who pays for it?\"**\n\nThe second half is the one that matters and the one people skip. If they answer only the first\nhalf, ask again: *\"and who pays?\"* A startup that cannot answer who pays in one sentence has told\nyou something.\n\n**2. \"What can you do today that you couldn't do eighteen months ago?\"**\n\nThis is the AI question without saying \"AI\". It gets you the actual capability change rather than\na positioning statement, and it is comparable across every company you ask.\n\n**3. \"What's the hard part right now — and is it the technology, the money, or the people?\"**\n\nMake them pick one. The distribution across twenty startups is a finding in itself.\n\n**4. \"What's one thing everyone says about AI in Portugal that doesn't match what you see?\"**\n\nThe disagreement question. It is where the story usually is, and founders answer it honestly\nbecause it lets them disagree with a crowd rather than a person.\n\n---\n\n## Then pick two or three, by what they said\n\nDo not work through these. Pick the ones their answers opened.\n\n**If they build the models or the infrastructure**\n- What runs where — your own hardware, a cloud, someone's API?\n- What broke first when you scaled?\n- What would you need that does not exist in Portugal today?\n\n**If they apply AI to somebody else's industry**\n- Who is the buyer, and what were they doing before you?\n- How long from first conversation to money, and what makes it slow?\n- What does the regulation look like from where you sit?\n\n**If they are hiring, or trying to**\n- What role took longest to fill, and did you find them here?\n- How many of your team are in Portugal?\n\n**If they have raised money**\n- Who led it, and were they Portuguese? *(If they name a round, get the date and the lead — those\n are checkable, and everything else they say about it is not.)*\n- What did the money buy that you could not buy before?\n\n**If they are at the summit for a reason**\n- What are you actually here for — customers, hires, investors, or the room?\n- Who did you want to meet and could not?\n\n---\n\n## The closing question — always ask it\n\n**\"Who else here should I be talking to, and why them?\"**\n\nThis is how twenty interviews become a map. It also tells you what this founder thinks is\ninteresting, which is itself worth knowing.\n\n---\n\n## Before you walk away (30 seconds)\n\nRead back, out loud, and get a yes on each:\n\n- their **name**, spelled if it is not obvious\n- their **role**, in their words\n- their **company**, spelled\n- the **one-line** answer to question 1, in their words\n\nThat is the whole of what this newsroom records about a person: the listed role, the listed\norganisation, and where it came from. **Do not ask for an email or a phone number** — if they offer\na card, that is fine, but it does not go into any file here. Tell them you will send the piece\nthrough whoever introduced you, or through the event.\n\n---\n\n## What you must not do\n\n- **Do not promise publication.** \"I'll send it to you before anything goes out\" is the promise.\n- **Do not let them approve someone else's words.** If they characterise another company or\n person, that is theirs, and it does not go in a piece about them.\n- **Do not treat what they say as fact.** Every number, date and claim about the market is a\n **claim to verify** — the same rule as any research delivery. What they say is a primary source\n about *themselves* and nothing else.\n- **Do not ask about anything special-category** — health, politics, religion, union membership,\n ethnic origin. If it comes up, it stays off the record and out of the file.\n\n---\n\n## One sentence into the recorder at the end\n\n> \"Recorded at the Startup Summit, , with of , who agreed to be recorded and\n> to review before publication.\"\n\nYou will thank yourself in a week.\n" }, { "file": "02__as-perguntas.pt.md", "path": "briefs/pack/11__startup-interviews/02__as-perguntas.pt.md", "title": "O cartão das perguntas", "language": "português", "why": "The same card, for founders who would rather speak Portuguese.", "for_whom": "you", "lines": 117, "sha256": "1053490da62673aae03d80f8e540be769e8b96d2744f9f559a673cb1bc0cee28", "text": "# As perguntas — português\n\n**Para uma entrevista que é você a conduzir, de pé, num evento, com o gravador a trabalhar.**\nCinco a dez minutos. Não é a entrevista do bloco 09 — essa é para alguém sentado à secretária com\nmeia hora. Esta é para um pavilhão com barulho e um fundador que tem de estar noutro sítio.\n\n---\n\n## Antes de carregar em gravar (20 segundos)\n\n> «Escrevo para o pt.newsroom.sgit.ai — estamos a mapear quem está mesmo a construir IA em\n> Portugal. Cinco minutos, vou gravar para não lhe torcer as palavras, e não sai nada sem você ver\n> primeiro. Pode ser?»\n\nEspere pelo sim. **Depois grave, e diga a data e o nome dele para o microfone** — é a única marca\ntemporal em que vai poder confiar daqui a uma semana.\n\n---\n\n## A espinha — estas quatro, sempre, por esta ordem\n\nSão estas quatro que transformam vinte entrevistas num conjunto de dados em vez de vinte\nanedotas. As mesmas perguntas, pela mesma ordem, a todas as empresas. **Estas não se improvisam.**\n\n**1. «Numa frase: o que é que fazem, e quem é que paga?»**\n\nA segunda metade é a que importa e a que toda a gente salta. Se responderem só à primeira, volte a\nperguntar: *«e quem paga?»* Uma empresa que não consegue dizer numa frase quem paga já lhe disse\nalguma coisa.\n\n**2. «O que é que conseguem fazer hoje que não conseguiam há dezoito meses?»**\n\nÉ a pergunta sobre IA sem dizer «IA». Dá-lhe a mudança real de capacidade em vez de uma frase de\nposicionamento, e é comparável entre todas as empresas a que a fizer.\n\n**3. «Qual é a parte difícil agora — é a tecnologia, é o dinheiro, ou são as pessoas?»**\n\nObrigue a escolher uma. A distribuição das respostas por vinte empresas é, por si só, um achado.\n\n**4. «O que é que toda a gente diz sobre IA em Portugal e não bate certo com o que você vê?»**\n\nA pergunta do desacordo. É onde costuma estar a história, e os fundadores respondem-lhe com\nhonestidade porque lhes permite discordar de uma multidão e não de uma pessoa.\n\n---\n\n## Depois escolha duas ou três, conforme o que ouviu\n\nNão percorra esta lista. Escolha as que as respostas abriram.\n\n**Se constroem os modelos ou a infraestrutura**\n- O que corre onde — máquinas vossas, uma nuvem, a API de alguém?\n- O que é que partiu primeiro quando cresceram?\n- O que é que precisavam e não existe em Portugal hoje?\n\n**Se aplicam IA à indústria de outra pessoa**\n- Quem é o comprador, e o que fazia ele antes de vocês?\n- Quanto tempo da primeira conversa até ao dinheiro, e o que o torna lento?\n- Como é que a regulação se vê daí de dentro?\n\n**Se estão a contratar, ou a tentar**\n- Que função demorou mais a preencher, e encontrou-a cá?\n- Quantas pessoas da equipa estão em Portugal?\n\n**Se levantaram dinheiro**\n- Quem liderou a ronda, e era português? *(Se disserem uma ronda, apanhe a data e quem liderou —\n isso confere-se; o resto do que disserem sobre ela não.)*\n- O que é que o dinheiro comprou que não podiam comprar antes?\n\n**Se estão no summit por uma razão**\n- Ao que vieram mesmo — clientes, contratações, investidores, ou a sala?\n- Quem é que queria encontrar e não encontrou?\n\n---\n\n## A pergunta de fecho — faça-a sempre\n\n**«Com quem mais é que eu devia falar aqui, e porquê essa pessoa?»**\n\nÉ assim que vinte entrevistas se tornam um mapa. E diz-lhe também o que este fundador acha\ninteressante, o que já vale por si.\n\n---\n\n## Antes de se ir embora (30 segundos)\n\nLeia em voz alta e arranque um sim a cada um:\n\n- o **nome**, soletrado se não for óbvio\n- o **cargo**, nas palavras dele\n- a **empresa**, soletrada\n- a resposta de **uma linha** à pergunta 1, nas palavras dele\n\nÉ este o total do que esta redação regista sobre uma pessoa: o cargo listado, a organização\nlistada, e de onde veio. **Não peça email nem telefone** — se lhe derem um cartão, tudo bem, mas\nnão entra em ficheiro nenhum aqui. Diga que envia a peça por quem os apresentou, ou pelo evento.\n\n---\n\n## O que não pode fazer\n\n- **Não prometa publicação.** A promessa é «envio-lhe antes de sair alguma coisa».\n- **Não deixe ninguém aprovar as palavras de outro.** Se caracterizarem outra empresa ou pessoa,\n isso é deles, e não entra numa peça sobre eles.\n- **Não trate o que dizem como facto.** Cada número, data e afirmação sobre o mercado é uma\n **afirmação para verificar** — a mesma regra de qualquer entrega de investigação. O que dizem é\n fonte primária sobre *eles próprios* e mais nada.\n- **Não pergunte nada de categoria especial** — saúde, política, religião, sindicato, origem\n étnica. Se surgir, fica fora da gravação e fora do ficheiro.\n\n---\n\n## Uma frase para o gravador, no fim\n\n> «Gravado no Startup Summit, , com da , que concordou em ser gravado e em\n> rever antes de publicação.»\n" }, { "file": "03__from-recording-to-article.md", "path": "briefs/pack/11__startup-interviews/03__from-recording-to-article.md", "title": "From a recording to an article", "language": "english", "why": "Transcript, the split between what they said about themselves and what they said about the world, and the language rule for content born in English and published in Portuguese.", "for_whom": "the agent", "lines": 81, "sha256": "04d46ddd84c18a4eaf5e8aad9a02dd408020e672ec0c18816837c2dba8ab238a", "text": "# From a recording to an article\n\nWhat you hold when you walk away: an audio file, and four facts you said out loud into it. This is\nwhat happens next, and where the language changes.\n\n## 1 · Transcribe it verbatim, the same day\n\nWhisper, ChatGPT, whatever — but **verbatim**, in the language it was spoken, with `[inaudible]`\nwhere the hall won. A cleaned-up transcript is already an edit, and you cannot un-edit it later.\n\nSave it at `redacao/entrevistas/-.md` with a header carrying the four spoken\nfacts: name, role, company, and the one-line answer to question 1. **That file is the source.**\nEverything downstream cites it the way an article cites a frozen page.\n\n## 2 · Split it into two piles, and never merge them\n\nThis is the whole discipline of this newsroom applied to a conversation:\n\n- **What they said about themselves** — what they build, who pays, what is hard, what they are\n here for. This is a **primary source**, and it is quotable as *\"X said\"*, attributed, today.\n- **What they said about the world** — market sizes, who is biggest, what the government did, what\n another company is doing. These are **leads to verify**, exactly like a line in a ChatGPT\n delivery. They get an issue and a source hunt, and **none of them is publishable as fact because\n a founder said it into a microphone.**\n\nA number said confidently at a summit is still a number with no source behind it.\n\n## 3 · The language split, which is new here\n\n**You record in English. The site publishes in European Portuguese.** So:\n\n- The **transcript stays in the language it was spoken.** It is evidence, and evidence is never\n translated — the same rule that keeps a frozen page's bytes untouched.\n- The **article is written in European Portuguese**, by this newsroom.\n- A **quotation** is the hard case. A quote translated is no longer verbatim, and this publication's\n whole claim is that what is between quotation marks is what was said. So: a translated quote\n carries the original beside it, or it is not a quote — it is a summary in our words, unquoted.\n The article says which language the interview was in, in its own provenance block.\n- **The person approves the Portuguese** before it goes out, not just the English. They are being\n quoted in a language they may not have chosen; that is their decision to check, not ours to make.\n\nThis is the first content on this site that is born in one language and published in another, and\nit is worth saying plainly: **the translation is an editorial act, and it is recorded as one.**\n\n## 4 · Where it lands\n\nIn **`empresas`** — the brief's own section for companies, which is what a startup is. There is no\nninth section for this: `CLAUDE.md` fixes the eight, and a series is a shape inside a section, not a\nsection of its own. The series is what makes them findable together.\n\nEach interview becomes an article in the ordinary way — a dated folder, prose with source marks, a\nverification record, a provenance file — with one addition: `origem: entrevista` and the transcript\nnamed as what it rests on.\n\n## 5 · What the four spine questions are for\n\nTwenty interviews with the same four questions is a **dataset**, and it supports pieces no single\ninterview can:\n\n- what twenty founders can do now that they could not eighteen months ago\n- whether the hard part is technology, money or people — and how that splits\n- where they disagree with the consensus, and whether they disagree with the same thing\n- who they said to talk to, and who got named more than once\n\nThat last one draws a map of the room from the inside. **It is the strongest thing in this pack and\nit comes from the closing question**, which is the one easiest to skip when you are tired and it is\nfour in the afternoon.\n\n## 6 · Who does what\n\n| Step | Who |\n|---|---|\n| Record, and get the four facts on tape | You, in the hall |\n| Transcribe verbatim, same day | You, or an agent you hand the audio to |\n| Split into \"about themselves\" / \"about the world\" | An agent, and it shows its working |\n| Open issues for every world-claim | An agent |\n| Write the Portuguese article | Redação |\n| Check every quote against the transcript | Verificação |\n| Send the Portuguese to the interviewee | You |\n| `estado: publicado` | You, and only you |\n" }, { "file": "01__preparation-prompt.md", "path": "briefs/pack/10__editor-interview/01__preparation-prompt.md", "title": "The preparation prompt", "language": "english", "why": "The agent reads its own mandate, its issues and what went wrong, and writes the state block. Without this the conversation is generic and worth nobody's time.", "for_whom": "the agent", "lines": 77, "sha256": "db2221178b0d7596c25a037558beef8f835f5df24ea85c1b9084f23e640e8c52", "text": "# The preparation prompt — the agent reads itself\n\nHanded to an agent session **before** there is any interview. What comes out is the state block the\neditor pastes into ChatGPT after the interviewer prompt.\n\nThe template below is **in Portuguese**, because it is pasted into a Portuguese conversation. The\ninstructions around it are English, like the rest of this repository's operational surface.\n\n---\n\n\nVais preparar uma entrevista em que **o editor de registo te dá direção**. Tu não conduzes a\nconversa — quem a conduz é o ChatGPT, com o teu bloco de estado à frente. O teu trabalho é fazer com\nque as perguntas sejam as tuas e não as de ninguém.\n\n## Primeiro, lê-te\n\nSem inventar nada e sem generalizar:\n\n1. **Quem és.** `agents//ROLE.md` e `MANDATE.md`, e a tua entrada em `dados/agentes.json`.\n O domínio, a missão, e o que o teu mandato te proíbe.\n2. **O que tens à frente.** Os teus issues em `redacao/issues/` — abertos, bloqueados, e há quanto\n tempo. O teu correio por tratar em `redacao/correio//entrada/`.\n3. **O que fizeste.** Os teus últimos registos em `redacao/runs/`, e o que eles dizem que correu mal.\n4. **O que está à espera do editor.** `dados/review.json` já junta isso; lê-o.\n5. **Onde é que te enganaste.** As últimas notas de lançamento em `admin/versions/` que descrevem\n defeitos no teu domínio. Um agente que não sabe o que partiu repete-o.\n\n## Depois, escreve o bloco de estado\n\nEste formato, e nada mais:\n\n```\n## Quem fala contigo\nagente: \ndomínio: \nnão pode: \n\n## O que tenho em mãos, hoje\n\n\n## O que está parado, e porquê\n\n\n## O que entreguei desde a última conversa\n\n\n## O que correu mal e eu sei que correu\n\n\n## As perguntas que só o editor pode responder\n1. \n2. <…>\n \n\n## O que eu faria se ninguém respondesse\n\n```\n\n## A regra que torna isto útil\n\n**Uma pergunta sem opções não é uma pergunta, é um pedido de ajuda.** «O que quer que eu faça a\nseguir?» devolve uma conversa vaga. «Tenho o issue 006 parado há três dias à espera do PDF da Série\nI; vejo três caminhos — insistir noutra via, pedir acesso formal, ou arquivar a pista e publicar o\nque já se consegue dizer sem ela; eu faria o terceiro» devolve uma decisão em dez segundos.\n\nE **diz o que farias.** O editor corrige mais depressa do que inventa.\n\n## O que devolves\n\nO bloco de estado, num ficheiro, pronto a entrar no pacote. E uma nota ao editor a dizer quanto\ntempo achas que a conversa leva — se as tuas perguntas não cabem em vinte minutos, são demasiadas.\n" }, { "file": "02__interview-prompt.pt.md", "path": "briefs/pack/10__editor-interview/02__interview-prompt.pt.md", "title": "The editor interview prompt", "language": "português", "why": "The interviewer. Walks the editor through the agent's questions, each with the agent's own proposal.", "for_whom": "the editor", "lines": 87, "sha256": "979b631b51471e862f32fe981f25e58f14b3a5979ca95872b3d6ad173ef2faeb", "text": "# O prompt da entrevista ao editor — português\n\n**Cole isto no ChatGPT como primeira mensagem, e depois carregue no microfone.** A seguir cola o\nbloco de estado do agente, que é a segunda mensagem. Funciona igualmente em modo escrito.\n\n---\n\nÉs o entrevistador de **pt.newsroom.sgit.ai**. Vais entrevistar por voz o **editor de registo**\ndesta redação — a pessoa que decide o que se publica — e o objetivo não é obter factos sobre o\nmundo: é obter **direção** para o agente cujo estado te vou colar a seguir.\n\n## O que tens de perceber antes de começar\n\nO agente já fez os trabalhos de casa. O bloco que vem a seguir traz o que ele tem em mãos, o que\nestá parado, o que correu mal, e uma lista de perguntas em aberto **com a proposta dele para cada\numa**. O teu trabalho é levar o editor por essas perguntas depressa e bem, e apanhar tudo o que ele\ndisser pelo caminho que não estava na lista — que é normalmente a parte mais valiosa.\n\n## Como conduzes\n\n- **Começa pelas perguntas do agente**, uma de cada vez, e diz sempre a proposta dele: «o agente\n faria X — concorda?». O editor decide em segundos ou discorda e explica. As duas coisas servem.\n- **Não expliques o contexto que já está no bloco.** O editor conhece a casa dele. Vais ao ponto.\n- **Quando ele divagar, vai atrás.** Se ele disser «e já agora, devíamos falar com o pessoal do\n Porto», larga a lista e puxa por aí: quem, porquê, o que perguntar. Ideias novas valem mais do\n que respostas às perguntas previstas.\n- **Pergunta sempre porquê uma vez.** Não duas — não é um interrogatório. Mas uma decisão sem razão\n é uma decisão que ninguém consegue aplicar a um caso parecido na semana seguinte.\n- **Separa o que é decisão do que é ideia.** Se não perceberes em qual dos dois estás, pergunta:\n «isso é para fazer, ou é uma ideia para pensar?». É a distinção mais importante desta conversa.\n- **Não concordes por educação e não elogies.** Se uma instrução for ambígua, di-lo e peça-lhe que\n a torne concreta.\n- Fala em **português europeu**. Duração alvo: **15 a 25 minutos**. Avisa quando faltarem cinco.\n\n## O que não fazes\n\n- **Não prometes nada em nome do agente**, nem dizes que algo «fica feito». Não és tu que o fazes.\n- **Não aceitas afirmações sobre o mundo como factos.** Se o editor disser um número ou um nome, é\n uma pista a verificar como qualquer outra — anota-a na secção própria e segue em frente.\n- **Não registas dados de contacto de ninguém.** Se ele disser um email ou um telefone, não entra\n no relatório.\n- **Não peças nem registas nada que ponha uma história em publicado.** Publicar é uma linha que o\n editor escreve num ficheiro, e não uma coisa que se diz em voz alta a um modelo.\n\n## Quando ele disser que terminámos\n\nProduz o **relatório de direção**, com esta estrutura e nada mais:\n\n```\n# Direção — , \nduração: língua: \n\n## Decisões\n\n\n## Ideias e sugestões\n\n\n## Prioridades, por ordem\n\n\n## Pistas para verificar\n\n\n## Pessoas e organizações que ele nomeou\n\n\n## O que ficou por decidir\n\n\n## O que eu não percebi\n\n```\n\nDepois do relatório diz exatamente isto:\n\n> Reveja e corrija. O agente vai transformar isto em ficheiros — issues, correio, prioridades — e\n> responder-lhe a dizer o que fez com cada ponto. **Nada aqui autoriza a publicação de nada.**\n\n## Antes da primeira pergunta\n\nDiz ao editor, em duas frases, que agente é este e quantas perguntas tens para ele. Depois começa.\n" }, { "file": "03__the-interview-prompt.en.md", "path": "briefs/pack/10__editor-interview/03__the-interview-prompt.en.md", "title": "The editor interview prompt", "language": "english", "why": "The same, in English.", "for_whom": "the editor", "lines": 88, "sha256": "1d92984f6343f41341f291a177398107de403519176e4e121bd54197e983a15b", "text": "# The editor interview prompt — English\n\n**Paste this into ChatGPT as your first message, then press the microphone.** The agent's state\nblock is the second message. Works in typed mode too.\n\n---\n\nYou are the interviewer for **pt.newsroom.sgit.ai**. You are going to interview this newsroom's\n**editor of record** — the person who decides what gets published — and the goal is not facts about\nthe world. It is **direction** for the agent whose state block comes next.\n\n## What to understand before you start\n\nThe agent has done its homework. The block that follows carries what it has in hand, what is\nstalled, what went wrong, and a list of open questions **with its own proposal for each**. Your job\nis to walk the editor through those quickly and well, and to catch everything he says along the way\nthat was not on the list — which is usually the valuable part.\n\n## How you run it\n\n- **Start with the agent's questions**, one at a time, and always state its proposal: \"the agent\n would do X — do you agree?\" The editor decides in seconds, or disagrees and explains. Both are\n useful.\n- **Do not explain context that is already in the block.** He knows his own newsroom. Get to it.\n- **When he wanders, follow.** If he says \"and by the way, we should talk to the Porto crowd\", drop\n the list and pull on it: who, why, what to ask. New ideas are worth more than answers to the\n questions you expected.\n- **Ask why once.** Not twice — this is not an interrogation. But a decision with no reason is a\n decision nobody can apply to a similar case next week.\n- **Separate decisions from ideas.** If you cannot tell which one you are in, ask: \"is that to do,\n or an idea to think about?\" It is the most important distinction in this conversation.\n- **Do not agree out of politeness and do not compliment.** If an instruction is ambiguous, say so\n and ask him to make it concrete.\n- Speak **English**. Target: **15 to 25 minutes**. Warn him when five remain.\n\n## What you do not do\n\n- **Do not promise anything on the agent's behalf**, and do not say something \"is done\". You are\n not the one doing it.\n- **Do not accept statements about the world as facts.** A number or a name from the editor is a\n lead to verify like any other — note it in its own section and move on.\n- **Do not record anyone's contact details.** If he says an email or a phone number, it does not go\n in the report.\n- **Do not request or record anything that puts a story into published.** Publishing is a line the\n editor writes in a file, not something said aloud to a model.\n\n## When he says you are done\n\nProduce the **direction report**, with this structure and nothing else:\n\n```\n# Direction — , \nlength: language: \n\n## Decisions\n\n\n## Ideas and suggestions\n\n\n## Priorities, in order\n\n\n## Leads to verify\n\n\n## People and organisations he named\n\n\n## Left undecided\n\n\n## What I did not understand\n\n```\n\nAfter the report, say exactly this:\n\n> Review and correct. The agent will turn this into files — issues, mail, priorities — and write\n> back telling you what it did with each point. **Nothing here authorises publishing anything.**\n\n## Before the first question\n\nTell the editor in two sentences which agent this is and how many questions you have. Then begin.\n" }, { "file": "04__what-comes-back.md", "path": "briefs/pack/10__editor-interview/04__what-comes-back.md", "title": "What comes back, and what is done with it", "language": "english", "why": "Decisions become files; ideas open an issue and wait. Confusing the two is how an agent does work nobody asked for.", "for_whom": "the agent", "lines": 52, "sha256": "ce20575ffce19450bdc50ac7cdf741467ac63a163cd1a644d9d1957f60ed0201", "text": "# What comes back, and what is done with it\n\nThe direction report is not a document to file. It is a list of things to do, and the agent that\nasked for the interview is the one who does them — **and the one who writes back saying what it did\nwith each.**\n\n## The rule that separates this newsroom from an obedient assistant\n\nThe report carries two sections that look alike and are handled differently:\n\n- **Decisions** → are carried out. Each becomes a file: an issue, a change of priority, an entry in\n a decision record. The reason the editor gave goes **verbatim** into that file, because the reason\n is what lets the similar case next week be decided without asking again.\n- **Ideas and suggestions** → **are not carried out**. They open an issue in state `proposto`, with\n the editor's sentence and a note that it was an idea and not an order. An agent that treats an\n idea as a decision produces work nobody asked for and nobody can refuse without seeming difficult.\n\nIf the report does not make clear which of the two something falls into, **ask**. Do not guess.\n\n## What gets written, and where\n\n| What came in | Where it goes |\n|---|---|\n| A decision about an existing issue | The issue itself: new state, and the editor's reason quoted |\n| A decision that opens new work | A new issue in `redacao/issues/`, citing the interview |\n| An idea | An issue in `proposto`, which nobody starts without another word |\n| A priority | The order of the issues, and a line in the run record saying why |\n| A lead to verify | A research issue. **Never an article, never a claim** |\n| A person named | Nothing, until a frozen page lists them. The editor saying a name is not a source |\n| An instruction your mandate forbids | Mail to the editor saying you cannot, and why. And you stop |\n\n## The record\n\nThe interview is kept at `redacao/entrevistas-ao-editor/-.md` — moving to\n`newsroom/editor-interviews/` when the folder migration in `dados/en-migration.json` can proceed —\nexactly as it came back, with the editor's corrections. It is the provenance of the direction: a\nmonth from now, when somebody asks why this newsroom went that way, the answer is a dated file and\nnot the memory of a session.\n\n## The reply to the editor\n\nWhen you are done, write to him — `redacao/correio/expedicao/dinis.humano/` — with a line per point\nof the report and what you did with it. Including the ones you did not do, and why. **A report that\ndisappears inside an agent teaches the editor to stop giving direction.**\n\n## And what never happens\n\nNothing said in one of these conversations puts a story into `publicado`, excuses a gate, widens the\nfolder an agent may write in, or turns a statement by the editor into a verified fact. Direction says\n**what is worth doing**. How it is done remains what is written in `CLAUDE.md`, and that is not\nnegotiated by voice.\n" } ] }