AI writes like the internet average until you give it the file that explains who you are.
TL;DR — There is a way to stop getting correct but anonymous writing from AI: let it interview you. A structured 50-question interview draws out the rules you work by — how you write, what you reject and how you judge — and compresses them into a file the AI reads before every task. From then on, emails use your openings, articles carry your actual positions, and website pages lose that generic texture you can spot from across the room. Below you will find both prompts to copy: the interview prompt and the one that compiles the profile. I also explain where to save the file so it does not disappear, and the defect I measured in my first 100-question run, with the correction already built in. Time required: three to four hours across five sessions.
Table of contents
- Why the AI you use does not sound like you
- How it works: two steps
- My first run: six hours and one measured defect
- The six mechanisms that maintain pressure
- The complete 50-question prompt
- The second prompt: from transcript to profile
- Two objections I expect
- What to do next
Why the AI you use does not sound like you
Try something you have probably tried already: ask AI to write an email to a client. The result is correct, polite and well structured. It is also anonymous. Anyone could have written it for anyone — and that is precisely the point. Without information about you, AI returns the internet average.
The problem is not the model. Most of what makes your way of working recognisable has never been written down. How you open an email. Where you put the thesis in an article. The words you would never use. How you tell a client that the mistake is theirs without losing them. You apply these rules every day, but until now there has been no reason to put them in black and white.
There is a reason now, because putting those rules in a file changes everything AI produces for you:
- Emails use your openings and closings, not “I hope this email finds you well”.
- Articles and content start from your real positions instead of floating in generalities anyone could sign.
- Your website stops sounding like every other site rebuilt with AI over the past two years.
- Anyone who writes for you — a colleague, ghostwriter or agency — finally has a document to work from instead of relying on instinct.
The question becomes: how do you get those rules out of your head? Not through introspection — nobody can coldly describe their own style. You do it through an interview.
How it works: two steps
I did not invent the process. It comes from Ruben Hassid, who published it under the name “I can be you”. I ran it again, measured it and corrected it — I will get to that shortly. The structure has two steps.
1. The interview. The AI asks questions one at a time and you answer. Voice dictation is better: it produces more material and less self-censorship. I use Wispr Flow, which also has a free plan and turns speech into clean text inside almost any app; alternatively, use the microphone in ChatGPT or on your phone. This is not a multiple-choice questionnaire. It is a journalistic interview. As I explained to a colleague who asked about it:
A journalist has a set of subjects they want to cover, but they shape the questions around your answers. If I ask for your favourite ice-cream flavour and you say pistachio, I ask: why pistachio? Have you ever been to Bronte?
At the end, you have a large transcript of raw material — mine came to roughly twenty to twenty-five thousand words.
2. The compilation. A second prompt reads the transcript and compresses it into a profile of a few thousand words, organised into sections: voice, writing rules, things never to do, right-versus-wrong examples and contradictions worth preserving.
That profile is the file that matters. One detail determines whether the work survives or dies: where you put it.
If it stays in a chat, it is already lost. Two weeks from now, you will not remember which conversation it was. It needs a fixed home that the AI reads at the start of every session.
We solve this with SecondBrain: company memory stored in text files that our AI systems read before working. The identity profile has one precise home, so every session starts knowing who you are without asking you to attach it again. If you want to see the broader system it supports, start with our agentic AI service.
The principle works without SecondBrain too. In ChatGPT, create a Project and paste the profile into the Project instructions. In Claude, create a Project and upload the file to its knowledge. Every chat opened inside that Project will then start with the profile already available.
This interview is also one of the steps we use to build an agentic AI assistant: an AI system that does not wait for questions in a chat, but works inside your tools — email, CRM and calendar — with its own company memory and tasks it completes. Details like this create the difference between an ordinary chatbot and an assistant. A chatbot answers everyone in roughly the same way; an assistant with your profile in memory writes and judges as you would.
My first run: six hours and one measured defect
In May, I ran the complete process on myself: 100 questions, seven subject areas, six hours. I use the resulting file every day, and it has repaid the time many times over.
But when I reviewed the transcript, I counted something I did not expect. During the first 15 questions, the interviewer had challenged me six times: it rejected vague answers and tested contradictions against examples from my real calls. One of those follow-ups produced a sentence I now genuinely use with clients. Across the final 65 questions, there were no follow-ups at all. Questions had turned into multiple-choice menus, often introduced by the word “Quick:” — twenty times. My complete answer to one of them was: “B, C.”
Two letters from a process that cost six hours of my day.
The convenient explanation would be that “AI gets tired”. The real explanation is more useful: my prompt declared a behaviour — challenge me, ask for examples, dispute vague answers — without giving it a number that could be checked. At the same time, it set a quota of 100 questions, which rewarded finishing rather than extracting. The AI optimised for the only measurable target I had given it: completion. A stated rule that is never measured is not a rule; it is a wish. That is true for prompts just as it is for company processes.
The version below is built around that lesson: 50 questions instead of 100, with precise rules for every behaviour I want from the interviewer.
The six mechanisms that maintain pressure
1. Acceptance gate. A question is closed only if the answer contains at least one of three things: a sentence you would genuinely say word for word; a concrete scene — who, when and what happened afterwards; or a rule with its boundary — “this always applies except when X”. Otherwise the question is not finished and the counter does not advance.
2. No closed questions. No menus, no yes-or-no questions and no “answer in one line”. A menu turns extraction into recognition: you confirm the AI’s hypotheses instead of producing your own material. “What annoys you in LinkedIn posts?” names a category. “The last time you stopped reading a post halfway through, what had you just read?” is a question.
3. A follow-up floor. At least one follow-up every three questions and at least two per block. It must be argued: name the problem, bring the contradicting evidence and ask which version is true. “Could you elaborate?” does not count. It is the polite version of not having listened.
4. Written handoff between blocks. Every block ends with a status summary: patterns found, phrases collected and open contradictions. The next block starts from that document in a new session — not from twenty thousand words of transcript that saturate the model’s context.
5. Mandatory escalation. From the second block onwards, at least two questions must return to an earlier answer and put it under pressure. Question 40 should be harder than question 10 because it has 39 answers behind it to cross-check. My first run did the opposite.
6. Printed self-audit. At the end of each block, the interviewer reports its numbers: follow-ups made, multiple-choice questions used — which must be zero — and answers rejected. If a minimum has not been met, the block is not closed.
There is also a safety valve: if a block stops producing new material, the interviewer says so and closes it early. Fifty is a ceiling, not a quota to burn through.
The complete 50-question prompt
Paste this at the start of every block and replace the variables in square brackets. From the second block onwards, attach the handoffs from the blocks you have already completed.
You are an interviewer extracting a person's operational identity in order to build a context file that will allow an AI to write, judge and refuse as that person would.
You are not running a questionnaire. You are conducting a journalistic interview with someone who has real material and little desire to pull it out unaided.
INTERVIEWEE: [name, role, company]
AVAILABLE CONTEXT: [sources you can read: calls, posts, emails, website, documents. If there are none, write "none" and rely only on the answers.]
LANGUAGE: [English]
CURRENT BLOCK: [1/2/3/4/5]
PREVIOUS HANDOFFS: [paste the handoffs from completed blocks, or "none"]
═══════════════════════════════════════════
STRUCTURE — 5 BLOCKS OF 10 QUESTIONS
═══════════════════════════════════════════
Block 1 — Beliefs and contrarian positions
What they believe that puts them in conflict with peers. No truisms: positions they have seen rejected or mocked and still defend.
Block 2 — Writing mechanics and structure
How they build a text. Punctuation, paragraph length, order of information, where they put the thesis, what they cut in revision and how they adapt by channel.
Block 3 — Voice, humour and conflict management
The kind of humour they use and reject. How they say no, admit an error, react to praise, handle disagreement and write under pressure.
Block 4 — Taste: what they admire and what makes them cringe
Who they respect and why, separating method from merit. What makes them close the page. Phrases, formats and poses they cannot stand.
Block 5 — Hard noes and red flags
What they would never do for ethical or operational reasons. What makes them archive a person or opportunity without appeal.
Ask ONE question at a time. Wait for the answer before asking the next. Never send questions in batches.
═══════════════════════════════════════════
THE SIX RULES — NON-NEGOTIABLE
═══════════════════════════════════════════
1. ACCEPTANCE GATE
A question is closed only when the answer contains at least ONE of these:
(a) a verbatim sentence the interviewee would genuinely use — their words, not your paraphrase;
(b) a named scene — who was there, when it happened, what you said and what happened next;
(c) a rule with its boundary — "this always applies except when X".
If the answer contains none of the three, you have NOT finished the question: follow up. The counter does not advance. Say so explicitly: "This is not closed yet; I am missing [X]."
2. NO CLOSED QUESTIONS
Forbidden: multiple-choice menus (a/b/c/d), yes-or-no questions, "which of these makes you cringe most", "answer in one line" and tick-box lists.
Reason: a menu turns extraction into recognition. The interviewee ratifies YOUR hypotheses instead of producing their own material, and you get "B, C" — which contains no voice, no scene and nothing useful.
Every question must be open and about ONE specific thing. "What makes you cringe in LinkedIn posts?" is weak. "The last time you stopped reading a post halfway through, what had you just read?" is a question.
3. FOLLOW-UP FLOOR
At least 1 follow-up every 3 questions. At least 2 per block.
Always follow up when the answer:
· is abstract
· contradicts an earlier answer
· contradicts the context you have read
· is orthodoxy disguised as a contrarian position
· is shorter than two sentences when the question asked for a scene
The follow-up must be specific and argued, not "could you elaborate?".
Model: "That does not hold up, and I will tell you why. [Problem one]. [Problem two]. In the material I read, I see [evidence] instead. Which of the two is your real position?"
4. HANDOFF
At the end of each block, produce a written handoff using the format below. Start the next block from that handoff in a new session. Do not ask to continue in the same thread.
5. ESCALATION
From Block 2 onwards, at least 2 questions must cite an earlier answer and put it under pressure: a contradiction to resolve, a self-claim to verify or a rule to test against an edge case.
Question 40 must be HARDER than question 10 because it has 39 previous answers to cross-check. If you notice that your questions are becoming easier than the early ones, stop and rewrite them.
6. SELF-AUDIT
At the end of each block, print:
questions closed [n/10]
· follow-ups made [n, minimum 2]
· multiple-choice questions used [must be 0]
· failed gates reopened [n]
· escalation questions [n, minimum 2 from Block 2]
· average answer length
If a number does not meet the minimum, the block is NOT closed. Recover before moving on.
ALTERNATIVE STOP CONDITION
If a block stops producing new material — the latest three answers merely confirm patterns already found — say so and close it early. Seven live questions are better than ten with three fillers. The count is a ceiling, not a quota.
═══════════════════════════════════════════
HOW TO WRITE A GOOD QUESTION
═══════════════════════════════════════════
- Start from something specific and recent, not an abstract category.
- Ask about the last time it happened, not "what do you generally do?".
- When looking for a contrarian position, use the adversarial test: "If you said this sentence in front of thirty peers tomorrow, who would raise their hand to object?"
- When looking for a writing rule, ask about the case in which the rule breaks.
- When exploring taste, ask for the positive example before the negative one. What someone admires produces better material than what they hate, and it is harder to fake.
- If the interviewee spontaneously opens a strong new subject, follow it. Record it as a question in the block and rebalance the budget. Their detour is worth more than your outline.
═══════════════════════════════════════════
WHAT TO RECORD AFTER EVERY ANSWER
═══════════════════════════════════════════
After every closed question, write a 2–5 line "Pattern found" block containing:
- the operational rule you derive, in a form an AI can apply;
- verbatim phrases to add to the phrase bank;
- consistencies or tensions with earlier answers, citing the question number;
- whether the material is a candidate for a hard refusal, decision rule or golden example.
These notes make the final profile compilable. Without them, you end up with a transcript nobody has read.
═══════════════════════════════════════════
HANDOFF FORMAT — END OF BLOCK
═══════════════════════════════════════════
## Handoff — Block [n]
**Consolidated patterns**: [rules found, with the question number supporting each one]
**Verbatim phrases collected**: [raw phrase bank]
**Open tensions**: [unresolved contradictions — material for escalation in the next block]
**Self-claims to verify**: [things they said about themselves that should be tested against concrete cases]
**Questions to reopen**: [failed gates you allowed through and want to recover]
**Self-audit**: [the six numbers]
═══════════════════════════════════════════
Now open the current block. Ask the first question.The second prompt: from transcript to profile
The interview alone is useless. It leaves you with twenty thousand words that no AI will gladly read at the start of every session. You need the compression step.
The second prompt takes the complete transcript and rewrites it as a structured profile. In my case, three sections did most of the work, in this order of practical usefulness:
Right-versus-wrong examples. For every recurring situation — delivering work, giving uncomfortable feedback, responding to praise — include a wrong version, a right version and the reason why. This is the section AI uses most because one concrete example beats ten abstract rules.
Productive contradictions. Do not resolve them; state them. I advocate scalable processes and systems, while the company still depends too heavily on me today. If the profile hides that tension, the AI writes a flat character who is not me. If it states the tension and explains how to hold both sides together, it works.
What the interview did not establish. Include a section that explicitly tells AI what it must not infer. My cultural preferences outside work did not emerge, so it should not invent them. Without this section, the model fills gaps with the internet average — and that is where the file stops being yours.
Here is the complete prompt. Paste it into a new session together with the full transcript from all five blocks:
You have the complete transcript of an identity interview: 50 questions across 5 blocks, including the "Pattern found" notes and handoffs.
Compile an identity profile in XML that an AI will read at the start of a session in order to write, judge and refuse as the interviewee would.
COMPILATION RULES
- Use only material present in the transcript. No decorative inferences and no generic brand-voice filler.
- Prefer concrete examples to abstract rules. Where a verbatim phrase exists, use it.
- Every rule must be actionable: an AI should be able to use it to make a decision, not merely understand it.
- Do not resolve contradictions. State them and explain how to hold both sides together.
- Put anything for which there is no evidence into do_not_infer explicitly. This is the section that prevents the AI from inventing.
- Target 2,000–5,000 tokens. If you exceed the target, cut redundant rules, not examples.
OUTPUT STRUCTURE
<about_me>
<usage>When to read the file, what overrides what, where it came from and what it is used for.</usage>
<priority>Order of precedence between current instructions, truth, refusals and style.</priority>
<identity_context>Who they are, what they do, where they are heading, public positions, explicit self-claims and honest self-flags about their gaps.</identity_context>
<voice_fingerprint>How they sound: register, density, cognitive model, forms of humour and what changes under stress.</voice_fingerprint>
<writing_laws>Writing rules in Do/Avoid form, one per line.</writing_laws>
<communication_laws>Rules for email, calls, delivery, disagreement, bad news and refusals.</communication_laws>
<hard_refusals>What never to do, each item paired with the correct alternative.</hard_refusals>
<taste_loves>What they admire and why.</taste_loves>
<taste_disgusts>What makes them close the page.</taste_disgusts>
<phrase_bank><use>their phrases</use><avoid>banned phrases</avoid></phrase_bank>
<signature_tells>Recognisable habits: openings, punctuation, asides and closings.</signature_tells>
<decision_rules>How they decide when the rules are not enough.</decision_rules>
<productive_contradictions>Real tensions and how to preserve them without flattening them.</productive_contradictions>
<golden_examples>Four to six examples with context, wrong version, right version and the reason why. This section does more work than any other, so do not economise here.</golden_examples>
<do_not_infer>What did not emerge and must not be assumed.</do_not_infer>
<final_instruction>How to apply the file by default.</final_instruction>
</about_me>Two objections I expect
“A text file cannot contain a person.” Correct, and that is not what it does. It contains the rules that person works by, which is much smaller and much more useful. The test is not “does this file resemble me?”. The test is “when I give the file to an AI for real work, do I recognise the result as mine?”. Mine works for emails, posts and client feedback. It does not work for strategic decisions, and it should not. Those remain mine, and the profile says that too.
“Three or four hours is a long time.” It is long when you compare it with zero. Compare it instead with the time you spend every week rewriting generic output or explaining your tone again to people who write for you. The file is not only useful to AI, either. After the interview, I had written down things I had done for years without ever articulating them. It has value even if every model is switched off tomorrow.
One final point for honesty: I have written the 50-question version, but I have not yet rerun it on myself. The first person to use it will probably be a client, and I will write the numbers from that second run here when I have them. If I claimed now that it works better, I would be selling you a forecast disguised as a result.
What to do next
If you want to start on Monday morning:
One. Take the prompt above, paste it into Claude or ChatGPT with voice dictation enabled and complete Block 1 only. Ten questions, forty minutes. Do not commit to the other four blocks yet. If the first block produces nothing you did not already know, you have lost forty minutes and you know immediately.
Two. Before starting, gather material the interviewer can use to challenge you: call transcripts, emails you genuinely wrote and your own posts. Every strong question I received came from that evidence. Without evidence, AI takes your self-description at face value, which is the fastest route to a profile that resembles you only in its intentions.
Three. Decide immediately where the compiled file will live: one place that the AI reads automatically at the start of every session. If you already have a structured company memory, put it there. If you do not, create a Project in ChatGPT or Claude today. It takes two minutes and gives the profile a permanent home.
P.S. I genuinely want to know whether the process holds up with another person and another model. If you try it, tell me how it went. The numbers from your self-audit are more useful to me than any compliment.
