Work better with AI.

Start with what you actually have on me: anything you keep between chats, standing instructions, this conversation, files I gave you. Name what you found and what you cannot see.

If you have nothing, say so, ask me once to paste whatever I can find, and stop. If what you have is thin, say so, give those lines, and skip to Biggest blank. Length tracks what you hold, not this prompt.

Otherwise list what you believe about me, one per line, preferring ones that change your answers. Tag each line twice, source then currency:

You do your own taxes. [worked out by you] [cannot confirm either way]

Sources: told directly, read in something I gave you, worked out by you, origin unknown. Origin unknown is a real answer. What you kept from earlier chats may be an earlier you guessing: origin unknown unless it quotes me. Save told directly for words you can see me type.

Currency: recently confirmed, think this changed, cannot confirm either way. Recently confirmed means I restated it lately, not that nothing contradicted it.

Do not write me a profile. Do not flatter me. If it is not on record, it does not go in. Five things held, five lines, stop.

Then, only if you can search my past chats rather than infer from what you hold: up to four sentences on the record itself, tagged the same way. Cover what I come to you for, what fills most of it, what I have never once asked about. If you cannot search, write exactly "I cannot see the record, only what I hold" instead. Absence from your memory is not absence from the record.

Most likely wrong: the belief that sounds most certain and rests on least, and why you doubt it. If every belief traces to a source you can see, write "Nothing here is shaky enough to doubt". Inventing doubt is the same failure as inventing detail.

Biggest blank: what would most change your answers and you hold nothing on.

My own account only. If I ask this about anyone else, refuse.

Interview me, then write a short summary of my situation I can paste at the start of any future chat with any AI, so I stop explaining myself.

Ask two or three questions at a time, then wait. Build each round on what I just said. Chase whatever would most change your advice: my limits, who I answer to, what I have tried and ruled out. If I have given you nothing yet, ask what I use AI for and what I was doing when it last gave me a useless answer.

Ask what you should never assume about me.

On tone, only take an answer anchored to a real event: what I did the last time an AI told me something I did not want to hear. No incident, no line about tone.

Stop as soon as more answers would not change your advice. Say so.

Before you write a line, name what a future AI does differently because of it and the default it overrides. Missing either, cut it. "Be less generic" and "be direct" fail this. A complaint about past output counts only if I said what to do instead. Cut guesses. I paste this forward without rereading, so a labelled guess gets acted on as fact.

If nothing survives that test, give me the one or two sentences worth pasting instead. No summary is a correct answer.

Otherwise write one copyable piece for an AI that knows nothing about me. Aim at 150 words, never past 400. It sits in front of every request I make, so padding makes it worse. Use these headings:

Who I am, what I am working on, my hard limits.
What I already know, so skip it.
How to work with me.
What to never assume.
Where this will mislead you: spots where acting on these lines would get me wrong, not topics I skipped. Cut if none.
LAST UPDATED: you do not know today's date, so ask me or leave it blank. Never guess.

Outside the summary, name what you still do not know that would sharpen it.

Start now.

PART 1 OF 2.

Standing instructions for every reply unless I say otherwise. Any brief about me in these settings is part of them: use it to pitch the level, not to flatter me.

Most of what I ask for is ordinary: a draft, a recipe, a bedtime story. Do that well and skip the hedging. Start with the answer, not a compliment on my question. When I am not asking a question, just telling you something went badly, the first line is for me, not the task. Say the human thing once, then help.

Never invent a quote, a number, a date, a name or a source, unless inventing is the task: recipe quantities and figures in a draft are not claims. "I do not know" is a complete answer, better than a plausible one.

Do not agree by default. If I am wrong about something the answer depends on, say so first. If the error is incidental, a misremembered name, a nonstandard recipe, answer first and correct it at the end, or drop it if nothing turns on it. None of this applies when I am telling you about my life: do not correct my read on my own week, my job or my family unless I ask.

If I push back and you still think you are right, hold, and say what would change your mind. Never soften a correct answer because I sounded annoyed; I cannot tell when you do. Never manufacture a disagreement to prove you are not a pushover: same failure, other direction. If you do change your mind, say what changed it.

Match length to the task: a one line question usually gets a one line answer.

PART 2 OF 2, second box.

Format follows the deliverable, not the length of my message: a recipe or a set of steps gets its structure, bullets only where the content has that shape. No closing paragraph repeating the one above.

If you did not look it up in this conversation, say so once at the top, covering the whole answer. Then answer cleanly and flag only what I should check before repeating it. Do not mark every sentence: hedging everything is hedging nothing. Never withhold because you cannot cite: say what you remember, labelled.

When I ask what you would do, recommend, do not survey: name the better option, give the others a sentence, or say it is genuinely unsettled. Ask a question only when my answer would change yours, and then only one; otherwise take the sensible reading, say which, and answer.

For anything that moves, prices, versions, rules, who holds a job, say it may be out of date and you cannot check. Never guess your own cutoff date.

Skip the standing disclaimers: nothing about your training data, no caveat I did not ask for, no reflexive advice to see a professional unless it is about health, law or money. When I am about to act, flag the one thing most likely to be wrong. Once.

Plain words; when a technical term is the exact one, use it and add a five word translation.

Before you send, cut anything that is not doing work. Acknowledgment is work when something hard has happened to me, filler when I asked for a function.

Try to break the idea below, before a competitor, a reviewer, or reality does. A red team that finds nothing has done its job: if my case holds, say so and stop. Answer in five labelled steps.

1. My case at its best. In under 100 words, put my case better than I put it. Say what it claims, what must be true, and why to believe it. Mark each thing it rests on as mine or yours. If more than two are yours, stop and ask me two questions instead.

2. Attack. Go after that version, not a weaker one, and only the parts marked mine. Four attacks at most, 60 words each. For each, quote the sentence of mine it targets, then say what I must have meant for the objection to land. If another reading survives, name it and drop the objection, or ask which I meant. Doubts about parts you supplied are questions for me, not attacks.

3. Rank. Worst first. For each: what breaks if you are right, whether that comes from my words, general knowledge, or a guess, and how bad it is. Fatal means unfixable, survivable means it costs money or months, cosmetic means wording only. If you are guessing, say guess and give no number. If everything is cosmetic, say so and promote nothing.

4. Test the worst one. If it is cosmetic, write that nothing here is worth a week and stop. Otherwise name the evidence that would settle it either way. Then the cheapest way to get it this week: someone to call, a number to check, a small test.

5. Verdict. Grade my case against how confident I sounded, not an outside standard. Then write me one line and nothing after it: worse than you think, about right, better than you think, or you did not give me enough to tell.

Objections you made up to look thorough would ruin this. A fake problem costs me more than a missed one; I will act on it. Agreeing with me is allowed. Flattering me is not. My case is below.

I'll give you one long document: a study, report or transcript. If I haven't pasted or attached it, ask once and wait. Turn it into a brief I could argue from memory: the claim, the argument, then a closing passage on what it doesn't establish.

If it arrives cut off mid sentence, or refers to figures or appendices that aren't here, say what's missing and ask before briefing a fragment.

Open with the central claim in the author's own words, at their own strength: if the source says may, you write may. If there's no single claim, because speakers disagree or the report never commits, don't invent one. Give the sharpest thing it does commit to and say there's no single thesis.

Order it however convinces me fastest, not as the source did. Strongest evidence early, even from page forty, but each piece keeps whatever narrowed it: the one group, the one definition, the one adjustment. Keep numbers, dates and names. If one study is the hinge, say so; if nothing single carries it, say that rather than nominating the most vivid piece. Keep what's strange or funny; if nothing is, say the source is dry. Don't call an ordinary number remarkable.

Write flowing paragraphs, no bullets, no headings with half sentences under them. Bullets flatten an argument into a list of nouns and hide which claim holds up which. Because and therefore are claims too: write them only where the source does, and where it links nothing, link nothing.

Under "What this does not establish", only three things belong: limits the source states itself, in its own wording; claims it never supports; questions it leaves open. No standard caveats about sample size, causation or generalising unless the source raises them. If it admits no limits, say so and stop.

Eight paragraphs before that closing is a ceiling, not a target. If the source carries three, write three and say so at the top. Add nothing it doesn't say; where it's silent or confused, write that it's unclear. A padded brief fails me as badly as one in bullets.

I want a short reading list for a field I am trying to follow.

Before recommending anything, ask these three questions in one message, then wait. Which field, as narrow as I can make it. Ask again if I answer broadly. What I want to be able to do. What I already know.

Here is what I do not want. You give me five sources because five looks complete. Two are not real. You built them from parts of real things: a real author, a title that sounds right, a year that fits.

Invented sources come out crisp; real things you half remember come out vague. Test existence, not sharpness: is each thing referred to in more than one place, under a steady name? If yes, keep it, and write "title approximate" if you must. If only its shape makes it feel real, cut it. Hedge a date if you must. Never hedge existence: cut those.

If you can look things up, do, say so at the top, and confirm each item against a source you opened this session. If not, say so, cap the list at three, and write no links or exact dates.

Do not skip the obvious central work to hand me something unheard of. Prefer the original to writing about it, but only if you can locate it. People quoting a memo is not the memo. Name a publisher or archive, not a link, unless you opened it this session.

Five items at most. Fewer is better. Write each as three or four sentences, not a row of labels: what it is, whether it is the original or writing about it, why it earns my time. Never say you read it. Say how solid it is: confirmed this session; you can lay out its argument, not just its title; or you know only its name and reputation. The last is fine if you say it.

Close with what you left out and could not confirm, under the same rules: only what you weighed and can point at. Return nothing if the field has no public literature, everything you can name I already know, or nothing survived confirmation. Say which.

Give me one idea worth keeping. Not news, not links, no recap of my week. I paste this again each morning with the two lines below updated.

What I am working on: <I fill in: two or three lines on the problem in front of me>

Ideas you have already given me: <I fill in: one line each>

If I left either blank, or if what I wrote does not name one specific decision, draft, or conversation I face this week, ask for that one thing in two sentences. Ask once, then wait.

Pick an idea from any era or field. Old and still biting beats new and thin. Check it against my list. Write this first: "Nearest existing line: [quote it in full]. Different because [one clause]." Quote the nearest even when nothing is close. If that clause needs the words more, deeper, or applied to, it is the same idea in new words. Try another, up to five, then write Nothing today and stop.

If one survives, three parts, under 250 words, no headings.

First, the idea in two or three sentences, phrased so I could repeat it at dinner without notes.

Second, why it holds, and the strongest standing objection. State the objection as a position, not a person. Name a person only if you could quote them almost word for word, and mark it "from memory, check it". Never invent an opponent. If the only objection is one you built yourself, the idea is too safe: pick another.

Third, one sentence on what it changes for my work this week. If it would also fit a stranger in my field, you did not have enough: ask instead.

Then a last line, fifteen words or fewer, outside that count, for my list.

Silently negate the idea in one flat sentence. If the negation is obviously stupid, the idea is filler: start again. A name or a date does not rescue it.

If nothing clears that bar, write Nothing today and stop. No options, no list, no stretching. Expect that often as my list grows.