Work down the steps in order, or jump to the one you need. Each step has a prompt that runs the whole thing, and a prompt for each method inside it. Every prompt works from what is already in your chat and asks only for what is missing, so you can start a task with it or drop it into one under way.
Turn a topic into a question you can actually research, before any searching starts. An AI will not question a badly framed topic. A researcher reframes it first.
Work through the framing of this research with me before searching. Use the topic we have been discussing, or ask me for it if there is none yet. 1. Restate it as the decision it informs, the questions needed to make that decision, and the evidence that would answer them. 2. Tell me whether the question as posed is the right one, and whether a nearby question would be more worth researching. 3. Once we agree the question, mark which parts are background and which the conclusion will depend on, so we know where to go deep. Wait for me to confirm the framing before moving on.
Help me turn what I am researching into a properly researchable question. If we have already been discussing a topic in this chat, work from that. If not, ask me for the topic. Restate it as three things: 1. The decision this research is meant to inform. 2. The questions that would have to be answered to make that decision. 3. The kind of evidence that would answer each of those questions. If it cannot be restated this way, tell me why and suggest a version that can. Also tell me if there is a nearby question that looks more worth researching than the one on the table, and why. Wait for me to confirm the question before searching.
Collect material and work out what can be trusted, one item at a time. This is where an AI on its own falls furthest short: it reads the top few results and stops.
Research this question, working to these rules the whole way through. Use the question we have settled on, or ask me for it. - As you read, flag leads worth following: named people, referenced but unlinked documents, conflicting numbers, and things that should be present and are missing. - Treat weak sources (marketing, social posts, unattributed articles) as leads only. Take the checkable detail out and confirm it against a primary document, official body, peer-reviewed work, or named specialist reporting. - Record every claim that matters in a table: the claim, whether the conclusion depends on it, its status, its best source and what kind it is, and what you searched. - Do not rest anything important on a single source when a second could be found. After three failed searches for a detail, mark it unconfirmed and move on. Show me the table as you build it, and tell me what you could not confirm.
Look at the material for leads, the way a researcher would rather than a summariser. Use whatever material is already in this chat. If there is none, or you want me to point at something specific, ask. Point out every detail worth following up, including: - a named person, especially one moving between roles or organisations - a document, filing or study referred to but not linked or quoted - a number that conflicts with another one here or with something well known - anything that should be present given the subject and is missing For each lead, say what it is and what following it up might tell me. Do not summarise. I only want the leads.
There is a detail here that needs verifying against a better source. If it is clear from our chat which detail I mean, work with that. If not, ask me which one, and where I saw it. Treat the current source as a lead, not as evidence. Find a source that can actually support the detail: a primary document, an official body, a peer-reviewed source, or named specialist reporting. Search for the detail itself, not the general topic. Tell me one of three things: - Confirmed, with the source and what kind of source it is. - Contradicted, with the source. - Could not confirm it after a genuine attempt. Do not settle for another version of the same weak source repeating the claim.
Keep a running record of every claim that matters to the conclusion. Start from the claims already raised in this chat, then keep it updated as we go. Present it as a table with these columns: - Claim - Does the conclusion depend on it? (yes / no) - Status (confirmed / contradicted / unconfirmed) - Best source so far, and what kind of source it is - What was searched When we reach conclusions, state as fact only the claims marked confirmed. Anything unconfirmed must be flagged as such in the write-up, not stated plainly and not left out.
Produce findings that are not sitting in any single source. The core move is combination: putting two things side by side and reading what comes out.
Help me analyse the material we have, rather than summarise it. I want findings that are not obvious from any single source. Use the material already in this chat, or ask me where it is. - Where two things can be set against each other (a policy against a lifecycle, one option against another), lay them out and read the shape of the result: where things spread, bunch, crowd or leave gaps, and what that implies. - Where the two are peers, tell me where they differ: what each treats as central, where one is specific and the other vague. - If a comparison does not hold together, say so and suggest a better one rather than forcing it. For each finding, tell me what in the material supports it.
Put two things side by side and tell me what the comparison reveals. If it is clear from our chat what the two things are, use them. If only one is clear, suggest a good thing to set it against. If neither is, ask me. Lay them out in a table. Then, and this is the important part, read the shape of the result rather than describing each cell. Tell me: - where the elements spread out evenly, and where they bunch together - where one side is crowded and the other is empty - what those patterns imply If most of one will not line up against the other, do not force it. Tell me they do not fit well together, and suggest a better pairing.
Test the finding and put it in a form you can trust. Search for evidence against it, state what the checked material supports, and show what it rests on.
We are ready to conclude. Work from the research in this chat, and work through these in order. 1. State what the verified material actually supports. Not the direction most of the material seems to point, but what has actually been confirmed. 2. Try to disprove that conclusion. Search for evidence against it, worded to find the opposite. Address the strongest objection you find. 3. Write the conclusion so it names the specific claims it rests on, each confirmed. 4. Check the result: every fact traceable to a good source, anything unconfirmed marked as such, the contrary case addressed. Show me the conclusion and the claims it depends on.
Take the conclusion we have reached and try to disprove it. Use the conclusion from our chat. If it is not clear which one I mean, ask. Search for evidence that it is wrong, not more evidence that it is right. Word the searches to find the opposite: if the conclusion is that something works, search for reports of it failing, criticism of it, or cases where it was overstated. Then take the strongest thing you find and address it directly. If it changes the conclusion, say how. If it does not, explain why. Do not give me a vague "there are other views" hedge. I want the strongest actual objection and a real response to it.
Check the draft against these standards and report every place it falls short. Use the draft already in this chat. If there is none, ask me for it. Do not rewrite it; just tell me what fails and where. - Every stated fact traces to a source good enough to support it (a primary document, official body, peer-reviewed work, or named specialist reporting). - Anything not confirmed is marked as unconfirmed, not stated as fact and not quietly left out. - The conclusion names the specific claims it rests on, and each is confirmed. - The strongest case against the conclusion is addressed, not avoided. For each failure, quote the sentence and say which standard it misses.
Want all of this on every task without invoking it each time? The full standing instruction carries the four steps and these standards into a project or agent.