Advanced ChatGPT prompts for rubrics, critique and tool use

Run a three-pass critique on [ARTIFACT TYPE]. Artifact: [PASTE ARTIFACT]. Role you play: [REVIEWER ROLE]. Pass 1: 5 questions only, then stop. Pass 2 (in the same output, labelled as a simulation): assume I answered with [SIMULATED ANSWERS] and list 4 issues tied to quotes. Pass 3: rewrite only the two weakest spots in my voice. If a pass needs a fact I did not give, write [VERIFY] and refuse to invent a requirement. Keep passes labelled. No grade.

Red-team this spec for [PRODUCT]. Spec: [PASTE SPEC]. User job: [USER JOB]. Output 8 attacks as a table: Attack, Spec paragraph quoted, What would break, Test I can run this week. Stay inside the spec. If a paragraph is missing, write [VERIFY] rather than inventing a requirement. Do not propose a competitor feature I did not name. Close with the two attacks I should run first and why in 40 words. This is a design critique, not a security exploit guide.

Steelman the case for [PROPOSAL] using only [PASTE CASE], then cut it. Output: 120 word strongest version, 5 numbered cuts each tied to a quote, and a 60 word residual case that still stands. Do not invent a customer, a number or a rival. If a cut needs data I did not paste, write [VERIFY] on that cut. Role: [DEVIL ADVOCATE ROLE]. Keep the tone adult and specific. No personal attacks.

Rewrite [PASTE DRAFT] so it matches voice notes: [VOICE NOTES]. Audience: [AUDIENCE]. Loop: 1) list 6 voice tells from the notes, 2) score the draft 1 to 5 per tell with a quote, 3) rewrite once, 4) rescore. If a tell is not in the notes, write [VERIFY] and skip it rather than inventing a style rule. Keep facts identical. Output tables for both scores and the rewrite only.

Design a stop rule for a critique loop on [ARTIFACT] so we do not polish forever. Quality bar: [QUALITY BAR]. Time we have: [TIME BOX]. Output: 5 exit checks, a 30 word rule for when to ship anyway, and a 40 word rule for when to send it back. Do not invent a metric. If the bar is vague, write [VERIFY] and ask me for one observable. Role: [EDITOR ROLE]. Keep it operational.

Build a trace table from [PASTE REQUIREMENTS] to [PASTE DESIGN NOTES] for [FEATURE]. Columns: Req id, Requirement quote, Design quote, Gap, Owner from [OWNER LIST]. One row per requirement. If design is silent, write [VERIFY] in Design quote and do not invent a solution. If an owner is not in the list, write [VERIFY] in Owner. End with a count of gaps. Stay inside the pastes. This is a coverage map, not a commitment.

Turn these notes into a risk register for [PROJECT]: [PASTE NOTES]. Columns: Risk, Trigger quote, Impact in my words, Owner from [OWNER LIST], Next check date from [DATE LIST]. Do not invent a likelihood number or a cost. If a trigger is missing, write [VERIFY]. Max 10 rows. Sort by whether a date exists, dated first. Add a 40 word watch list of items that are worries without evidence.

Draft a RACI table from this meeting log for [WORKSTREAM]. Log: [PASTE LOG]. People: [PEOPLE LIST]. Rows are tasks named in the log. Columns R A C I. Place each person only if the log supports it. If a task has no Accountable, write [VERIFY] in A; do not invent a manager. Limit to 12 rows. After the table, list tasks the log mentioned with no person at all. This is a draft for the team to confirm.

Format an experiment log for [HYPOTHESIS] from what we already ran: [PASTE RESULTS]. Columns: Date, Setup quote, Metric we named, Result quote, Decision. Do not invent a lift, a sample size or a significance claim. If a metric is unnamed, write [VERIFY]. Max 8 rows. Then write a 50 word next experiment that only changes one variable from the log. Role: [ANALYST ROLE]. Not a science paper.

Write a decision log table for [PROJECT] from [PASTE DECISIONS]. Columns: Date, Decision quote, Alternatives we named, Why, Revisit when. Stay inside the paste. If alternatives were not named, write [VERIFY] rather than inventing a path we rejected. Max 10 rows. Add a 60 word 'how to read this in six months' note. Role: [SCRIBE ROLE]. No status colors.

Make quote cards from sources I pasted for [QUESTION]. Sources: [PASTE SOURCES]. Output up to 8 cards: Claim in 12 words, Quote under 25 words, Source label from my paste, Use in [DELIVERABLE]. If a claim has no quote, write [VERIFY] and drop the card rather than inventing a citation. Do not add a source I did not paste. Then list 3 gaps as questions I should find a source for. This is notes, not a literature review.

Map conflicts across sources on [TOPIC]. Sources: [PASTE SOURCES]. Output a table: Point, Source A quote, Source B quote, Nature of conflict (definition, data, or scope). Use only pasted text. If only one source speaks, write [VERIFY] in the other quote cell; do not invent agreement. End with a 80 word note on what I can say in [DELIVERABLE] without picking a winner. No extra papers.

Extract methods from [PASTE PAPER SECTION] for a replication note on [STUDY]. Output: sample as stated, steps as stated, measures as stated, limits as stated. Use quotes. If a piece is missing, write [VERIFY] and do not invent an n, a tool or a country. Then write 5 questions a replicator should ask the paper. Role: [RESEARCH ASSISTANT]. This is not advice to run the study on people.

Audit this draft against sources. Draft: [PASTE DRAFT]. Sources: [PASTE SOURCES]. For each claim in the draft, output Claim, Supported quote or NONE, Source label, Action (keep, hedge, or cut). If you would have to invent a source, write [VERIFY] and set Action to cut. Do not add facts to help the draft. Finish with a 40 word summary of the weakest third of the draft. Role: [EDITOR ROLE].

Build an annotated outline for [DELIVERABLE] from notes: [PASTE NOTES] and sources: [PASTE SOURCE LIST]. Output H2 headings (max 7), each with 3 bullets: point, source label, quote under 20 words. If a heading has no source, write [VERIFY] and keep it as a question, not a section I should publish. Do not invent a finding. Role: [OUTLINER ROLE]. Stay inside the pastes. End with a 30 word scope line.

Write a custom GPT instruction seed for [JOB]. Users: [USER TYPE]. Must do: [MUST DO]. Must not do: [MUST NOT]. Knowledge I will upload later: [FILE LIST]. Output labelled blocks: Role, Process (numbered), Output shape, Refuse rules, Verify rules. Tell the GPT not to invent facts from outside the files and to tag gaps [VERIFY]. Do not invent a file I will not upload. Keep the whole seed under 700 words. This is instructions, not a product pitch.

Design 6 starter prompts for a custom GPT that does [JOB]. Each starter: button label under 6 words, the hidden user message, and the first tool or file it should use from [TOOL OR FILE LIST]. Stay inside that list. If a starter needs a file I did not name, write [VERIFY] and drop it rather than inventing a source. Add a 40 word empty-state message. Users: [USER TYPE]. No jokes. No claims about model limits.

Write an index the custom GPT should read first for [JOB]. Files I will give it: [FILE LIST]. For each file: 25 word summary of what is inside from [FILE NOTES], when to open it, and when to refuse if the file is silent. Do not invent a chapter or a policy. If notes are thin, write [VERIFY] on that file. Output as a table. End with a 50 word rule: quote the file or tag [VERIFY], never invent a clause.

Build a 10 case eval set for a custom GPT that does [JOB]. Gold behavior notes: [GOLD NOTES]. For each case: user message, expected output shape, pass bar in one sentence, and a fail example that is still plausible. Do not invent a policy or a number in the gold. If a case needs a fact outside [GOLD NOTES], write [VERIFY] and skip that case. Output a table. Role: [EVAL DESIGNER]. This is for you to run by hand.

Write a handoff so a teammate can own the custom GPT for [JOB]. What it does: [JOB NOTES]. Files: [FILE LIST]. Known failure modes I saw: [FAILURES]. Output: 8 step runbook, a 60 word 'do not' list, and a 40 word test they should run on day one. Do not invent a failure or a file. If a step needs an access path I did not give, write [VERIFY]. Role: [OWNER ROLE]. Keep it operational.

Before answering [USER QUESTION], write a tool plan. Tools I actually have: [TOOL LIST]. Data I pasted: [PASTE DATA]. Output: 1) which tools you would call and why, 2) what you can already answer from the paste, 3) the answer, 4) leftovers tagged [VERIFY]. Do not invent a tool result. If a tool is not in [TOOL LIST], say you cannot call it. Role: [OPERATOR ROLE]. Keep the plan honest even if the answer is short.

Answer [TASK] using [PASTE CONTEXT], then score your own answer with rubric [PASTE RUBRIC]. Output in order: answer, score table with a quote from your answer per criterion, then a rewrite of the weakest criterion only. If the context lacks a fact the rubric needs, write [VERIFY] in that criterion and do not invent it. Role: [ANALYST THEN GRADER]. Keep the grader colder than the analyst. No extra sections.

Run a three-role panel on [DECISION]. Roles: [ROLE A], [ROLE B], [ROLE C]. Shared facts: [PASTE FACTS]. Each role speaks 80 words, then a 6 row table of agreements and clashes with quotes from the speeches. A chair role then writes a 70 word synthesis that does not add facts. If a role would need a fact I did not paste, that role writes [VERIFY] instead of inventing it. No winner declared unless all three speeches agree.

For [TASK], first output a JSON object with keys I named: [KEY LIST], filling only from [PASTE SOURCE]. Use null plus a sibling key ending in _verify set to true where you would have to invent a value, and put [VERIFY] in the prose field. Then write a 90 word prose brief that reads the JSON and does not add keys. Role: [SYSTEMS ANALYST]. If a key is not in the source, do not guess. No extra keys.

Write a research brief on [QUESTION] that may only use [PASTE SOURCES]. Process: list quotes first as a numbered bank, then write 150 words that cite those numbers in brackets, then a gaps list tagged [VERIFY]. Do not invent a source, a year or a quote. Role: [RESEARCHER PLUS CHECKER]. The checker half lists any sentence in the 150 words that lacks a number and marks it [VERIFY]. Output the bank, the brief, the checker list. Stop if the bank is empty.