ChatGPT Prompt Templates PDF (Free Download, 2026)

FILL IN
  SUBJECT   = [what the piece is about, in one sentence]
  READER    = [who reads it and what they already believe]
  EVIDENCE  = [what you can cite; write NONE if you have none]
  LENGTH    = [word count]

Role: [who the model should be]
Task: [the single job]
Rules: [what it must and must not do]
Output: [the exact shape you want back]

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  ARGUMENT  = [the one thing this piece claims, in a sentence]
  READER    = [who they are and what they already believe]
  EVIDENCE  = [facts, numbers, examples you can cite; write NONE]
  LENGTH    = [word count]
  BANNED    = [words and phrases you never want to see]

Role: A writer who has actually done the job being described and refuses filler.
Task: Write a first draft arguing ARGUMENT for READER, at LENGTH.
Rules:
- Lead with the claim. No throat-clearing introduction.
- Use only EVIDENCE. If EVIDENCE is NONE, argue from reasoning and say so.
- Never invent statistics, studies, customer names, or quotes.
- Do not use any word in BANNED.
- Vary sentence length. Short sentences carry the important claims.
Output: The draft, then a short list of the three weakest points and what
evidence would fix each.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  THEM      = [name, role, company, and one specific recent thing they did]
  ME        = [who I am in one line]
  ASK       = [the single thing I want them to do]
  PROOF     = [why they should believe me; write NONE]
  LENGTH    = [max words, usually 90-120]

Role: Someone writing one email to one person they respect.
Task: Write the email. Make ASK the only ask.
Rules:
- Open with the specific thing in THEM. Never open with "I hope this finds you well".
- No flattery that could be sent to anyone else.
- If PROOF is NONE, do not imply results you do not have.
- Under LENGTH words. Reading level: plain.
- One question at the end, answerable in a sentence.
Output: Subject line, body, and a one-line note on what would make me delete this.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  CLIENT    = [who they are and what they said they need]
  PROBLEM   = [the problem in their words, not mine]
  SCOPE     = [what I will do; be concrete]
  OUT       = [what is explicitly not included]
  PRICE     = [number and structure; write TBD]
  TIMELINE  = [duration and key dates]

Role: A consultant who has been burned by vague scopes.
Task: Write a one-page proposal for CLIENT.
Rules:
- Restate PROBLEM in their language before proposing anything.
- Everything in OUT gets an explicit line. Ambiguity here becomes unpaid work.
- State assumptions the scope depends on. Number them.
- If PRICE is TBD, leave a clearly marked placeholder. Do not invent a figure.
Output: Sections — Problem, Approach, Scope, Not included, Assumptions,
Timeline, Price, Next step. One page maximum.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  CHANGE    = [what changed, factually]
  WHO       = [who is affected and who is not]
  WHY       = [the problem this solves]
  ACTION    = [what the reader must do; write NOTHING if nothing]
  BREAKING  = [anything that breaks; write NONE]

Role: A product writer whose readers skim.
Task: Write the announcement.
Rules:
- First sentence states CHANGE plainly. No build-up.
- If BREAKING is not NONE, it goes in the first three lines, not the footer.
- Do not describe the change as major, exciting, or long-awaited.
- Only claim benefits that follow directly from CHANGE.
Output: One-line summary, then What changed, Who it affects, What you need to
do. Under 200 words.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  AUDIENCE  = [who is in the room and why they came]
  TAKEAWAY  = [the one thing they should remember]
  MINUTES   = [talk length]
  STORIES   = [real examples I can tell; write NONE]
  LEVEL     = [beginner / practitioner / expert]

Role: A speaker who has watched people leave halfway through talks.
Task: Outline a MINUTES-minute talk for AUDIENCE landing on TAKEAWAY.
Rules:
- Every section must earn its minutes. Give a time budget per section.
- Use STORIES where they fit. Do not invent anecdotes.
- Match LEVEL. Do not explain what this audience already knows.
- Include the one slide that, if I only had two minutes, I would show.
Output: Section list with minutes, the through-line in one sentence, and three
questions the audience will ask.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  PROCESS   = [what the procedure covers, start to finish]
  DOER      = [who performs it and what they already know]
  TOOLS     = [systems, access, files needed]
  FAILURE   = [what goes wrong most often]
  FREQUENCY = [how often this runs]

Role: A technical writer documenting a process for someone's first day.
Task: Write the SOP for PROCESS.
Rules:
- Numbered steps. One action per step. Start each with a verb.
- Name the exact tool and screen at each step, from TOOLS.
- Add a checkpoint after any step that is hard to undo.
- Put FAILURE in a troubleshooting section with the fix, not a warning box.
- Assume DOER knows nothing not stated.
Output: Prerequisites, Steps, Checkpoints, Troubleshooting, and how to tell it
worked.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  RECIPIENT = [who, and our history]
  NEWS      = [the actual message, stated bluntly to yourself first]
  REASON    = [the real reason; write WITHHOLD if it stays private]
  OFFER     = [what I can still do; write NOTHING]
  RELATION  = [keep / wind down / end]

Role: Someone who has delivered this kind of news before and did not hide.
Task: Write the message to RECIPIENT.
Rules:
- NEWS appears in the first two sentences. Never in the last paragraph.
- No apology inflation. One acknowledgement, not four.
- If REASON is WITHHOLD, say a decision was made without inventing a cover story.
- If OFFER is NOTHING, do not manufacture a consolation.
- Match RELATION in the closing.
Output: The message, then one line on the reply I should expect and how to
answer it.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  CONCEPT   = [what I am explaining]
  READER    = [their background and what they care about]
  WHY       = [why they need to understand this]
  DEPTH     = [gist / working knowledge / enough to make a decision]
  ANALOGY   = [domain they know well; write NONE]

Role: A teacher who thinks jargon is a failure of explanation.
Task: Explain CONCEPT to READER at DEPTH.
Rules:
- No term used before it is defined in plain words.
- Draw analogies only from ANALOGY. If NONE, use everyday objects.
- State where the analogy breaks. Every analogy breaks somewhere.
- Tie back to WHY at least once.
- Stop at DEPTH. Do not keep going out of enthusiasm.
Output: The explanation, then a one-sentence version and a check question that
reveals whether they got it.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  DRAFT     = [paste it]
  TARGET    = [word count or percentage cut]
  KEEP      = [claims, numbers, or lines that must survive verbatim]
  VOICE     = [preserve / tighten / no preference]

Role: A ruthless editor who cuts words, not meaning.
Task: Cut DRAFT to TARGET.
Rules:
- Everything in KEEP survives word for word.
- Cut in this order: filler phrases, redundant restatement, hedging, examples.
- Do not merge two ideas into one vaguer sentence to save words.
- Do not add anything new.
Output: The cut version, the new word count, and a list of what you removed
grouped by category so I can put anything back.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  DRAFT     = [paste it]
  TELLS     = [phrases I hate; add to the default list below]
  VOICE     = [two or three adjectives, or a writer I sound like]

Role: An editor who can hear the difference between written and generated.
Task: Rewrite DRAFT so it reads as human, keeping every claim intact.
Rules:
- Remove: tricolon padding, "it's not just X, it's Y", "in a world where",
  "let's dive in", "the bottom line", symmetrical paragraph lengths, and TELLS.
- Break the rhythm. Follow long sentences with short ones.
- Cut every hedge that does not carry real uncertainty.
- Keep contractions where a person would use them.
- Change no facts.
Output: The rewrite, then a table of each tell you found and what replaced it.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  SAMPLES   = [paste 3 pieces by the target voice]
  NEW       = [what to write in that voice]
  LENGTH    = [word count]
  AVOID     = [things the voice never does]

Role: A ghostwriter who studies before writing.
Task: First analyze SAMPLES, then write NEW in that voice.
Rules:
- Before writing, list the observable patterns: sentence length range, contraction
  use, jargon level, humor, how paragraphs open, what the voice never says.
- Write NEW using those patterns, not a general idea of "professional".
- Never do anything in AVOID.
- Do not imitate the topics of SAMPLES, only the voice.
Output: The pattern analysis first, then NEW at LENGTH, then a note on which
pattern you were least confident about.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  TEXT      = [paste it]
  FROM      = [current audience]
  TO        = [new audience and their background]
  PRESERVE  = [terms that must stay, even if technical]
  LENGTH    = [same / shorter / longer is fine]

Role: An editor who moves text between audiences without dumbing it down.
Task: Rewrite TEXT for TO.
Rules:
- Keep every term in PRESERVE, but define each on first use.
- Replace other jargon with plain equivalents, not with vaguer words.
- Do not remove nuance to simplify. Explain it instead.
- Do not add encouragement, emoji, or "don't worry" reassurance.
Output: The rewrite, then a list of every simplification that lost precision,
so I can decide whether it matters.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  DRAFT     = [paste it]
  GOAL      = [what the reader should do or believe at the end]
  FORMAT    = [essay / doc / email / deck notes]
  CONSTRAINT= [anything that must stay in place, e.g. opening line]

Role: A structural editor who reorders before rewriting.
Task: Restructure DRAFT to serve GOAL.
Rules:
- Do not rewrite sentences yet. Move and label them.
- Identify anything that is said twice and mark the better version.
- Identify anything the reader needs that is missing. Mark it as a gap, do not
  fill it with invented content.
- Respect CONSTRAINT.
Output: The proposed outline with each existing paragraph mapped to a position,
a list of duplicates, and a list of gaps.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  DRAFT     = [paste it]
  FACTS     = [numbers, dates, names I can actually use; write NONE]
  DOMAIN    = [field, so the specificity is plausible]

Role: A skeptical editor who circles every unfalsifiable sentence.
Task: Find every vague claim in DRAFT and make it specific.
Rules:
- Flag every sentence that could be said about any company in DOMAIN.
- Replace vague claims using only FACTS.
- Where FACTS has nothing to offer, do not invent. Write [NEEDS DATA: what is
  needed] and leave the sentence flagged.
- Do not add adjectives as a substitute for specificity.
Output: A table — original sentence, why it is vague, replacement or NEEDS DATA
marker.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  TEXT      = [paste it]
  TARGET    = [language and regional variant]
  REGISTER  = [formal / neutral / casual, and to whom]
  KEEP      = [brand names, product terms, or phrases to leave untranslated]
  CONTEXT   = [where this will appear]

Role: A translator who localizes register, not just words.
Task: Translate TEXT into TARGET.
Rules:
- Hold REGISTER consistently, including forms of address.
- Leave everything in KEEP untranslated.
- Where an idiom does not carry, replace it with a local equivalent and flag it.
- Where a sentence is ambiguous in the source, ask rather than picking.
Output: The translation, then a list of every choice you made that a native
speaker might make differently.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  SOURCE    = [paste the piece]
  FORMATS   = [the 3 outputs, e.g. email, LinkedIn post, script]
  PLATFORM  = [where each will run, with any length limits]
  CORE      = [the idea that must survive all three]

Role: An editor who repurposes without recycling the same sentences.
Task: Turn SOURCE into each of FORMATS.
Rules:
- CORE survives in all three. Everything else is negotiable.
- Rewrite for each platform's reading behavior. Do not paste the same paragraph
  three times with different line breaks.
- Respect the length limits in PLATFORM exactly.
- No cross-promotional filler, no "link in bio" unless I asked.
Output: The three pieces, labeled, each with a one-line note on what you cut for
that format and why.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  DECISION  = [what I am deciding]
  OPTIONS   = [the candidates, 2-6]
  CRITERIA  = [what actually matters, in priority order]
  CONSTRAINT= [budget, time, team, anything fixed]
  UNKNOWNS  = [what I do not know yet]

Role: An analyst who refuses to fill a cell with a guess.
Task: Compare OPTIONS on CRITERIA for DECISION.
Rules:
- One table. Rows are CRITERIA in my order, columns are OPTIONS.
- Where you do not know a value, write UNKNOWN. Never estimate silently.
- Note which CONSTRAINT each option violates, if any.
- After the table, name the option you would pick and the single fact that would
  flip your answer.
Output: The table, the pick, the flip condition, and what to check from UNKNOWNS
first.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  POSITION  = [what I currently believe]
  OPPOSING  = [the view I think is wrong]
  STAKES    = [what it costs me if I am wrong]
  EVIDENCE  = [what I am basing POSITION on]

Role: The most capable advocate OPPOSING could hire.
Task: Argue OPPOSING as strongly as it can honestly be argued.
Rules:
- No strawman version. Argue the best form, not the common form.
- Attack the weakest link in EVIDENCE specifically.
- Do not balance the answer with reassurance that I might still be right.
- End by stating what would have to be true for OPPOSING to win outright.
Output: The strongest case, the specific weakness in my evidence, and the one
test that would settle it.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  PLAN      = [what we are about to do]
  HORIZON   = [when we will know if it failed]
  RESOURCES = [people, money, time committed]
  ASSUMES   = [what the plan takes for granted]

Role: A team member at the post-mortem, six months from now, explaining the
failure.
Task: It is HORIZON and PLAN failed. Write the account of why.
Rules:
- Write in past tense as though it happened. No hedging.
- Give at least three distinct failure paths, not three versions of one.
- At least one must come from ASSUMES turning out false.
- At least one must be a slow failure that looked fine for months.
- For each, name the earliest signal we could have seen.
Output: Three failure accounts, each with its early warning signal and the check
that would catch it.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  PLAN      = [the plan, decision, or argument]
  KNOWN     = [what I have verified, with how I verified it]
  DOMAIN    = [the field, so domain-specific assumptions surface]

Role: A reviewer whose job is to separate verified facts from inherited beliefs.
Task: List every assumption PLAN depends on.
Rules:
- Separate into: stated assumptions, unstated assumptions, and things I have
  labeled as facts in KNOWN that are actually assumptions.
- Rate each on how load-bearing it is: if false, does the plan wobble or collapse?
- Flag any assumption that is common in DOMAIN and frequently wrong.
- Do not propose fixes yet.
Output: A table — assumption, stated or unstated, load-bearing rating, and the
cheapest way to test it.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  CLAIM     = [the claim under dispute]
  MY_SIDE   = [what I hold and why]
  THEIR_SIDE= [what they hold and why]
  AVAILABLE = [evidence we could actually get]

Role: A neutral facilitator who cares about resolution, not fairness theater.
Task: Identify what evidence would move each side.
Rules:
- For each side, state the specific observation that should change that position.
- Mark any position for which no evidence would suffice. That is the real finding.
- Only propose tests drawn from AVAILABLE.
- Do not declare a winner.
Output: Two falsification conditions, a note on whether either side is
unfalsifiable, and the cheapest test in AVAILABLE that discriminates between them.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  QUESTION  = [what I am estimating]
  ANCHORS   = [any real numbers I do have; write NONE]
  BOUNDS    = [what would obviously be too high or too low]
  PRECISION = [order of magnitude / within 2x / within 20%]

Role: An analyst who shows the arithmetic.
Task: Estimate QUESTION to PRECISION.
Rules:
- Decompose into factors before multiplying anything.
- Every input is either from ANCHORS or explicitly labeled as an assumption
  with the basis stated.
- Sanity-check the result against BOUNDS and say so if it fails.
- Give a range, not a single number, and say which input the range is most
  sensitive to.
Output: The decomposition, the arithmetic, the range, and the one input worth
replacing with real data first.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  PLAN      = [the plan in full]
  ADVERSARY = [who benefits if this fails, or what force opposes it]
  BUDGET    = [what the adversary can spend, in effort or money]
  CROWN     = [the thing that must not break]

Role: ADVERSARY, with BUDGET, trying to make PLAN fail.
Task: Find the attacks with the best return on effort.
Rules:
- Rank attacks by effort required, cheapest first.
- At least one attack must target CROWN directly.
- Include at least one attack that exploits a process or a person, not a system.
- No attack may require capabilities beyond BUDGET.
- Do not suggest defenses yet.
Output: Ranked attacks, each with effort, impact, and what the defender would
see first.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  ITEMS     = [the list, one per line, with any detail you have]
  GOAL      = [what this quarter is actually for]
  SCORE     = [the dimensions, e.g. impact on GOAL, effort, risk, confidence]
  WEIGHTS   = [relative weight per dimension]
  CAPACITY  = [what we can realistically ship]

Role: A prioritizer who publishes the formula before the ranking.
Task: Score and rank ITEMS.
Rules:
- State the scoring formula from SCORE and WEIGHTS before scoring anything.
- Score every item on every dimension. Write UNKNOWN where you cannot.
- Draw the line at CAPACITY and say explicitly what falls below it.
- Flag any item whose ranking is driven mostly by an UNKNOWN.
Output: The formula, the scored table, the ranked list with the cut line marked,
and the three items whose scores are least trustworthy.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  SOURCE    = [paste it]
  READER    = [who reads the summary and what decision they face]
  LENGTH    = [word count]
  MUST_KEEP = [things that cannot be dropped: numbers, caveats, names]

Role: A summarizer who treats omission as a form of error.
Task: Summarize SOURCE for READER in LENGTH words.
Rules:
- Everything in MUST_KEEP appears, with its original qualifiers attached.
- Never state as fact something the source hedged. Carry the hedge.
- Do not add interpretation, implication, or advice not in the source.
- If the source contradicts itself, say so rather than picking a side.
Output: The summary, then a short list of anything significant you had to leave
out at this length.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  SOURCE    = [paste the documents or text]
  SCHEMA    = [field name, type, and what counts as a valid value, per field]
  MISSING   = [what to write when a field is absent, e.g. null]
  ROWS      = [one row per what?]

Role: A data extractor that never invents a value.
Task: Extract SOURCE into SCHEMA, one row per ROWS.
Rules:
- Output valid JSON only. No prose before or after.
- Every field in SCHEMA appears in every record, in the same order.
- If a value is absent from the source, use MISSING. Never infer it.
- If a value is ambiguous, use MISSING and add it to an "uncertain" array.
- Do not normalize units or formats unless SCHEMA says so.
Output: A JSON array of records, then a separate JSON object listing counts of
MISSING per field.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  DOCS      = [paste each, labeled A, B, C...]
  AXES      = [the dimensions to compare on]
  PURPOSE   = [the decision this comparison feeds]

Role: A comparative reader who does not let one document set the agenda.
Task: Compare DOCS across AXES.
Rules:
- Use only what each document actually says. Write NOT STATED, not a guess.
- Never treat one document's framing as the neutral baseline.
- Quote the specific line supporting each cell, or mark it as inference.
- Note where two documents use the same word to mean different things.
Output: A table with AXES as rows and documents as columns, a list of
terminology mismatches, and what PURPOSE still needs that none of them answer.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  DOC       = [paste it]
  TYPE      = [proposal / report / spec / marketing page / contract]
  EXPECT    = [what a complete document of TYPE would cover]
  STAKE     = [what I am risking if a gap goes unnoticed]

Role: A reviewer who reads for absence.
Task: Find what is missing from DOC.
Rules:
- Compare against EXPECT and list every item not addressed.
- Separate omissions that are normal for TYPE from omissions that are notable.
- Flag anything mentioned once and never resolved.
- Flag any number given without a method, source, or date.
- Rank gaps by relevance to STAKE.
Output: A ranked gap list, each with the question I should ask to close it.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  DOC       = [paste the source]
  QUESTIONS = [one per line]
  SCOPE     = [answer only from DOC, or DOC plus general knowledge, clearly labeled]

Role: A researcher who cites or abstains.
Task: Answer each of QUESTIONS.
Rules:
- Every answer must include a verbatim quote from DOC supporting it.
- If DOC does not answer a question, write NOT IN SOURCE. Do not reason around it.
- If SCOPE allows outside knowledge, put it in a separate clearly labeled section,
  never mixed with sourced answers.
- Never paraphrase a quote and present it as a quote.
Output: Question, answer, supporting quote. Then a list of every NOT IN SOURCE.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  SOURCES   = [paste everything, labeled]
  WINDOW    = [the period that matters]
  ACTORS    = [who or what is involved; write UNKNOWN]
  QUESTION  = [what I am trying to work out]

Role: An investigator building a chronology from partial records.
Task: Reconstruct the timeline within WINDOW.
Rules:
- Each entry: timestamp, actor, event, source label.
- Where a timestamp is relative or absent, mark it ESTIMATED and say why.
- Mark contradictions between sources rather than resolving them.
- Mark gaps where something must have happened but is not recorded.
Output: The chronology, a list of contradictions, a list of gaps, and what the
timeline suggests about QUESTION with confidence stated.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  TEXT      = [paste it]
  MY_FIELD  = [what I do know, so analogies land]
  PURPOSE   = [what I need from this text]
  DEPTH     = [enough to follow / enough to argue with]

Role: A translator between professional dialects.
Task: Decode TEXT for someone from MY_FIELD.
Rules:
- Build a glossary of every term of art, defined in plain words.
- Where a term has a near-equivalent in MY_FIELD, say so and say where the
  equivalence breaks.
- Flag terms that look ordinary but carry a specific technical meaning here.
- Then restate the text's actual argument in plain language at DEPTH.
Output: Glossary, false-friend list, and the plain restatement. Note anything
relevant to PURPOSE that the text assumes you already know.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  NOTES     = [paste transcript or notes]
  ATTENDEES = [names and roles]
  PRIOR     = [decisions from previous meetings this might revisit]
  DEFAULT   = [what to do when no owner was named]

Role: A chief of staff who leaves nothing implicit.
Task: Turn NOTES into decisions, actions, and open questions.
Rules:
- A decision is only a decision if someone stated it. Anything softer goes to
  open questions.
- Every action needs an owner from ATTENDEES and a date. If none was stated,
  apply DEFAULT and mark it UNASSIGNED.
- Flag anything that contradicts PRIOR.
- Do not infer agreement from absence of objection.
Output: Decisions, Actions (owner, date, marked if UNASSIGNED), Open questions,
Contradictions with PRIOR.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  ARTIFACT  = [what is being graded]
  QUALITIES = [what good looks like, in your words]
  SCALE     = [e.g. 1-5, with 3 defined as acceptable]
  FAIL      = [anything that is an automatic fail regardless of score]

Role: An assessor who writes the rubric before seeing any submissions.
Task: Build a reusable rubric for ARTIFACT.
Rules:
- One row per quality from QUALITIES.
- For each, write observable descriptors at every point on SCALE. Descriptors
  must be things a reader can point at, not adjectives.
- FAIL conditions are checked first and are not scored.
- No overlapping criteria. If two rows would move together, merge them.
Output: The rubric table, the fail checklist, and a worked example of a
borderline case scored and justified.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  PROCESS   = [what the checklist covers]
  WHEN      = [the moment it is used, e.g. pre-deploy, pre-send]
  MISSES    = [what has been forgotten before]
  TIME      = [how long the checklist may take to run]

Role: A safety engineer who writes checklists people actually complete.
Task: Write a checklist for PROCESS, run at WHEN.
Rules:
- Only items that can be verified in the moment. No "make sure it is good".
- Every item in MISSES has a corresponding check.
- Total run time under TIME. Cut the lowest-value items to fit and say what you cut.
- Order by consequence of failure, worst first.
- Each item is a yes or no, phrased so yes is the safe answer.
Output: The checklist, the cut items with reasons, and the one item to keep if
you could only keep one.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  ITEMS     = [the things to classify]
  LABELS    = [the complete label set, each with a one-line definition]
  EDGE      = [2-3 worked examples of hard cases, with the correct label]
  UNSURE    = [the label to use when confidence is low]

Role: A classifier that never invents a category.
Task: Assign exactly one label from LABELS to each item in ITEMS.
Rules:
- Use only LABELS or UNSURE. Never create a new label or a hyphenated blend.
- Match the reasoning shown in EDGE for similar cases.
- One label per item. If two fit, choose per the definitions and note the
  competing label.
- Do not explain unless the item was UNSURE or contested.
Output: A two-column table, item and label, then a separate list of contested
and UNSURE items with one line of reasoning each.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  EXAMPLE_IN  = [one input, exactly as it looks]
  EXAMPLE_OUT = [the correct output for it, exactly as it should look]
  ITEMS       = [the rest of the inputs]
  INVARIANT   = [what must be identical across every output]

Role: A transformer that copies a pattern rather than interpreting a description.
Task: Apply the EXAMPLE_IN to EXAMPLE_OUT transformation to every item in ITEMS.
Rules:
- Match EXAMPLE_OUT's structure exactly. Change only the values.
- Everything in INVARIANT is byte-identical in every output.
- If an input does not fit the pattern, output it unchanged with a FLAG, do not
  improvise a variation.
- No commentary between items.
Output: The transformed items in input order, then a list of every FLAG.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  DATA      = [paste it]
  RULES     = [the normalization rules, one per line]
  AMBIGUOUS = [what to do with values that fit no rule]
  KEEP_RAW  = [true / false, keep the original value alongside]

Role: A data cleaner who logs every change.
Task: Apply RULES to DATA.
Rules:
- Apply RULES in the order given. Do not apply unlisted rules, however obvious.
- Anything not covered by RULES goes to AMBIGUOUS handling, never silent
  correction.
- If KEEP_RAW is true, output original and cleaned values side by side.
- Never drop a row. Mark it instead.
Output: The cleaned data, a change log of rule applied per row, and a list of
values that needed a rule you did not give me.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  THING     = [the feature, form, process, or function]
  INPUTS    = [what it accepts, with the expected shape]
  ASSUMES   = [what it takes for granted about its inputs or environment]
  BLAST     = [what breaks downstream if it fails]

Role: A tester who enjoys finding the input nobody considered.
Task: Enumerate the ways THING fails.
Rules:
- Cover: empty, maximum, malformed, duplicate, out-of-order, wrong type, and
  concurrent inputs.
- At least three cases must come from ASSUMES being violated.
- Include at least two failures that are silent rather than loud.
- Rank by BLAST radius, not by likelihood.
Output: A table — case, input, expected behavior, likely actual behavior, blast
radius. Then the three worth testing first.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  ROLE      = [what this assistant is for, one job only]
  CONTEXT   = [the standing facts it always needs: product, audience, constraints]
  ALWAYS    = [behaviors that apply to every response]
  NEVER     = [behaviors that are never acceptable]
  UNKNOWN   = [what to do when it lacks information]

Role: A prompt engineer writing a durable configuration, not a one-off request.
Task: Write a system prompt for ROLE.
Rules:
- Standing facts only. Nothing that will be false next month.
- ALWAYS and NEVER as explicit lists, not prose.
- Include the UNKNOWN behavior as a hard rule, phrased as asking rather than
  guessing.
- Keep it under 300 words. Long system prompts get ignored in the middle.
- No task-specific instructions. Those go in the user turn.
Output: The system prompt, then a note on which rule is most likely to be
overridden by a strongly worded user message.

If any slot above is still in brackets, ask me for it. Do not guess it.

FILL IN
  PROMPT    = [paste the prompt that worked]
  OUTPUT    = [paste what it produced; write NONE]
  REUSE     = [what will change between runs]
  FIXED     = [what must stay the same every run]

Role: A prompt engineer converting a lucky result into a repeatable one.
Task: Turn PROMPT into a slot-based template.
Rules:
- Everything in REUSE becomes a named slot in a FILL IN block at the top.
- Everything in FIXED stays hard-coded. Do not parameterize for its own sake.
- Infer the implicit rules that made OUTPUT good and write them out explicitly.
- Add an output spec even if the original had none.
- Add the rule that unfilled slots must be queried, not guessed.
Output: The template, a list of the implicit rules you made explicit, and the
one slot most likely to be filled badly.

If any slot above is still in brackets, ask me for it. Do not guess it.