Most bad AI answers are not the assistant's fault. They are a prompt with empty brackets – the model guessed at your role, your reader, your length and your facts, and guessed wrong. The fix is a template, not a smarter model.
Here is the whole thing: role, bracketed context, numbered steps, then output format and rules last. Same order every time. The output line governs everything above it, so it goes last.
By the end you will have the master template in a notes file, plus 15 filled-in copies for the tasks that keep coming back wrong – the PDF summary, the invoice table, the voice rewrite, the fact-check.
| Prompt | Best-suited tool | Minutes by hand | AI's own time | Minutes to review |
|---|---|---|---|---|
| Master template fill-in | ChatGPT | 8 | 20-90 sec | 6 |
| Role line | Claude | 5 | 20-60 sec | 4 |
| PDF summary with page cites | Claude | 60 | 1-2 min | 12 |
| Voice rewrite from sample | Claude | 20 | 30 sec | 6 |
| Invoice table with CHECK | ChatGPT | 50 | 60 sec | 15 |
| Fact-check verdict table | Perplexity | 45 | 90 sec | 20 |
1. The fill-in formula, four parts in order
Forty seconds after hitting enter, the reply came back as three labeled parts: five bullets, a two-sentence summary, and one question about the missing launch date. The same launch note written cold, with no skeleton, had taken two attempts and about 25 minutes of rewrite.
Prompt to copy
You are a copy editor for B2B software. Context: [paste the material – the draft, the notes, the audience, the goal]. Do this in order: 1. [first step]. 2. [second step]. 3. [third step]. Output: [length and format, e.g. five bullets under 20 words each, then a two-sentence summary]. Rules: If something is missing or unclear, ask me up to 3 questions before answering. Mark anything you are not sure of as UNVERIFIED. Do not invent sources, quotes, links or figures. Use plain American English.
Sequence is the whole trick. Put the output and rules lines above the numbered steps and they stop governing what comes back, reading as background instead.
Keep them last and every step above answers to them. The four names never move, only what sits inside the brackets, job to job.
A document summary keeps the same skeleton with an analyst role and steps that ask for a page number beside each point.
An invoice job swaps in extract steps and a CHECK rule. Two jobs, one saved template, four parts with new bracket contents each time.
By the numbers: about 12 minutes by hand to write and rewrite one answer, 20 to 90 seconds for the reply, and 5 to 20 minutes of your review.
Save the template in a notes file and the next bad answer gets a refill, not a fresh prompt.
💡 Tip: Save the template with the brackets still in it. Filling brackets is faster than composing a prompt from scratch.
The four parts, decoded
2. The role line that kills generic answers
Prompt to copy
You are a [specific role] with [X years or specialty] working for [type of company]. You are writing for [audience] who already knows [what they know]. Skip the basics and assume [level]. Your bias is toward [what this role cares about most – speed, accuracy, cost].
The two slots that do the work are audience and what that audience already knows. Years, company type and bias adjust the tone a little.
A filled version: a copy editor with ten years in B2B software, writing for a procurement manager who knows the pricing tiers, basics skipped, weighted toward accuracy.
The catch: a bare title, or "you are an expert," still opens with "In today's fast-paced business world." A role that names no reader gives the model nothing to aim at.
Swap the brackets and the same line serves a board memo or a support reply, in one setup of about two minutes.
⚠️ COMMON MISTAKE
Redact before you paste
On free and personal plans, chats may be used to improve the company's models unless that is switched off in settings. Strip names, account numbers, passwords, client files, and medical or financial records before anything goes into a prompt.
3. Context slots that pin your material
Six brackets, none empty: that is what turned a three-page draft into a reply a VP of sales could read with small edits instead of a rewrite.
Before you paste, take out names, account numbers and passwords, and leave client files or health and money records out entirely – chats on free and personal plans may train the company's models unless switched off.
Prompt to copy
Context: [paste the document or notes], [who will read the output], [what decision it feeds], [what good looks like], [hard length limit], [things to avoid].
Source, reader, goal, length: fill those four first and the answer already lands in the right place.
Good and avoid: add what good looks like and what to avoid once the four are set.
In the example, 400 words was the hard cap, and jargon and hedges sat on the avoid line.
4. Numbered steps that force an order
Ask for a draft with no sequence stated and the assistant picks its own order, so facts, caveats and the actual ask land in whatever position it likes.
The safe version numbers the work and asks to see it: four steps, run in order, before any prose is written.
Prompt to copy
Do this in order and show each step: 1. List the 3 claims this piece must make. 2. For each claim, name the evidence I gave you or write NO EVIDENCE. 3. Draft only the claims that have evidence. 4. Re-read the draft and cut anything not supported by steps 1-2.
Each numbered line opens with a verb, so the instruction is the sequence, not a description of it.
On a launch note with three claims, the second had no evidence behind it, so step four dropped it. The fix belonged in the source document, not the prose.
Cutting is step four because a model adds by default and rarely deletes. Without that step the draft runs longer than the material it came from.
Run time is 20 to 60 seconds; the edit here took about 6 minutes of the 5 to 20 the job owns.
5. The output line that ends wall-of-text
The tedious part is a question worth three lines coming back as nine paragraphs, with the answer buried somewhere in the middle. What the prompt has not yet said is what a good reply looks like: its length, its order, its shape.
Prompt to copy
Output: one line that answers the question, then 3 bullets of no more than 15 words each, then a table with columns [A | B | C], then one question you still need answered. No preamble, no restating my prompt, no closing offer to help further.
The no-preamble line is the one doing the work. It removes the opening 'Certainly! Here's…' and the offer of more help at the end. Those two blocks are what make a five-line answer read like an essay, and they go on request.
Quick check: a word cap of 15 is a number you can count. 'Be concise' is not, and the model's reading of concise is a 200-word paragraph. Name the length without the order and you get 150 words with the answer in the last sentence.
💡 Tip: Count the bullets and the words. If either is over, re-prompt with the number and ask for the cut, not a fresh draft.
6. The rules block that blocks invented facts
Give it material to answer from and two ways out. Asked plainly whether a statistic is right, an assistant agrees and produces a plausible-looking source in seconds.
Prompt to copy
Rules: Answer only from the material above. If the answer is not in the material, write 'Not in the material' and list what you would need. Every figure, quote and link must come from the material or be marked UNVERIFIED. No invented citations, no invented page numbers. No hedging filler.
ChatGPT, Claude, Gemini, Copilot and Perplexity will all state a false fact or a dead link without hesitation.
UNVERIFIED tells you which line to open, not that the line is true. With nothing pasted above the rules block, the whole reply reads 'Not in the material'.
7. Ask me three questions before answering
Rules keep the assistant honest about material you already pasted. A brief with holes in it is a different gap, and the tidy fix is one line that makes the assistant ask first instead of filling the holes with guesses.
Prompt to copy
Before you answer, ask me up to 3 questions whose answers would most change your answer. Wait for my replies. Then answer, and at the end name which of my answers changed what.
The waiting line carries the weight. Drop it and the assistant invents a deadline of "next month" and a discount of "15%", then drafts a launch email around both.
Spend the 30 seconds answering instead. Your three replies land in half a minute, before a wrong draft exists.
Save it for a brief you cannot fully fill in or a task you have not done before. A quick rewrite with a clear brief just gets slower by one round trip.
The finished draft closes with a short list of which answers shifted what, so the assumptions behind it sit in the open.
Pick the prompt for the job in front of you
8. Rewrite in my voice from a sample
Retyping a draft so it stops reading like a press release is the tedious part here.
Paste a sample of your own writing and the same four parts handle it: role and context show the assistant your rhythm, and the output line asks it to say what it copied.
Prompt to copy
Here is a sample of my writing: [paste 150-300 words I wrote]. Rewrite the draft below in that voice: same average sentence length, same formality, same vocabulary. Keep these terms unchanged: [list]. Draft: [paste]. Output: the rewritten text, then 3 bullets naming the voice choices you made.
The sample has to be yours, 150 to 300 words of it. Hand it a page of another chatbot's output and it learns that flat register and hands it straight back.
The limit is the genuine one: this copies surface habits, not the judgment behind them. Claude tracks voice most closely, about 30 seconds for the rewrite and 6 minutes to read it against the sample.
What that means in practice: any product name, job title or branded term left off the keep-list comes back swapped for something near enough.
A "Growth Plan" turns into a "growth package" and a "Head of Ops" becomes an "operations lead". Put every one on the list before you paste the draft.
Done that way, a 20-minute rewrite by hand becomes a light edit against the sample. You still read every line, because the assistant copies vocabulary, not the call you would make about which sentence to cut.
💡 Tip: Keep a saved 200-word sample of your own writing in the same notes file as the template. Paste it every time.
9. Summarize a long PDF with page cites
Every page number in an AI summary is a guess until you have opened the page yourself.
So demand them anyway, and check each one before the brief goes to anyone. On its own, a summary of a 40-page document is a smooth paragraph you cannot trace.
Prompt to copy
You are an analyst summarizing a document for [role]. Read the attached [PDF]. Do this in order: 1. Give a 5-sentence summary. 2. List the 7 decisions or numbers that matter, each with the page number it appears on. 3. List what the document does not answer. Output: the summary, then a table [Point | Page | Why it matters], then the gaps. Rules: page numbers only where they appear; write UNVERIFIED where you cannot find one.
Asking for seven specific figures, each pinned to a page, is what turns the brief into something you can audit in 12 minutes flat.
The step most summaries skip is the third, which names the questions the document leaves hanging. Claude takes the upload and reads it in a minute or two.
Against 60 minutes of reading by hand, that is a one-page brief with a trail back to the source. Open every cite first. A page number pointing at the wrong clause is worse than no number, because it looks checked.
10. Tighten to 150 words without losing the ask
About 20 seconds after you paste it, a 600-word draft comes back at 150. The keep-list has to be typed before the draft goes in. A list made after the cut just names whatever happened to survive, which tells you nothing.
Prompt to copy
Cut the draft below to 150 words or fewer. Keep exactly these points: [list]. You may drop anything else. Keep the call to action as the last line. Output: the cut version, then a word count, then one line naming what you cut. Draft: [paste].
A cut with no keep-list deletes the ask and keeps the warm-up, so the call to action is pinned as the last line.
By hand this runs about 20 minutes; review is 5. ChatGPT is fine here, and the short version still makes the same request.
💡 PRO TIP
Name what to keep on every repair
When an answer is close but one part is wrong, name the parts that worked and ask for only the broken section to be redone. A repair that names nothing returns the whole answer again with the fix buried somewhere inside it.
11. Critique my draft like an editor
The notes come back as five numbered fixes, and every one quotes a line you can find on the page. It takes 45 seconds to get them.
The one-time part is saving this prompt next to your drafts, so it is there whenever a piece feels finished but you suspect it is not.
That is a different job from trimming. A draft can sit inside its word count and still argue weakly, bury its point or open on a warm-up. Length control cannot see any of that. An editor's read can.
Prompt to copy
You are an editor at [publication or company]. Read the draft: [paste]. 1. Name the single biggest problem in one sentence. 2. List 5 line-level fixes, each quoting the exact line. 3. Say what the draft does well. 4. Rewrite only the opening 3 sentences. Rules: quote the real text; no advice like 'add more detail'.
By hand this takes about 35 minutes; review takes 10, spent working the list in one pass. Claude gives the sharper notes. Leave out the quote rule and the fixes arrive as headings that point at no line.
12. Compare two options in a decision table
The one that wins more rows, and the table tells you which rows matter for your situation. That is a comparison, not an essay.
By hand, an honest vendor comparison runs about 45 minutes of sorting quotes. The prompt takes 40 seconds and review takes 8 minutes. Gemini's search-and-table output suits it.
Prompt to copy
You are a buyer's advisor in [domain]. Compare [Option A: paste specs or quote] with [Option B: paste specs or quote]. Do this in order: 1. Build a table: [criterion | A | B | which wins]. 2. Mark the criteria where they tie. 3. Say which to pick for [situation 1] and which for [situation 2]. 4. List the 3 facts I should confirm with each vendor. Rules: use only the figures I gave you; write 'not stated' rather than guessing.
The same criteria run down both sides, or the table proves nothing. A winner named on each row stops the answer sliding into a paragraph of trade-offs. The three facts for the vendor are the part you leave with.
13. Extract numbers into a clean table
Prompt to copy
Extract data from the attached [files] into a table with columns: [date | vendor | amount | due date | notes]. Rules: one row per document; copy amounts exactly as printed with no rounding; if a field is missing or unreadable, leave the cell blank and mark it CHECK; then list every CHECK cell. Output: the table, then the CHECK list, then a one-line total.
A comparison table survives because its columns never move, and invoice extraction needs that same discipline plus one rule. Paste this into ChatGPT with the files attached.
Sixty seconds later you get rows, a CHECK list and a total. Expect about 15 minutes checking the flagged cells, against roughly 50 minutes typing it by hand in a spreadsheet.
An invoice whose amount field is smudged is where extraction fails quietly. Left to smooth the gap, the model writes a number near enough to look right.
A blank cell tagged CHECK points you at the smudge instead. One is a hidden error you find at tax time, the other a task you finish in a minute.
That clean total at the bottom is the trap. It makes a half-read stack of receipts look settled, so work the CHECK list before trusting it. A finished table should tell you exactly where to open the original, not reassure you that nothing needs opening.
💡 Tip: Do not sort or sum the table until every CHECK cell is cleared. A total over a guessed row is worse than no total.
14. Fact-check the answer before you use it
Fact-checking is the step where you confirm each claim against a source you opened yourself. Clean extracted numbers can still rest on a wrong figure, so this last pass walks each claim back to a real page.
Prompt to copy
You are a fact-checker. For each claim below, find a primary source. Claims: [list]. Output: a table [Claim | Verdict: supported / contradicted / not found | Source link | Quote from the source | Source date]. Rules: link only pages you actually opened; if you cannot find a source, write NOT FOUND – never guess a URL, study or quotation.
Paste it into Perplexity with your claims listed. Asking "is this right?" gets you a friendly paragraph agreeing.
A verdict column makes it pick supported, contradicted or not found for every row. About 45 minutes by hand, 90 seconds for the AI, then 20 minutes clicking each link.
The two times: 45 minutes by hand against about 22 minutes with the tool. Your final table can travel with the claim, showing exactly who checked what.
🔧 Repair a wrong answer in four moves
Name what was wrong
Say which part failed – too general, wrong on X, missed Y – in one line.
Name what worked
List the parts to keep so the model does not throw away a good draft.
Redo only the broken part
Ask for the section or question, not the whole answer again.
Add the new constraint
Name the audience, length and format that changed, and ask for no repetition.
15. Re-prompt a bad answer without starting over
One wrong section, one reply, about 30 seconds: that is what repairing a draft in place costs instead of retyping from scratch. Fact-checking catches the flaw before anything ships.
When the flaw is a single section of an otherwise good answer, fix that section and keep the rest.
Step 1 – name the damage. Stay in the same ChatGPT or Claude chat where the original answer arrived. Say which part failed and which parts earned their place, so the good opening paragraph survives the repair.
Prompt to copy
Your answer was [too general / wrong on X / missed Y]. Keep the parts that worked: [name them]. Redo only [the section or question]. New constraints: [audience], [length], [format]. Do not repeat what you already said.
Step 2 – read the reply. You should see the corrected section alone, not a fresh full draft.
Because the rules line still governs this chat, the repair only has to describe what changed.
Watch for the trap: drop the no-repeat line and you get the whole answer back, fix buried somewhere inside.
Step 3 – reread the kept parts. They should match the draft you approved, unchanged. If the ending now contradicts the repAired section, say so in one more short message. That is still faster than a new prompt and another wrong draft.
💡 Tip: Keep the repair in the same chat. Starting a new one drops the rules block and puts you back at square one.
The prompt behind the article – at a glance
🎭 Part 1
Role
📋 Part 2
Context brackets
🔢 Part 3
Numbered steps
📐 Part 4
Output + rules last
⏱️ AI time
20-90 sec
✏️ Review time
5-20 min
The one rule
If a bracket is empty, the model fills it by guessing – and the guess is the bad answer.
Frequently Asked Questions
Do I need to pay for a subscription to use these prompts?
No. Every prompt here works on the free tiers of ChatGPT, Claude, Gemini, Copilot and Perplexity, within their usage limits. The limits change often, so check the tool's own pricing page before relying on a long session.
Do I need to know how to code?
No. All 15 prompts are plain text you paste into a chat box. The only technical habit is filling brackets before you hit send, and saving the template in a notes file.
Is it safe to paste a work document into ChatGPT or Claude?
Only after redacting. On free and personal plans, chats may be used to improve the company's models unless that is switched off in settings. Strip names, account numbers, passwords, client files, and medical or financial records first.
Which assistant is best for these prompts?
Claude handles long PDFs and voice matching well. ChatGPT is reliable for extraction and rewrites. Perplexity suits the fact-check table because it links sources. Gemini's search-and-table output works for the comparison prompt.
How long does the whole formula take to learn?
About 15 minutes to save the template and fill it once. After that, filling the brackets for a new task takes 3 to 8 minutes, depending on how much source material you paste in.
What if the answer is still bad after I fill all four parts?
Use the repair prompt in #15 before you write a new one. Name the failure, list what to keep, and redo only the broken section. Nine times out of ten the problem is one missing constraint, not the wrong tool.
Paste the skeleton, fill two brackets, add the three questions
Open your notes file, paste the master template, and fill the role and context brackets for the next task on your list. That is the whole move, and it takes about 5 minutes.
Then add the three-questions step from #7, so the model asks before it drafts. When the answer arrives, run the check that matters: open every figure and link.
For the work in front of you this week, start with the PDF summary if you have a long document, the invoice table if you have files to reconcile, and the fact-check table before anything goes out under your name.