The prompting patterns that save the most time are the same structures a good manager uses with a new colleague: a brief that says what done looks like, a fact list that bounds what can be claimed, a checklist the output is scored against, a review loop where the model critiques its own draft, and a counter-argument request before any decision. Each takes a minute to set up and removes an hour of fixing output you cannot trust.
Here are the five, with a template for each and where it fits.
Pattern 1: the brief
A brief answers four questions before the model writes anything: who is this for, what will they do with it, what does done look like, and what is out of scope.
Template:
"Audience: [who reads or uses this]. Purpose: they will use it to [action]. Done means: [the shape of the output: length, format, sections, the one thing it must contain]. Out of scope: [what to leave out]. Constraints: [tone, banned words, sources allowed]."
Use it for: anything longer than a paragraph. The out-of-scope line is the one most people skip, and it is the one that stops the model padding.
Pattern 2: the fact list
Give the model the facts it may use, and say that nothing else is allowed. It turns "write about X" into "arrange these true things about X".
Template:
"Facts you may use, and the only facts you may use: [numbered list]. If something needed is not in the list, write [MISSING] in its place rather than guessing."
Use it for: resumes, product descriptions, bios, anything about real people or real products, reports from data you supply. The [MISSING] marker gives you a list of what to look up instead of an invention to find later.
Pattern 3: the checklist
Write the acceptance criteria before the draft, then ask the model to score its draft against them before you read it.
Template:
"After writing, check the draft against this list and report pass or fail for each line, quoting the part that passes: 1. [criterion] 2. [criterion] 3. [criterion]. Fix any fail and show the fixed version."
Use it for: specs (image sizes, word counts, required sections), tone rules (no exclamation marks, no superlatives), compliance (no claims about results). A checklist of five to eight lines catches most of what you would catch on a first read, so your read is a second pass.
Pattern 4: the review loop
One draft is rarely right. Instead of editing it yourself, run a short loop: draft, critique, revise, with the critique prompt written by you.
Template:
"Now act as [a specific critic: a sceptical buyer, a copy editor, a senior engineer]. List the five weakest points in the draft, each with the sentence it refers to. Then revise the draft to fix them. Do not add length."
Two loops is usually enough; a third tends to polish rather than improve. The "do not add length" line stops the revision bloating.
Pattern 5: the counter-argument
Before acting on a recommendation, ask for the case against it. Models agree with the framing they are given, and this is the cheapest correction.
Template:
"Give the three strongest arguments against [the recommendation], as someone who disagrees would make them. Then say which, if any, changes the recommendation."
Use it for: choosing between options, go or no-go decisions, any output that ends with "you should". If the model produces no real counter-argument, the recommendation was never examined.
Putting them together on one task
Say the task is a product description for a store listing.
- Brief: audience is a founder scanning listings on a phone; done is 120 words, what they get, what they send, delivery time, one plain closing line; out of scope is pricing comparisons and promises of results.
- Fact list: the deliverables, the turnaround, the revision count, the file formats. Nothing else.
- Checklist: under 130 words; every deliverable from the fact list present; no word from the banned list; no claim of outcomes; closing line present.
- Review loop: critic is a buyer who has been burned by vague listings; one loop.
- Counter-argument: "argue that this listing will attract the wrong buyers". Usually surfaces an ambiguous line to fix.
Twenty minutes, and the result needs a read rather than a rewrite.
Habits that make the patterns work
- Save the brief and checklist as text files per recurring task; the second time takes two minutes.
- Keep the fact list outside the conversation, in a document you own, and paste it in. It stays the source of truth as the conversation drifts.
- Read the pass or fail report before the draft. A failed criterion tells you where to look.
- When the output is still wrong after a loop, the brief was wrong. Fix the brief, not the draft.
Related reading: When to trust an AI agent's output and how to verify it for the verification side, and Rewrite your resume with AI without sounding like AI for the fact-list pattern applied to one document.
FAQ
What is the most useful prompting pattern? The brief: audience, purpose, what done looks like, what is out of scope. Most bad output comes from a missing brief rather than a weak model.
How do I stop AI from making things up? The fact list pattern: supply the only facts allowed and ask for a [MISSING] marker where a fact is absent, instead of a guess.
Should I ask AI to check its own work? Yes, against a checklist you wrote, with a pass or fail per line and the quoted evidence. Then read the report before the draft.
How many revision rounds are useful? Two. After that the model polishes rather than improves; if it is still wrong, change the brief.
How do I get an AI to disagree with me? Ask for the three strongest arguments against the recommendation, voiced by someone who disagrees, and whether any of them changes the answer.
Do these patterns work with agents as well as chat? Yes. An agent brief adds the sources it may read and the actions it may take; the checklist and verification steps matter more, since the agent's errors reach further.
If you want research done this way, with a brief, a fact list with sources and a verification pass, the lead list research gig on my store is one place it is applied: https://rmgo.sites.tab.bot/leads?s=blog
Latest comments
0