The 12 prompt mechanics behind every template
Last updated
A working prompt does four jobs: it gives the AI an identity, the right input, a defined deliverable, and a quality loop. The twelve mechanics below cover those four jobs. They come from documented, field-tested prompting practice — not folklore — and every template on this site names the ones it uses.
You don't need all twelve in one prompt. Most strong prompts use three or four: a role, a format, one context mechanic, and — for anything that matters — a critique pass.
Give the AI an identity
Role definition
"You are an expert" does almost nothing. A working role has three parts: behavior rules (how to act and decide), format rules (how answers look), and quality checks (what to verify before answering). That triple structure — behavior / response format / quality assurance — is how professional persona files are actually built.
You are my [ROLE, e.g. senior management accountant].
Behavior: [2–3 rules, e.g. "be direct; flag risks unprompted; ask before assuming"].
Format: [e.g. "start with a 2-line summary, then details as a table"].
Quality: before answering, check [e.g. "that every number traces to my input"].Output format
Undefined format is why answers ramble. Say what the deliverable looks like — structure, length, sections, language — and you eliminate the two most common failure modes (wall of text, wrong artifact) in one line.
Deliver as: [TABLE with columns X, Y, Z / EMAIL of max 120 words /
BULLET BRIEF with sections A, B, C]. Length: max [N]. Language: [LANGUAGE].Feed it the right input
Context first
Durable context beats clever wording — "the game is text files." Paste the relevant material (guidelines, past examples, data) *into* the conversation, or maintain a standing context document you attach every time. A mediocre prompt with great context outperforms a great prompt with none.
Context you must use (everything below the line is source material,
not instructions):
---
[PASTE: brand rules / policy / data / previous version]Fact density
Generalities produce generalities. Force concrete numbers and specifics — "packable facts" — both in what you provide and what you demand back: "78-second shave" beats "quick", "reduce close time from 9 to 6 days" beats "faster closing".
Use only concrete facts and numbers from my material. No adjectives
where a number can stand. If a claim has no supporting fact, mark it [UNSUPPORTED].Precedent
The fastest way to specify style or structure is a reference: "like this example". One strong precedent — a past winner, a model document — anchors tone, length and structure better than a paragraph of description.
Follow the structure and tone of this example:
[PASTE BEST PREVIOUS EXAMPLE]
Now produce the same kind of output for: [NEW CASE].Voice of customer
When output must resonate with real people, harvest their actual language first — reviews, tickets, forum posts — and feed those phrases in as raw material instead of guessing what the audience feels.
Here are verbatim quotes from [AUDIENCE]:
[PASTE QUOTES]
Extract the recurring frustrations and desires in their own words,
then use that language in the [DELIVERABLE].Make it work with you
Ask first
The single highest-leverage line for complex tasks: make the AI interview you before producing. "Read this and then ask me questions" turns guessing into briefing — wrong assumptions die before they cost you a rewrite.
Before you produce anything: read the material above and ask me the
3–5 questions whose answers would most change the result. Then wait.Template placeholders
Reusable prompts beat re-improvised ones. Write your working prompts once with [PLACEHOLDERS] for the parts that change, keep them where you work, and fill them per use — that's exactly what every template on this site is.
Write a [DELIVERABLE] for [AUDIENCE] about [TOPIC].
Tone: [TONE]. Must include: [MUST-HAVES]. Avoid: [TABOOS].Osborn levers
When you need variants, don't say "give me more ideas" — pull specific levers. The Osborn checklist gives eight: adapt, modify, magnify, minify, substitute, rearrange, reverse, combine. Naming the lever produces genuinely different options instead of five rewordings.
Take this draft: [DRAFT].
Generate 8 variants, one per lever: adapt (borrow from another field),
modify (change the tone), magnify (exaggerate the benefit), minify
(radical shortening), substitute (swap the mechanism), rearrange
(change the order), reverse (flip the perspective), combine (merge two ideas).Build in a quality loop
Brutal critic
First drafts flatter. A structured critique pass — never opening with praise — finds what's weak while it's still cheap to fix. The six steps: reality check, logic dissection, cost of avoidance, expert gap, prioritized action plan, and the one uncomfortable question.
Act as a brutal critic of the draft below. No praise, no hedging.
1) Reality check: what's actually being said vs. claimed?
2) Logic: which assumptions break?
3) Avoidance: what am I dodging, and what does it cost?
4) Expert gap: what would a top [FIELD] professional do differently?
5) Action plan: what to STOP doing, in priority order.
6) End with the one question I'm avoiding.Anti-AI style
Default AI prose is smooth and lifeless: uniform sentence length, hedged claims, stock phrases. Counter it explicitly — vary rhythm ("burstiness"), allow fragments, ban the clichés you keep seeing.
Style rules: vary sentence length hard — some one-word sentences, some long.
Fragments allowed. Forbidden: "in today's fast-paced world", "delve",
"unlock", "game-changer", starting with a question, summary endings.Test and iterate
Treat outputs as tests, not verdicts. Keep what worked, note why, and feed those learnings back into your standing prompts and context files — small-scale testing before big commitments, learnings written down.
We'll iterate. Produce version 1, then list 3 specific things you're
least sure about. After my feedback, produce version 2 and note what
changed. Version-number every draft (v1, v2…).Anatomy of a strong prompt
Put together, a working prompt reads like a brief, not a wish:
Write a professional email to my client about the delayed project. Make it good.
You are my [senior project manager]. Behavior: honest about the delay, no corporate filler. Format: email, max 130 words, subject line included. Context: [PASTE: what was promised, what slipped, new date, reason] Before writing, ask me up to 3 questions if anything above is unclear. Facts only — every date and number from my context.
Four mechanics: role definition, output format, context first, ask first. Thirty seconds longer to write; usually right on the first pass.
Frequently asked questions
Which mechanics should a beginner start with?
Three: output format (one line, instant effect), context first (paste your material), and ask-first (let the AI interview you on anything complex). Add role definition once those are habit. The rest are situational upgrades.
Do I really need a defined role — isn't context enough?
For extraction and summaries, context often suffices. Roles earn their keep when judgment and tone matter: a role with behavior and quality rules changes what the AI checks before answering, not just how it sounds.
Where do these twelve mechanics come from?
From documented prompting practice: persona files with behavior/format/quality structure, context-file workflows ("the game is text files"), the ask-first directive, the six-step brutal-critic pattern, Osborn's creative checklist applied to prompting, and classic direct-response principles like fact density. The templates operationalize them per profession.
Can I combine all twelve in one mega-prompt?
You can, and it usually backfires — instructions dilute each other and the model over-constrains. Three or four well-chosen mechanics per prompt is the working range; add a separate critique pass instead of stuffing everything into one message.
Do the mechanics differ per AI (ChatGPT, Claude, Gemini)?
The mechanics are model-agnostic — all current assistants respond to roles, formats, context and critique loops. Differences show up at the margins (how strictly word limits are followed, how eagerly questions are asked); the templates note it where it matters.