← Insights Prompting

The Prompt Anatomy Check

Cody Reppert · A working pattern from the cookbook

Most advice for getting better results from AI is a list of principles. Be specific. Give examples. Set the role. All true, and all useless in the moment, because you forget half of them every time you actually sit down to type.

So I stopped trying to remember them. Instead I wrote one short instruction that makes the model grade my prompt before it answers. It checks eight specific things, and if something important is missing, it says so and tells me what it is going to assume. Then it does the work.

It is the single highest-leverage paragraph in my setup. It runs on every project. It has cut my revision cycles roughly in half, because the model catches the gap before it produces the wrong thing, not after.

The eight parts

Every prompt that produces a usable result on the first try tends to answer eight questions. When one is missing in a way that matters, the output drifts. Here is the full anatomy.

PartWhat it answersCommon failure
OutcomeWhat is the actual deliverable?"Help with the deck." Which deck, for what?
Success criteriaWhat does "done well" look like?No definition of done, so the model guesses.
StyleTone, voice, register.Default AI voice, which is not your voice.
FormattingSections, length, layout.A wall of prose when you wanted a table.
EvidenceInputs, constraints, preferences.The model invents facts it was never given.
Search limitsWhen to look things up.Over-searches, or never searches when it should.
Drafting guardrailsWhat not to invent.Confident hallucinations.
ValidationA self-check before final.Output quietly violates a constraint you stated.

The paragraph you paste in

This is the whole thing. Paste it into your CLAUDE.md at the project root, or the top of any system prompt. It works in Claude Code, Cowork, Claude Desktop projects, and ChatGPT custom instructions. Anywhere you have a persistent instruction.

## Prompt Anatomy Check, Always On

Silently grade every prompt against the 8-part anatomy: Outcome, Success
Criteria, Style, Formatting, Evidence, Search Limits, Drafting Guardrails,
Validation.

- If all eight are clear (or absent in a way that doesn't matter), stay
  silent and do the work.
- If one or more are missing in a way that would materially change the
  output (different file, different scope, hallucinated content, missed
  constraint), surface a short flag block before doing the work, then
  proceed with stated defaults. Don't block.

Flag format:

Prompt gaps:
- [Part]: [what's missing, why it matters]

Default: [what I'll assume]

Material = would filling this gap change what gets made, what gets
included, or whether output is usable first try. Casual chat,
continuations, and obvious-context prompts don't get flagged.

About 150 words. Paste and go.

Why it works

The trick is not the eight parts. Plenty of checklists exist. The trick is the threshold: would filling this gap actually change the output?

Without that threshold, the model nags on every message, including "thanks" and "keep going," and you turn it off within a day. With it, the model stays quiet on casual chat and only speaks up when answering the gap would save you a revision. It checks you exactly when it matters and never when it doesn't.

This is a small example of a larger idea: the highest-leverage AI work is rarely a flashy automation. It is usually a small, durable instruction that quietly raises the quality of everything that comes after it. Multiply that across a team's daily work and you have real capacity back.

This is one pattern from the system I install inside firms. If your team is spending hours on work AI should be helping with, that is what a Capacity Audit is for.

Request an AI Capacity Audit