A three-step method to define quality standards for recurring AI tasks

Executive overview

AI outputs feel wrong but you can't explain why — so you tweak prompts endlessly and either give up or settle. The root cause is never defining what "good" looks like in terms the AI can act on.

The three C's framework — Capture, Convert, Check — extracts quality criteria from your own best work, turns them into binary pass/fail questions, and uses that checklist to diagnose and fix AI outputs systematically.

Your standards are trapped in your head; externalising them as binary questions is the fix.

Capture: extract criteria from your best examples

• Pull three to five diverse examples of outputs you already consider good.
• Feed them to the AI with a prompt asking it to identify specific, concrete patterns common across all examples.
• Request the top five criteria only, with one to two sentences per criterion — no essays.
• Diversity across examples matters; it prevents the AI from over-fitting to a single style.

Convert: turn criteria into binary questions

• Restate each criterion as a yes/no question with no room for interpretation.
• Test: if two people reading the same output could disagree on the answer, the question is too vague.
• Bad example: "Did the AI give a professional output?" (subjective). Good example: "Does the first paragraph directly answer what was asked?" (binary).
• Whittle the full list down to six to eight questions — more than eight creates too much friction and people revert to old habits.
• Mark two to three questions as must-pass: if any must-pass answer is no, the output cannot be used.

Check: evaluate output and improve the system

• After the AI produces output, run it against the binary checklist.
• Rule of thumb: if a question takes more than three seconds to answer yes or no, treat it as a no.
• A slow answer means either the question is unclear or the output is not good enough.
• Two root causes for a "no": the system prompt is missing an instruction, or the AI lacks context.
• Fix missing instructions by updating the system prompt; fix missing context by adding files to the project knowledge base.

Three common mistakes

1. Checklist too long — over eight questions guarantees abandonment.
2. Binary questions too vague — any subjectivity breaks the yes/no contract.
3. No must-pass questions — the AI can pass everything and still miss critical non-negotiables like including pricing in a proposal.