Compare and score versions

Choose four to six criteria tied to the task. For a project update:

• Accuracy: Claims match the supplied evidence.
• Completeness: Required decision, status, risk, and next step appear.
• Usability: The reader can find the main action quickly.
• Audience fit: Terms and detail suit the recipient.
• Tone: Direct and accountable without blame.
• Safety: No private data or unsupported commitments.

Define what each criterion means before scoring. A simple 0–2 scale is often enough: 0 fails, 1 partly meets, 2 meets. Scores create structure, not mathematical truth. Include notes that quote evidence from the draft.

Some failures should not be averaged away. Mark privacy leakage, fabricated facts, incorrect calculations, missing required warnings, or unauthorized actions as hard gates. A draft that fails one cannot ship even if its style scores highly.

Write gates before comparison. Otherwise, a polished answer can tempt you to relax the standard after seeing it.

Read V1 and V2 in alternating sections rather than relying on memory. Highlight additions, removals, and changed claims. Ask:

• Did the requested change occur?
• What useful content disappeared?
• Did any number, name, attribution, or commitment change?
• Did the revision introduce a new assumption?
• Is the improvement robust or merely different wording?

For structured data, use field-by-field comparison. For long documents, compare outlines first, then high-risk passages.

A reusable prompt needs more than one success. Prepare a small test set representing normal, difficult, and incomplete cases. A meeting-summary prompt might receive clear notes, notes with no owners, contradictory deadlines, and a transcript containing irrelevant personal conversation.

Score each result with the same rubric. Look for patterns. If the prompt performs well only when every field exists, revise its missing-information rule. If it handles short notes but drops details in long input, reduce input scope or split the workflow.

An output can look good on screen and fail downstream. JSON must parse and match the schema. A checklist must fit the actual process. An email must preserve the relationship and make the ask clear. A study explanation must improve recall rather than simply feel easy.

When possible, test the next step: parse the data, have a colleague locate the decision, answer questions from the summary, or complete the task using only the instructions. Human behavior provides stronger evidence than aesthetic preference.

An assistant can help apply a rubric, but it should not be the sole judge of its own answer. It may repeat the same mistaken assumption or reward surface fluency. Human review is essential when facts, context, relationships, or consequences matter. External sources are necessary for factual claims.

If you ask AI to compare versions, require criterion-by-criterion evidence and then inspect it. Do not accept a confident total score without quotations.

Keep a short experiment log: task, input version, prompt change, expected effect, rubric result, regressions, and decision. For personal tasks this can be two lines. For a team template, it becomes useful documentation.

Select the version that passes gates and best balances the criteria. Sometimes that is V2 rather than V5. Iteration is not a staircase where every later rung is higher.