autorefine

name

autorefine

description

Self-improving content refinement loop with a user-defined AI judge. Use when the user wants to iteratively improve any content (copy, slides, emails, briefs, scripts, headlines) through repeated generate → judge → iterate cycles. Triggers on phrases like "refine this until it's good", "run autorefine", "keep improving this", "iterate on this with a judge", "make this better automatically", or any request to repeatedly improve content against defined quality criteria. The user owns the criteria — this skill ships with zero hardcoded taste.

autorefine

A self-improving loop for any content. You define what "good" looks like. The skill handles the rest.

How It Works

Load criteria → Generate versions → Checklist filter → AI judge scores → Winner saved → Repeat

Every iteration starts from the previous winner. The loop runs until the threshold is hit or max iterations reached. The judge never sees who generated the content — clean separation, no ego.

Setup (First Run)

1. Check if criteria.md exists in the current workspace or project folder
2. If not → copy assets/criteria-template.md to the working directory and ask the user to fill it in before running
3. If yes → load it and proceed

The Loop — Step by Step

Step 1: Generate

• Produce 3 versions of the content using the user's brief
• Variations should differ meaningfully — not just synonyms. Try different angles, structures, tones.

Step 2: Checklist Filter (binary)

• Load the MUST-HAVE section from criteria.md
• Each version either passes or fails — no partial credit
• Versions that fail any must-have are eliminated before scoring
• If all 3 fail → revise all and repeat from Step 1 (counts as one iteration)

Step 3: AI Judge (scored)

• Take surviving versions and score each criterion from the SCORED section in criteria.md
• Score each criterion 1–10, then compute weighted average
• Return: { score, reasoning, pass, improvement_notes }
• Pass threshold: default 7.0 unless user specifies otherwise in criteria.md

Step 4: Decision

• Score ≥ threshold → winner found, exit loop
• Score < threshold → use improvement_notes as input for next generation
• Max iterations default: 5 (user can override)

Step 5: Output

• Winning content (clean, no annotations)
• Iteration log: what was tried, what scored what, why
• Prompt the user: "Want to add anything to criteria.md based on this run?"

Rules

See references/rules.md for the full do's and don'ts. Core rules:

DO:

• Separate generation from judgment — never let the same agent judge its own output in a single pass
• Always run the checklist before scoring — it's the quality floor
• Log every iteration — losing versions are learning data
• After human correction → update criteria.md immediately
• Start each iteration from the previous winner, not scratch

DON'T:

• Use vague criteria like "make it good" — criteria must be testable
• Run more than 7 iterations without a human checkpoint
• Skip the checklist to save time — it exists for a reason
• Discard losing versions — keep the log
• Hardcode your taste into the skill — criteria.md is where taste lives

Taste Accumulation

The judge gets better over time because criteria.md grows:

Base criteria → Run → Human corrects → Append correction → Next run is smarter

Every time the user says "this isn't right because X" → append that as a new criterion:

- [date] Avoid passive voice in headlines (user correction)
- [date] Market size must cite a source, not just a number (user correction)

Over time the judge sounds like the user — because it literally is the user, accumulated.

References

• Rules + do's/don'ts: references/rules.md
• How to write good criteria: references/how-to-write-criteria.md

Assets

• Blank criteria template: assets/criteria-template.md — copy this to your project and fill it in