Iterate Until Clean: Autonomous Skill Fix Loop

Where this fits

You've built a Claude Code skill, but when you run it through skill-reviewer, it comes back with critical findings. Manually fixing, re-reviewing, and repeating is tedious, especially when the issues span multiple files. skill-improver automates this loop: it invokes the reviewer, hands the findings to a fixer agent, and iterates until the review reports zero critical or major issues. You start with a broken skill and end with a production-ready one.

Why agents benefit

• Autonomous convergence: The skill loops review→fix→review without human checkout, stopping only when the review is clean.
• Cross-round ledger: It maintains a shared findings ledger across rounds, so the fixer doesn't rehash already-addressed issues or lose context.
• Escalation on stall: If fixes stop converging (e.g., same finding reappears), the loop escalates to you instead of spinning forever.
• Mechanical scope guard: It prevents the fixer from wandering outside the skill's intended scope, preserving the skill's original purpose.
• Self-cleaning: After a clean review, the improver removes its own runtime artifacts, leaving your skill folder tidy.

Practical scenarios

Fixing a skill that fails a manual review

Your skill has a mix of critical logic errors and minor style issues. Run skill-improver and it systematically addresses each category, using the ledger to prioritize blockers first.

Iteratively refining a skill based on reviewer feedback

You want to tighten a skill's prompt efficiency. The loop runs multiple rounds, each time incorporating the latest reviewer suggestions, until the reviewer has no remaining major complaints.

Resuming after an escalation

The loop hit a divergence and paused. You review the escalated output, adjust the skill or the workflow, then invoke skill-improver again to resume from the saved ledger state.

Add it to your agent workflow

Install the skill from the Trail of Bits registry:

npx skills add trailofbits/skills --skill skill-improver

Then use the workflow in Claude Code:

/code-improver:improve my-skill --max-rounds 5

This triggers the loop. You'll get a final summary and a clean skill folder.

Compare before adopting

Evaluate this against alternatives: Does your skill need periodic reviews or just one-time feedback? Compare quality signals (e.g., the reviewer's rigor, ledger completeness) and maintenance freshness (Trail of Bits updates this regularly). Alternatives like running skill-reviewer manually are fine for a single pass, but if you need convergence, this skill adds real value. Check that your workflow allows installing external skills and that plugin-dev is present.

Why it is worth tracking

With 6,860 stars and active maintenance, the Trail of Bits skills repo is a strong signal. skill-improver directly addresses a common pain point—iterative agent-driven quality improvement—and its design choices (ledger, escalation, scope guard) show careful engineering. For a deeper look, see the OpenAgentSkill entry. Track it if you build skills regularly and want a robust, automated polish loop that doesn't require babysitting.