Auto-Improve

Most skills, agents, and documentation work about 70% of the time. The other 30% produces inconsistent, shallow, or wrong output. The fix is not a full rewrite — it is letting an autonomous loop run the target repeatedly, score every output against binary criteria, tighten the prompt until that 30% disappears, and keep a complete research log of every mutation attempted.

Memories are different: they degrade silently. Facts go stale, gaps accumulate, entries duplicate. The fix is a structured audit followed by targeted rewrites.

This skill handles both patterns under one entry point.

It is not request-routed. The trigger is what the agents actually did: files changed, mistakes repeated, user directions clarified, workflows that felt awkward, docs that were missing, and gaps between expected behavior and actual behavior.

It should behave like a lightweight hyperagent, not a one-shot optimizer:

• the task-level target and the meta-level improvement procedure are treated as part of the same editable system
• successful intermediate variants are stepping stones, not disposable artifacts
• memory should store synthesized insights, causal diagnoses, and forward-looking plans, not just scores
• lessons from one run should be reusable in later runs and across nearby targets

entry point