Repository navigation
[FYI] Automated, natural language Skillopt skill #236
Description
I built a skill (/skill-opt) to automate the entire evaluation, reflection, and patch deployment cycle through an interactive, natural language conversation with your agent.
Features
- Generates a self-contained Python optimizer on the fly using standard libraries (no pip install or external repo needed).
- Probes recent session transcripts to turn real developer corrections into regression benchmarks.
- Optimize a single file or point it at an entire folder of skills to tune them all in sequence.
- Can run on a recurring schedule to continuously optimize skills against recent real-life agent runs.
- Only commits patches with measured score gains on held-out tasks.
- Runs across Gemini, Claude, OpenAI, or OpenRouter with live progress updates in chat.
- Saves timestamped backups before modifying source files.
LMKWYT!
Activity
Yif-Yang commentedon Sep 30, 2026
PR #267 proposes a narrow implementation through the bundled Claude Code command and existing Sleep pipeline. It preserves explicit staging/adoption rather than adding a generated optimizer or a second execution engine.
The PR has not changed since the last review. The remaining request is the routing acceptance table and safe command traces already listed there, particularly quoted explicit arguments, ambiguous targets/lookback periods, scope changes, and separation of preview from adoption or persistent scheduling. CLA is also still queued. Python test results alone do not validate agent interpretation of these Markdown instructions. No new implementation or delivery date is committed by this status update.