Skill optimizer
AdvancedLocation: /studio/[project]/skillopt
The SkillOpt (AI skill optimization) tab automatically iterates on a user-edited 'skill document' (a scenario-improvement instruction). The optimizeSkill loop generates skill candidates each step, applies them to the current scenario, scores via the existing evaluator (/api/owntent/evaluate), and accepts an edit only when the held-out (craft) axis improves. With chosen epochs and an LR budget it shows initial-to-best scores and a per-step accept/reject log, and you can save the best skill or apply it to the project brief.
- 1Open the SkillOpt tab and edit the skill document (improvement instruction) or seed it from the brief.
- 2Choose epochs and an LR budget, then run Optimize.
- 3Watch the per-step accept/reject log and the initial-to-best score change (cancellable mid-run).
- 4Name and Save skill, or Apply to brief to use it in the project.
- 5Reload a saved skill to re-optimize, or delete it.
Scoring during optimization uses the same LLM backend as Evaluate: CLI agent (Codex/Claude Code) preferred or BYOK (OpenAI/Anthropic/Google). Saved skills are stored locally per project.
Optimization requires both a scenario and an LLM backend and incurs cost per evaluation step; start with small epochs/LR budget to confirm the score actually rises.