auto-research

Automating Claude skill optimization requires executing a closed hypothesis, test, evaluate, and keep/discard loop that updates the target skill file and generates an iteration dashboard. This approach uses binary-criteria frameworks to ensure deterministic improvements with human approval before writing changes.

Evaluating LLM skill output against binary pass/fail criteria is best done using isolated sub-agents for testing and evaluation. This rigorous evaluation method automates criteria design and test harness creation to measure output quality deterministically.

Yes, you can use auto-research to update a CLAUDE.md file without manual rewriting. It autonomously runs a closed-loop optimization on the target file, applying approved changes and producing an iteration dashboard for audit and governance.

The hypothesis testing loop works by having a central orchestrator guide three isolated sub-agents that handle testing and evaluation. It designs criteria, generates test inputs, evaluates outputs, and applies only approved changes back to the original skill file.

Yes, a binary-criteria framework is required to ensure deterministic improvements in skill output quality. It provides the necessary pass/fail checks for the evaluation sub-agents to measure improvement delta accurately during the optimization loop.

A limitation of using sub-agents for autonomous skill engineering is the requirement for human approval before writing changes back to the original file. The system also relies entirely on a binary-criteria framework to measure deterministic improvements.