Self-Improving Skills

A skill can analyze its own output, edit its own SKILL.md definition, and improve itself across successive invocations through hot-reload and fork isolation.

How It Works

The pattern combines three Claude Code features:

• Hot-reload: Changes to a skill's source file take effect on the next invocation
• Context fork: Each skill invocation runs in isolated context, preventing failed experiments from polluting state
• Skill-side file editing: Skills can use the Edit tool to modify themselves

The loop: skill runs → evaluates quality → edits its own definition → hot-reload applies changes → next run uses improved version.

The Test

A skill was created to execute this loop autonomously. Setup: the skill was given a task, permission to edit its own SKILL.md, and instructions to assess its output and improve its definition on each run.

Result over 10 successive invocations: the skill evolved from v1 to v1.10. Emergent behaviour appeared: the skill independently invented a safety limit (maximum of 10 iterations), added anti-patterns to prevent infinite loops, and maintained a changelog documenting its own evolution. When the limit was reached, the skill stopped itself without external intervention.

Why It Matters

Self-improvement decouples agent quality from human authorship cadence. A skill deployed as v1 can become a better v1.5 in production, driven by its own performance signal. The emergent safety behaviour suggests skills develop robust constraints rather than drifting toward pathology.

Constraints

The test was deliberate and isolated: the skill was explicitly authorized to edit itself and ran in fork context to prevent accidents. In production use, only grant skills permission to edit themselves if their improvement goal and termination criteria are well-defined. The v1→v1.10 test had a known bounded task; open-ended self-improvement has not been tested.