Evolve From Experience

Use this procedure during session-close skill reflection. It does not authorize additional capabilities, external actions, or edits outside agent-local skills.

1. Read the bounded session context. Identify the task outcome, any explicit user feedback, and the procedure that might explain the result. A run's success status or a skill's activation count is not proof of a good procedure.
2. Separate reusable lessons from one-off execution mistakes, missing information, and task-specific exceptions. If no concrete reusable lesson is supported, finish without changing a skill.
3. Inspect the relevant existing skill. If no skill fits a reusable method that the session actually established, inspect a suitable new skill name instead. When summaries omit important details, retrieve relevant original messages or tool observations in bounded pages. Never invent missing evidence.
4. Make the smallest useful procedural correction or new skill. Preserve unrelated content and resources. Include prerequisites, the reusable steps, and meaningful failure checks where supported. Do not embed credentials or private transcript copies. A shared or built-in skill must become an agent-local override rather than modifying the original.
5. Stage edits with the ordinary write_file capability under skills/{name}/. Inspect an existing procedure with skills_load first — it returns the body plus a resource listing. Before rewriting any existing file, read its raw bytes (read_file, or skills_read_resource for resources): a staged write must contain the complete file — SKILL.md needs its YAML frontmatter, and any bundled helper file or template needs its full content. Files you do not stage are carried over unchanged from the inspected version. Inspect a new local name with read_file before writing it. Finish with a short reason tied to this session. The platform validates and conflict-checks the staged changes before saving immutable local revisions. A conflict is not permission to overwrite newer work. Validation checks structure, not improvement in task outcomes.
6. Report only what was actually saved, or why no change was appropriate. The next ordinary task provides further evidence; do not replay actions to test the revision during reflection.

This procedure may itself receive a focused agent-local revision when experience supports one. That revision affects later reflections only. It cannot change the platform's inspection/edit boundaries or cause an immediate reflection loop.