AutoSkill

Problem

Reusable insights often appear in model conversations: user corrections, durable preferences, failure fixes, and repeated setup. These notes usually remain trapped in chat history instead of becoming maintainable local skills. Creating skills immediately can create duplicates, over-triggering, and unconfirmed behavior changes. AutoSkill turns when to extract, whether to keep, how to improve, and where to write into a lightweight, non-blocking decision process.

How it works

It operates on local SKILL.md files and explicit skill roots, without requiring databases, vector stores, daemons, or schedulers. It processes candidate experience through a staged path:
- Silent scan: trigger an extraction check when the user corrects an approach, states a durable preference, reveals a reusable fix, repeats non-trivial setup, or asks to save a workflow.
- Candidate evaluation: assess evidence, value, priority, recurrence, and check for similar existing skills.
- Confirmed write: create, update, merge, or delete a skill only after the candidate passes the reuse boundary, the target path is clear, and the user approves the full proposal or diff.
For ambiguous topics, it first presents 2-5 candidate titles with one-line evidence and asks the user to choose or define a custom topic.

Boundaries

It is suited to maintaining a personal or small-team local skill library, not to automatically installing third-party capabilities. Before writing, it should surface the target path, evidence, duplicate check, full SKILL.md content, or unified diff. Scripts, network access, credentials, destructive commands, or shared policy changes require explicit consent. One-off outputs are not promoted, and it should not silently modify itself during routine maintenance.