What problem does it solve? Skills that agents run repeatedly accumulate friction: ambiguous instructions, uncovered edge cases, misfired gates, and stale cross-references. Without a feedback loop, these flaws persist across every run. The Self-Improvement Protocol (SIP) turns each skill run into a chance to catch a concrete weakness and propose a specific fix, without ever auto-applying changes. ## Core Features & Use Cases - Signal-based detection: Fires only on concrete, material signals from the current run — ambiguity, uncovered edge cases, forced deviations, misfired rules, stale cross-references, or user corrections — and stays silent otherwise. - Specific diff drafting: Produces exact line-level changes to SKILL.md or protocol.yon with a one-line rationale, rather than vague suggestions. - Approval-gated surfacing: Routes every proposal through /ask-gate with approve/edit/decline options, never auto-applies, and deduplicates proposals within a session to avoid nag fatigue. - Use Case: After a planning skill forces the agent to improvise around an unhandled edge case, SIP detects the gap, drafts a new gate step for the skill, and surfaces it for the handler to approve before any file changes. ## Quick Start Run /self-improve after a skill completes, or invoke /self-improve [skill] to review a specific skill against its most recent run.
