self-improvement

SKILL: Skill and Agent Self-Improvement

Purpose

Control updates to existing skills and agent profiles so the framework improves from evidence without accumulating noise, redundant rules, or broader prompts. This skill is a Markdown-only refactoring gate for AI operational assets.

Activation Triggers

Load this skill before changing any skills/*/SKILL.md, agent profile, skill catalog, trigger list, or agent-specific operating rule when:

• a skill or agent failed, was ambiguous, did not activate, or produced avoidable rework;
• two or more sessions repeated the same clarification or mistake;
• a validation gap allowed hallucination, duplication, monolith creation, unsafe cleanup, or context loss;
• a new framework rule makes an existing skill incomplete or contradictory;
• a proposed edit claims token savings, better triggers, or safer execution for existing assets;
• the task uses Portuguese equivalents such as melhorar skill, melhorar agente, ajuste de gatilho, gatilho falhou, skill não acionou, agente não acionou, auto melhoria, melhoria de prompt, or regra ambígua.

Inputs

Concrete failure or drift evidence, owning canonical documents, current skill and catalog entry, overlapping skills, representative replay task, and active plan/change record.

Improvement Gate

Improvement is allowed only when at least one evidence metric and all safety metrics pass:

Metric	Pass threshold
Failure evidence	2 independent failures/ambiguities, or 1 P1/P2 incident.
Rule drift	A canonical document changed and the skill now contradicts or misses it.
Validation gap	Existing exit criteria did not detect a real defect.
Token ROI	Expected skill or base-document reduction of at least 15%, or removes repeated clarifying prompts.
Scope preservation	The edit narrows or clarifies scope; it does not expand responsibility without agent-factory.
Backward compatibility	Existing valid triggers and outputs remain valid or have a migration note.