skill-innovation-retrospective

• Detect skill gaps, quality issues, user suggestions, or repeated friction.
• Recommend updating an existing skill or creating a new one.
• Never modify or create skills automatically.

• Task finished (final output delivered or conversation clearly done).
• Task blocked or issues encountered (errors, tool failure, uncertainty, partial completion, or multiple clarification loops).
• Quality feedback (user dissatisfaction, redo/refactor request, "wrong/missed/didn't follow").

• Quality: negative sentiment, format mismatch, missing constraints, incorrect facts/outdated info,

missing citations where required, hallucination risk.

• Process: too many clarifying questions, forgot constraints, no plan/options when needed.
• Tooling: tool errors, wrong tool usage, missing troubleshooting steps.
• User contribution: suggested better workflow/template/tone.
• Additional must-include cases: unclear skill steps, insufficient edge cases, cross-skill conflict,

recurring task type, domain nuance, delivery-model nuance.

• Observed Signals (bulleted, brief paraphrase).
• Root Cause Hypothesis (1-3 bullets).
• Recommendation: update existing skill or create new skill.
• Proposed Change Summary: additions/removals/clarifications; checklist/guardrails/examples;

new skill scope/triggers/standardization.

• Ask for Confirmation (mandatory): "Do you want me to run $skill-creator to (create/update)

the skill now?"

• Keep concise (8-20 lines unless user asks for more).

• Update if within scope of existing skill and needs clarity/guardrails/examples/edge cases.
• Create if no existing skill fits or pattern recurs and needs standardized structure.

1) Identify target: existing skill name/path or proposed new skill name/path (e.g.,

.codex/skills/.md).

2) Prepare change package: title, problem statement, triggers, inputs/outputs, step-by-step

algorithm, guardrails/anti-patterns, at least one positive and one negative example, and a

Definition of Done checklist.

3) Confirm again if scope is large or impacts many conversations.

4) Execute $skill-creator; summarize changes and improvements.

• Never run $skill-creator without explicit confirmation.
• Never include sensitive or personal data in examples.

• Don't derail the main answer; run only at end or on issue/feedback.
• Don't produce long essays or blame the user.
• Don't propose changes without evidence.
• Don't auto-edit skills without approval.