AgentSkillsCN

skill-reflection

当用户要求对当前对话进行反思时，包括识别已使用哪些技能、为现有技能提出通用改进意见，以及建议新的可重复技能或命令时，务必启用此技能。

SKILL.md
--- frontmatter
name: skill-reflection
description: MUST be activated when the user asks for a reflection on the current conversation, including identifying which skills were used, suggesting general improvements to existing skills, and proposing new repeatable skills or commands.

Skill Reflection

Use this skill to produce a high-signal retrospective on the current conversation focused on skills: what was used, what was missing, and what would make future runs faster and safer.

Output Contract

Return a short, structured reflection. Prefer bullets over prose.

Include these sections when applicable:

  1. •Conversation recap (1–3 bullets): what the user wanted and what happened.
  2. •Skills used: list explicitly activated skills and any implicitly-triggered ones (name + evidence).
  3. •Skill improvements: actionable recommendations to update existing skills.
  4. •New skills to consider: candidates for repeatable workflows discovered in this conversation.
  5. •Commands to consider (optional): whether a slash command would reduce friction.

How to Identify Skills Used

  • •Explicit usage: look for direct skill activation (for example, skill: <name> tool calls) and commands that instruct activation.
  • •Implicit usage: infer from the agent’s behavior and/or command prompts (for example, a commit message workflow likely used commit-messages).
  • •If uncertain, say so explicitly (for example, “Likely used general-software-engineering (pattern match), but not explicitly activated”).

How to Recommend Skill Improvements (Keep It General)

Your goal is to improve the skill package, not to solve the one-off task.

When suggesting a change, include:

  • •Problem pattern: what repeated friction or failure mode appeared.
  • •Proposed change: what to add/remove/clarify in the skill (description triggers, workflow steps, safety constraints, validation commands, templates).
  • •Why it helps: how it reduces ambiguity, prevents mistakes, or saves iterations.

Guardrails:

  • •Avoid recommendations that are overly specific to a single repo, file path, or one-off request.
  • •It’s fine to cite the conversation as an example, but phrase the change as a general technique.
  • •Prefer small edits that preserve the skill’s intent and minimize token bloat.
  • •If the gap is better solved by a command (argument validation + small context blocks), say so.

How to Propose New Skills

Propose a new skill only when it clears this bar:

  • •The workflow is likely to recur.
  • •The conversation showed repeated friction (clarification loops, missed context, or recurring mistakes).
  • •The fix can’t be captured as a small edit to an existing skill.

For each candidate skill, provide:

  • •Candidate name: kebab-case.
  • •Scope suggestion: global if broadly reusable; local if repo-specific.
  • •Activation CTA: one sentence in the form “MUST be activated when…”.
  • •Core workflow (3–7 bullets): the minimal steps the skill should encode.
  • •Non-goals: what it explicitly should not do (to prevent overreach).

When to Recommend a Slash Command

Recommend a new command when it would:

  • •Standardize a common argument shape (reduces clarification loops).
  • •Reliably inject minimal, high-signal context (for example, git status, git diff, recent logs).
  • •Enforce a safe workflow (for example, validation checks before edits).

If recommending a command, specify:

  • •Command name (kebab-case).
  • •Expected $ARGUMENTS shape (or explicitly “no arguments”).
  • •2–4 context blocks (!), each small and directly relevant.

Scope

This skill is about reflection and recommendations. Do not perform code edits unless the user explicitly asks you to apply the proposed changes.