AgentSkillsCN

reflect

回顾近期工作,提出可优化未来工作流程的新技能、新钩子、新规则,或新代理建议。

SKILL.md
--- frontmatter
name: reflect
description: Review recent work and suggest new skills, hooks, rules, or agents that could improve future workflows
argument-hint: "[optional focus area]"
allowed-tools: Bash, Read, Grep, Glob, AskUserQuestion

Reflect

Analyze recent activity and the current Claude Code setup to suggest workflow improvements.

Steps

Step 1: Gather Context

Read the following in parallel:

  • ~/CLAUDE.md — global instructions
  • ~/.claude/settings.json — permissions, hooks
  • ~/.claude/skills/*/SKILL.md — all global skills
  • Recent git history in the current project (if in a git repo): git log --oneline -20
  • Project-level CLAUDE.md, .claude/rules/, .claude/skills/, .claude/agents/ (if in a project)

If the user provided $ARGUMENTS, use it to focus your analysis on that area.

Step 2: Analyze

Look for opportunities in these categories:

TypeSignalExample
SkillA multi-step process that's been done manually more than once"You frequently run build + test + lint in sequence — extract a /check skill"
HookA repetitive action that should happen automatically on an event"You always run tests after editing — add a PostToolUse hook on Edit"
RuleA convention or preference that's been corrected or stated verbally"You keep reminding Claude about your test framework — add a .claude/rules/ file"
AgentA complex investigation pattern that benefits from isolated context"Deep architecture analysis keeps filling context — extract an agent"
CLAUDE.mdA behavioral instruction that applies globally or project-wide"You always want snake_case — add it to CLAUDE.md"

Also check for:

  • Existing skills/hooks/rules that might be stale or redundant
  • Global settings that could be tightened or relaxed
  • Patterns in git history that suggest missing automation

Step 3: Present Suggestions

Present your findings as a concise list, grouped by type. For each suggestion, number it and include:

  • What you observed
  • What you'd create or change
  • Why it would help

Step 4: Select

After presenting the list, use AskUserQuestion with multiSelect: true to let the user pick which suggestions to implement. Each option should be:

  • label: Number and short label (e.g. #1 /check skill, #3 Auto-lint hook)
  • description: One-line summary of what it does

The user can also use "Other" to modify instructions or refine specific items before proceeding.

Step 5: Implement

For each selected suggestion, implement it. If the user provided custom instructions via "Other", follow those. After implementing, briefly summarize what was created or changed.