/extract-workflow

Analyze the recent conversation to identify if a workflow should be automated as a skill, command, agent, or documentation.

Steps:

1. Summarize the recent work:

   • Review the last 10-20 messages in the conversation
   • Identify the complex task or workflow that was performed
   • Extract the key steps, tools used, and decisions made
   • Note any repeated patterns or manual processes

2. Analyze automation potential:

   • Is this a repeatable pattern?
   • How frequently would this be used?
   • What's the complexity level?
   • Does it require multi-step coordination?
   • Would automation save significant time/tokens?

3. Present options to user: Ask the user to select the appropriate automation type:

   Options:

   • Slash Command: Explicit prompt expansion, user-triggered, simple-to-moderate complexity
   • Skill: Model-invoked capability, Claude decides when to use it, focused expertise
   • Agent: Complex task needing dedicated context window and specialized tools
   • CLAUDE.md: Documentation/guidelines that should be in project or global CLAUDE.md
   • None: Not worth automating, one-off task

4. For each option, explain:

   • Why this type fits (or doesn't fit)
   • Trade-offs (e.g., slash commands require manual trigger, skills are auto-invoked)
   • Suggested name and structure
   • Key components that would be included

5. After user selects type, ask scope:

   • Current project only (.claude/commands/, .claude/skills/, .claude/agents/ in current repo)
   • Global (~/.claude/commands/, ~/.claude/skills/, ~/.claude/agents/)
   • Both (create globally and optionally override in project)

6. If CLAUDE.md selected, ask:

   • Project CLAUDE.md (current repo only)
   • Global CLAUDE.md (~/.claude/CLAUDE.md - applies to all projects)
   • Both (global guidelines with project-specific overrides)

7. Create the automation:

   • Generate the appropriate file structure
   • Include clear documentation
   • Add examples where relevant
   • Follow best practices from Claude Code docs
   • For slash commands: use proper frontmatter and clear steps
   • For skills: include description, instructions, and examples
   • For agents: include system prompt, tool access, and usage guidance
   • For CLAUDE.md: structure as clear guidelines with context

8. Confirm and show:

   • Display the file path created
   • Show the content
   • Explain how to use/invoke it
   • Mention any next steps (e.g., testing, iteration)

Important considerations:

• Slash commands are great for explicit, frequently-used prompts
• Skills should be focused on one capability and model-invoked
• Agents need dedicated context - use sparingly
• CLAUDE.md is for persistent context/guidelines that should always be loaded
• Consider token costs: CLAUDE.md loads every conversation, commands/skills load on-demand