generate

generate — Research Skill Generator

Scaffold a standalone research skill customized for a specific domain.

When to Use

• You want a permanent research skill for a specific review niche
• You want to customize the loop behavior for your workflow
• You want a skill that works independently (without auto-research installed)

When NOT to Use

• You want to run a one-off research loop (use /research-loop directly)
• You want to understand how auto-research works (read the README)

Tool Access

Bash — reading source template files. Read — reading reference materials to inline. Write — creating generated SKILL.md and reference files. Edit — updating existing plugin.json if output target is an existing plugin. All four are minimum necessary for the generator to scaffold a complete standalone skill.

Built-in Templates

Template	Best for
gh-review	GitHub PR review comments — groups by file, outputs PR comments
spec-review	Spec/RFC feedback — groups by theme, outputs markdown report
plan-review	Implementation plan feedback — groups by theme, outputs annotated plan

Process

Step 1: Identify Template

If the user specified a template name in the argument, use it.

If not, or if "custom", ask: "What kind of findings will you be researching? (PR review comments, spec feedback, plan critiques, or describe your niche)"

Step 2: Interview (5 questions)

Ask these using AskUserQuestion, adapting based on the template chosen:

1. Input source: "Where do findings come from?"

   • GitHub PR comments (via gh api)
   • A markdown file with structured findings
   • Jira ticket comments
   • Pasted text in conversation
   • Custom (describe)

2. Output target: "Where should the generated skill live?"

   • New plugin directory (ask for path and name)
   • Existing plugin directory (ask which one)
   • Local .claude/skills/ in a project (project-local, no plugin)
   • Current directory (for testing)

3. Output format: "How should results be delivered?"

   • PR comments (one per cluster)
   • Markdown report file
   • Annotated plan
   • Inline code suggestions
   • Custom (describe)

   Reference ${CLAUDE_SKILL_DIR}/references/output-formats.md for details.

4. Exploration scope: "What should the Research agent explore?"

   • Changed files + imports + tests (typical for code reviews)
   • Full codebase + docs (typical for spec reviews)
   • Referenced files + git history (typical for plan reviews)
   • Custom (describe)

5. Skill name: "What should the skill be called?"

   • Suggest a default based on template (e.g., pr-research, spec-research)
   • Validate: must match [a-z0-9]+(-[a-z0-9]+)*, max 64 characters
   • If invalid, explain the constraint and ask again

Step 3: Read Source Materials

Read these files to assemble the generated skill:

1. Template: ${CLAUDE_SKILL_DIR}/../../templates/{{template_name}}.md
2. Loop protocol: ${CLAUDE_SKILL_DIR}/../research-loop/references/loop-protocol.md
3. Agent prompts: ${CLAUDE_SKILL_DIR}/../research-loop/references/agent-prompts.md
4. Skill template: ${CLAUDE_SKILL_DIR}/references/skill-template.md

If any file is not found, stop and report which file is missing. This likely means the auto-research plugin directory was moved or partially installed. The user should reinstall via claude plugin marketplace add or verify their plugin path.

Step 4: Generate Files

Create the output structure:

{{output_target}}/
├── .claude-plugin/
│   └── plugin.json           # Only if output target is a plugin
└── skills/
    └── {{skill_name}}/
        ├── SKILL.md           # Orchestration skill (~250 lines)
        └── references/
            ├── loop-protocol.md    # Full state machine (copied from source)
            └── agent-prompts.md    # Niche-customized prompts

SKILL.md: Use skill-template.md as the structure. Fill placeholders with:

• Niche context from the chosen template
• Input parsing instructions from the template
• Grouping strategy from the template
• Output format instructions from interview answer

loop-protocol.md: Copy verbatim from the research-loop source. Do not modify.

agent-prompts.md: Copy the agent prompt templates, replacing {{niche_context}} placeholders with the niche-specific context from the chosen template.

plugin.json (if applicable):

{
  "name": "{{plugin_name}}",
  "version": "0.1.0",
  "description": "{{niche}} research skill — iterative verification loop"
}

Step 5: Verify

After writing all files:

1. Read back the generated SKILL.md and confirm it references the correct relative paths
2. Verify loop-protocol.md was copied completely
3. Verify agent-prompts.md has niche context filled in (no remaining {{placeholders}})
4. Report to user: "Generated skill at {{path}}. Test with: claude --plugin-dir {{path}}"

Customization Guidance

If the user chose "custom" for any interview question, help them define:

• Custom input parsing: What format are their findings in? How to extract structured records?
• Custom grouping: What dimension should findings be clustered on?
• Custom output: What sections, metadata, and delivery method do they need?
• Custom scope: What parts of their codebase are relevant?

Write their custom definitions directly into the generated skill (no placeholder references to templates that won't exist in the standalone skill).