research

claude-research

Deep research to eliminate all assumptions before implementation.

What It Does

Normal Claude: "I'll use library X for this..." (assumption)

With /research: Verifies every assumption through web search and codebase exploration before proceeding.

How It Works

1. Draft approach - Outline initial plan
2. Extract assumptions - Flag everything unverified
3. Research each - WebSearch, codebase exploration, documentation
4. Log findings - Document evidence in state file
5. Update approach - Adapt plan based on findings
6. Re-scan - Look for new assumptions, repeat until none remain

Installation

Via Marketplace

/plugin marketplace add nrempel/claude-plugins
/plugin install research@nrempel-plugins

Local Development

claude --plugin-dir ~/Projects/claude-research

Usage

/research Add authentication to the API

Commands

Command	Description
/research <task>	Start new research session
/research	Continue existing session, or prompt for task
/research status	Show assumptions and progress
/research cancel	Stop and clean up

To stop mid-research: Say "stop", "done", or "that's enough".

Example

> /research Add Stripe subscription billing

[Drafts initial approach]
- Use Stripe Billing API
- Webhooks for events
- Store status in DB

[Extracts assumptions]
- [ ] Stripe supports usage-based billing
- [ ] Project has webhook infrastructure
- [ ] DB can store subscription metadata

[Researches each]
✅ Stripe supports metered billing (verified in docs)
❌ No webhook infrastructure exists (searched codebase)
✅ Postgres with Drizzle - can add columns

[Updates approach based on findings]
Need to add webhook endpoint - wasn't in original plan

[Re-scans for new assumptions]
- [ ] Next.js API routes can handle webhooks
→ Researches, verifies

## Research Complete

Verified approach with 0 assumptions:
1. Use Stripe metered billing API
2. Create /api/webhooks/stripe endpoint (NEW)
3. Add subscription_id column to users table
...

Compaction-Safe

This plugin survives context compaction:

State file: Creates .claude/research/<branch>-<timestamp>.md containing:

• Original goal
• Current approach
• Assumptions (verified/unverified/refuted)
• Research log with findings

Resume: Run /research again after compaction. It will:

1. Find the state file
2. Read the goal and progress
3. Continue from unverified assumptions

Cleanup: State file is deleted when research completes.

Gitignore

Add to your .gitignore:

.claude/research/

What Gets Verified

Type	How
External libraries	WebSearch for docs, npm/pypi
API behavior	Official documentation
Codebase patterns	Grep, Glob, Read tools
Version compatibility	Changelogs, compatibility docs
Service limits	Official docs, pricing pages

Plan Mode Integration

This plugin integrates with Claude Code's plan mode:

1. Hook on plan mode entry: When you enter plan mode, the plugin displays a reminder:

💡 Tip: Run /research to eliminate assumptions before finalizing your plan.

2. Writes to plan file on completion: When research completes, verified findings are written to the active plan file at ~/.claude/plans/. This ensures your plan is backed by verified research.

vs /iterate

/iterate	/research
Execution	Planning
Does the work	Verifies the approach
Tests/builds verify	Docs/code verify
After planning	Before execution

Recommended workflow:

1. Enter plan mode
2. Run /research to verify approach
3. Finalize plan with verified facts
4. Exit plan mode
5. Run /iterate to execute

License

MIT