autoresearch

ai-plugins

Early work-in-progress. These plugins work and produce high-quality output, but the experience is rough around the edges. Expect lots of permission prompts to accept, some UI jank, and a workflow that isn't yet polished. The results, however, are genuinely good.

A monorepo of Claude Code plugins that make Claude research before it codes.

Plugins

/research — autoresearch

Autonomous research-driven coding. Instead of jumping straight to implementation, Claude researches multiple options, scores them, presents you with a ranked comparison, and only implements after you pick a winner.

Inspired by Andrej Karpathy's autoresearch concept and pi-autoresearch by davebcn87.

Pipeline:

1. Brief — Scans your codebase, discovers 3-5 candidate approaches, presents a brief for confirmation
2. Research — Dispatches parallel research agents (one per candidate) that deep-dive into docs, bundle size, maintenance health, API design, community activity
3. Evaluate — Synthesizes all research into a ranked recommendation with trade-off analysis
4. Decide — You pick the winner (or accept the recommendation)
5. Implement — Executes in an isolated git worktree so your working tree stays clean
6. Verify — Runs tests, typecheck, lint, bundle analysis
7. Auto-fix — If verification fails, automatically retries (up to 3 attempts)
8. Report — Final summary with the branch ready to merge

Example:

/research replace moment.js with a lighter alternative

/spec — autospec

Spec-driven coding. Paste a spec from anywhere — Google Docs, Jira, Slack, an email from your boss — and Claude will review it for gaps, ask clarifying questions, then implement it.

Pipeline:

1. Intake — Analyzes the spec, classifies it as prescriptive / directional / open-ended, identifies gaps
2. Interrogation — Asks you targeted questions to fill gaps (skipped if spec is airtight)
3. Routing — Based on prescriptiveness:
   • Prescriptive → straight to implementation
   • Directional → light research on specific technical questions
   • Open-ended → full research pipeline (delegates to autoresearch agents)
4. Plan — Presents an implementation plan for approval
5. Implement — Executes in an isolated git worktree
6. Verify — Tests, typecheck, lint, bundle analysis, auto-fix on failure

Example:

/spec Here's the PRD my PM sent: [paste spec]

Installation

Prerequisites

• Claude Code CLI installed and working
• A Claude Code plan that supports plugins (Pro, Team, or Enterprise)

Install the plugins

From your terminal:

# Install autoresearch
claude plugin add /path/to/ai-plugins/autoresearch

# Install autospec
claude plugin add /path/to/ai-plugins/autospec

Or if published to a registry:

claude plugin add autoresearch
claude plugin add autospec

After installing, the /research and /spec slash commands become available in any Claude Code session.

Verify installation

# Should list both plugins
claude plugin list

Usage

Basic usage

# Start a Claude Code session in your project
claude

# Run autoresearch
/research replace our REST client with something better

# Run autospec
/spec implement user authentication with OAuth2 and refresh tokens

What to expect (honest version)

1. Lots of permission prompts. The plugins dispatch multiple subagents that read files, search the web, write to disk, create git branches, and run shell commands. You'll be accepting tool permissions frequently. This is the biggest UX pain point right now.

2. It takes a while. A full research pipeline with 4 candidates can take 5-10+ minutes. It's doing real work — web searches, documentation analysis, codebase scanning — but it's not instant.

3. The UI can be noisy. You'll see agent dispatches, state file writes, and structured log updates scrolling by. The intermediate output isn't pretty, but it's transparent.

4. The output is genuinely good. Despite the rough experience, the final implementation quality is high. Research is thorough, trade-off analysis is real, and implementation happens in isolated branches so nothing breaks your working tree.

Session management

Both plugins maintain state in their respective directories (.autoresearch/ and .autospec/). Sessions are resumable — if Claude Code crashes or you close the terminal, the plugin picks up where it left off on next launch.

# Clear a session and start fresh
/research clear
/spec clear

Configuration (optional)

Create a config file to override defaults:

# For autoresearch
echo '{"maxRetries": 5, "maxResearchers": 6}' > .autoresearch/config.json