Deep Recon

Deep Recon

Research swarm that scours the internet for existing solutions before you build anything.

Time budget: 10 min total. Abort if no useful results after 5 min of searching.

When to Use

• Before building anything non-trivial
• When brainstorming identifies a hard problem
• When you suspect the problem has been solved before
• Before writing an implementation plan
• When evaluating build vs buy decisions

Quick Start

/recon [problem description]

Examples:
/recon real-time collaborative editing for a web app
/recon email deliverability monitoring system
/recon AI agent orchestration with tool use

The problem description is passed as $ARGUMENTS. If empty, ask the user what to research.

The Process

Phase 1: Decompose (Lead Agent)

Parse $ARGUMENTS into 3-5 searchable questions. Also identify:

• Domain: What field is this in?
• Frameworks: What tech stack constraints exist?
• Core challenge: What's the hard part?

Phase 2: Dispatch Research Agents (parallel)

Launch 3 agents in parallel using the Task tool with subagent_type: "general-purpose" and model: "sonnet".

Agent 1 — Solution Scout: Find packages, repos, and services.

• WebSearch for "[problem] library [framework]", "[problem] open source github", "[problem] SaaS API"
• WebFetch top results for details (stars, last commit, license)
• Output: Ranked list with name, URL, pros/cons, maintenance status

Agent 2 — Community Intel: Find real developer discussions.

• WebSearch targeting reddit.com, news.ycombinator.com, stackoverflow
• WebFetch top 3-5 threads for detailed reading
• Output: Community consensus, pitfalls, "I wish I'd known" quotes

Agent 3 — Doc Collector: Find docs, tutorials, educational content.

• WebSearch for "[solution] docs", "[problem] tutorial [year]", "[problem] best practices"
• WebFetch top doc pages for summaries
• Output: Curated links ranked as must-read / useful / deep-dive

Phase 3: Code Analyzer (sequential, after Phase 2)

This runs AFTER Phase 2 because it needs Solution Scout results as input.

Launch one agent (sonnet) to analyze the top 2-3 solutions found:

• Use WebFetch on GitHub repos to read READMEs and key source files
• Use WebSearch for "[package] architecture" or "[package] how it works"
• Report: Architecture pattern, key abstractions, how they handle the core challenge, adoptable patterns

Phase 4: Synthesize (Lead Agent)

Merge all agent outputs into a research brief using the template at templates/research-brief.md.

Save to: /research/YYYY-MM-DD-[topic]/research-brief.md

Be opinionated — recommend ONE approach and defend it with evidence.

Phase 5: Feed into Planning

The research brief becomes input for writing-plans:

• Package solves 80%+ → recommend using it
• No good solution → plan includes key patterns from research
• Open questions → items to resolve during planning

Error Handling

• No search results: Broaden terms, drop framework constraint, try synonym. After 2 empty rounds, note gap and move on.
• Agent returns garbage: Discard, note in brief as "insufficient data for [area]".
• Problem too broad: Split into sub-problems, recon each separately (max 2 splits).
• All solutions are bad: That's a valid finding. Brief should say "build from scratch" with rationale.

Success Criteria

• [ ] At least 3 existing solutions evaluated (or documented why fewer exist)
• [ ] Community sentiment captured
• [ ] Recommended approach with justification
• [ ] Research brief saved to artifacts
• [ ] Open questions identified

Integration Points

| Skill | Relationship | |-------|-------------| | brainstorming | Triggers recon for non-trivial problems | | writing-plans | Receives research brief as input | | app-teardown | Code Analyzer can use teardown methods for deep analysis | | package-research | Solution Scout can delegate to this for npm/pypi specifics | | compound | Analysis phase triggers recon when needed |

Model Routing

• Lead agent (decompose + synthesize): opus
• Solution Scout / Community Intel / Doc Collector: sonnet
• Code Analyzer: sonnet

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Lucface
• Source: Lucface/claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.