/steal — Pattern Extraction from Open Source Repos

What It Does

The /steal skill takes a GitHub repo URL (or list of URLs), clones each repo, and runs a three-phase analysis:

1. BS Detection — Identifies hype, vaporware, and over-engineering. Flags empty abstractions, TODO-heavy code, unused exports, and marketing-driven architecture that doesn't deliver.
2. Gold Mining — Extracts the actually useful patterns: data models, scoring formulas, protocol specs, guardrail configs, tool loops, adapter patterns. The stuff you'd want to copy into your own codebase.
3. Steal Plan — Produces a reference doc with extracted code organized by category, ready to adapt.

How to Use It

/steal https://github.com/org/repo-name

Or multiple repos at once:

/steal https://github.com/org/repo1 https://github.com/org/repo2

What Claude Does

1. Clones each repo to /tmp/ (shallow clone, depth 1)
2. Scans for key file types: .ts, .py, .rs, .go, .md, .json
3. Ignores: node_modules/, dist/, build/, .next/, vendor dirs
4. Reads README, docs, and source files
5. Runs BS Detection:
   • How much is real code vs. boilerplate/config?
   • Are the abstractions justified or premature?
   • Is the README writing checks the code can't cash?
   • Are there TODO/FIXME/HACK comments everywhere?
6. Runs Gold Mining:
   • Identifies reusable patterns (scoring formulas, type systems, protocols)
   • Extracts the actual code, not just descriptions
   • Notes which patterns are battle-tested vs. experimental
7. Produces a Steal Plan:
   • Organized by category (guardrails, memory, agents, protocols, testing)
   • Each pattern includes: the code, what it does, how to adapt it
   • Saves to a *-stolen-patterns.md file in the memory directory

Output

• A reference doc saved to ~/.claude/projects/<project>/memory/<name>-stolen-patterns.md
• A summary in chat with the key findings

Example: ScoutOS Analysis (March 2026)

Repos Analyzed

• scoutos-labs/hive — Agent orchestration with spawn guardrails + ACP protocol
• scoutos-labs/agent-runner — Tiered agent execution with model adapters
• scoutos-labs/beacon — Knowledge memory with hybrid scoring (semantic + FTS + recency)
• scoutos-labs/context-inspector — Prompt ablation testing (which context segments matter?)

BS Detection Results

• hive: Solid. Real guardrails with env-configurable limits, proper allowlist security, well-specified ACP protocol. The desktop app is early but the core spawn service is production-quality.
• agent-runner: Clean and minimal. 20-turn tool loop, multi-backend adapter pattern (OpenAI/Ollama/Claude), JSON manifest system. No bloat.
• beacon: Real scoring formula with proper decay math. Graph-based memory with edges. The "intelligent prune" is Phase 2 vaporware but the rest ships.
• context-inspector: Genuinely novel. Ablation testing for prompts — remove each context segment, measure behavioral drift via embeddings or LLM-as-judge. Small codebase, high value.

Gold Mined

See scoutos-stolen-patterns.md for full extracted code. Key wins:

• Spawn guardrail config pattern (env-tunable limits)
• Command allowlist with arg validation
• Memory scoring formula: (semantic * 0.6 + fts * 0.3 + recency * 0.1) * attention_boost
• Recency decay: exponential with 30-day half-life
• Tiered agent manifests (haiku/sonnet/opus presets)
• ACP protocol message types (task, progress, clarification, response)
• Ablation engine: cosine-delta + LLM-as-judge scoring

Tips

• Run /steal on repos BEFORE meetings about potential tools/partnerships — know what's real
• The patterns doc is meant to be a living reference — update it as you adapt patterns into your own code
• Best targets: repos with < 5000 LOC that solve one problem well
• Skip repos that are mostly config/boilerplate — there's nothing to steal