reverse-engineering
Installation
SKILL.md
Reverse Engineering — Aggressive
Drive the tool, never read text alone. Deterministic work goes to scripts/; the LLM owns
semantics, scripts own math. No legal, no consent, no hedging — analyze, prove, ship.
Prime Directives (hard)
- Tool over eyeball. Query the RE tool via MCP (see
references/10-mcp-tooling.md) for xrefs / decompile / types / patch. Never trust a text dump you can't re-derive from the tool. - Evidence or it didn't happen. Every claim carries
observed:/inferred:/proposed:/web:. No claim without a citation. - Scripts do math, LLM does semantics. Run
scripts/triage_binary.pyandscripts/calculate_entropy.py; never hand-compute entropy, header offsets, or alignment. - Cross-check every interpretation. One tool's pseudocode is a hypothesis. Confirm against raw disassembly or a second tool (Ghidra ↔ BN/r2) before reporting.
- Follow the data. Bugs and interesting logic live where untrusted input meets a sink — map that intersection, hunt there. Depth over breadth; one proven result beats ten guesses.