reverse-skill: a security research skill router for AI coding agents

Hand an AI coding agent an APK, a binary, or a piece of obfuscated frontend JS, and its first move is usually a guess: jadx, Frida, IDA or BurpSuite? Every task class plays by different rules, the tools live on different machines, and the same mistakes get made twice because nothing learned carries over. reverse-skill — a cybersecurity skills router pack for coding agents, demoed by WeChat Channels creator 青青-wang in a September 26 video — routes the task to the matching methodology first, checks the local toolchain, and walks a repeatable workflow: an authorization scope gate before any action, an evidence chain, and a report at the end.

Core features

• One source of truth for routing: task routing is driven by a single file, skills/config/routing.json — 44 rules (R0–R44); keyword hits are scored and the top-priority candidate becomes PRIMARY. The 178 cases in skills/tests/routing-benchmark.json run on Windows + Ubuntu CI for every push, and any routing change that breaks an expectation fails the build (counts checked 2026-09-28).
• Authorization before action: the scope gate in RULES.md confirms the authorized scope and network profile before any actual operation; case-init creates a case directory (scope, timeline, workitems), everything is recorded as Evidence → Finding → Path, and the run ends with a report and field journal. The README’s disclaimer is explicit: lawful security research, education, CTF, and systems you own or are authorized to test only.
• Scenarios to toolchains: 20-plus scenario modules cover APK/Android, iOS, binaries (exe/dll/so/elf), .NET, frontend JS with encrypted params, malware and YARA, firmware/IoT, API/GraphQL, supply-chain SBOM, LLM security, OLLVM deobfuscation and more; CTF gets its own orchestrator with 42 sub-skills. The refresh-tool-index script detects what is installed locally and bootstraps what is missing.
• Client-neutral: the routing core, regression suite and case workflow depend on no particular client — Claude Code, Codex, Cursor and OpenCode each load the repo through their own adapter, with an optional Codex plugin that registers no external MCP servers.
• Experience accumulates: field-journal keeps a running log, case-review does read-only evidence review and artifact checks, and the auto-generated skills/INDEX.md keeps navigation current — the same trap does not get stepped on twice.

Typical use cases

• Pentesters: on an authorized engagement, have the agent open a case, pass the scope gate, then route into the pentest-tools or attack-chain modules and produce a handover-ready report.
• CTF players: drop a challenge on the agent and let CTF-Sandbox-Orchestrator spin up the right sandbox and sub-skills for the category.
• Malware analysts: take the APK/binary/YARA routes, with IDA Pro, radare2 or Ghidra chosen by what the local tool index finds.
• Security teams onboarding juniors: case directories plus field journals make every analytical step reviewable.

Quick start

You need Java/JDK (for jadx and apktool), Node 22.12+, Python 3.x, and a code AI client. Clone, then refresh the tool index:

git clone https://github.com/zhaoxuya520/reverse-skill.git
bash skills/scripts/refresh-tool-index.sh   # Windows: skills/scripts/refresh-tool-index.ps1

Check skills/tool-index.md for what was detected, then have your agent read README_AI.md at the repo root and follow it. Kali Linux, Ubuntu/Debian and macOS each have dedicated docs, and skills/scripts/test-routing.ps1 re-runs the regression suite after any routing change.

Summary

For people doing security research under authorization: pentesters, CTF players, malware analysts, and teams that want their methodology captured as routing rules. The README draws the line plainly — unauthorized access, scanning and exploitation are prohibited, so confirm your engagement is in scope before using it. Packing domain methodology into coding agents is becoming a pattern of its own — html-explainer, reviewed here earlier, does the same for explainer videos. MIT-licensed (the CTF-Sandbox-Orchestrator submodule is GPLv3, and Pentest Swarm AI is AGPL-3.0, invoked via CLI/MCP only), built mostly on PowerShell, Python and Markdown contracts, created in May 2026, with 38,399 stars and 5,343 forks as of 2026-09-28 and the latest push on 2026-09-22. Caveats: the routing core ships no tools themselves — commercial software like IDA Pro is on you; PowerShell is the first-class citizen, so Linux/macOS have scripts but the ecosystem leans Windows; and the project’s rapid rise (Trendshift daily badges on the README) means docs and the issue tracker may lag the code.