Snapshot 2026-08-05 14:11:43 UTC · version 1
Navigation across the markdown files of zhaoxuya520/reverse-skill with measured metrics: 402 .md files, 83 SKILL.md, all in git mode 100644. Documentation is split by audience — 46 KB for humans, 157 KB for agents. Covers the routing chain and three license zones (MIT / GPL-3.0 / third-party MIT).
.md map — humans and agents, separatedzhaoxuya520/reverse-skill is a skill router for reverse engineering, authorized penetration testing and CTF. The repository holds 511 files, 402 of them .md. There is no product code: the expertise is packaged as text.
The defining architectural choice: the documentation is split by audience. Some files are written for a human to read; others are read only by the agent. Below is a map of both groups with measured metrics.
| Metric | Value |
|---|---|
| Total files (blobs) | 511 |
Of which .md |
402 (3,948,872 B ≈ 3.77 MiB) |
SKILL.md files |
83 |
| Directories | 253 |
| Git file mode | all 100644 — no executables, no symlinks, no submodules |
Scripts (.sh/.ps1/.py) |
36, none carry the +x bit |
| File | Bytes | Lines | ~Tokens | Purpose |
|---|---|---|---|---|
| README.md | 10,468 | 253 | ~2,577 | Project overview (English) |
| README_zh.md | 10,238 | 250 | ~2,666 | Overview (Chinese, 789 CJK chars) |
| README-kali.md | 1,029 | 33 | ~278 | Kali Linux install |
| CHANGELOG.md | 3,944 | 84 | ~979 | Version history |
| skills/CONTRIBUTING.md | 20,275 | 654 | ~5,583 | How to write and add a new skill |
| File | Bytes | Lines | ~Tokens | Role in the chain |
|---|---|---|---|---|
| README_AI.md | 24,929 | 575 | ~6,160 | Bootstrap. A README not for humans: what the agent does in the first second |
| CLAUDE.md | 1,374 | 43 | ~341 | Claude Code auto-loads it from the root |
| RULES.md | 23,934 | 417 | ~5,978 | Global gates: scope and authorization |
| RULES_zh.md | 33,777 | 596 | ~9,404 | Same gates (Chinese) — largest file in the group |
| skills/SKILL.md | 12,322 | 179 | ~3,415 | Entry point, master controller |
| skills/MASTER-ROUTING.md | 5,480 | 110 | ~1,455 | Fast triage ladder |
| skills/routing.md | 26,298 | 338 | ~6,631 | Full task→skill matrix |
| skills/routing_zh.md | 29,000 | 367 | ~7,859 | Matrix (Chinese) |
The agent-facing documentation is 3.4× larger than the human-facing one — 157 KB vs 46 KB. Humans get a short overview; the machine gets the detailed protocol. That inverts the usual ratio: in most repositories it is the other way around.
Second: loading all 13 navigation files into context costs ~53,300 tokens. So the real work runs on progressive disclosure — the agent reads MASTER-ROUTING.md (~1.5k tokens) and only opens the heavy routing.md when it has to.
Task → RULES.md check scope / authorization
→ MASTER-ROUTING.md fast classification of the task
→ routing.md task → skill matrix
→ tool-index.md what is actually installed (auto-generated, not in git)
→ <skill>/SKILL.md execute by the methodology
→ report Evidence → Finding → Path
The "MIT" badge on GitHub applies to the root only — the detector does not see nested LICENSE files. In practice the repository has three license zones:
| Zone | .md files |
License |
|---|---|---|
| Core and router (root) | 225 | MIT © zhaoxuya520 |
CTF-Sandbox-Orchestrator/ |
91 | GPL-3.0 (copyleft) |
skills/pentest-tools/src-hunter/ |
86 | MIT © MyuriKanao |
All 13 files in the tables above are under the root MIT. The router (SKILL.md, MASTER-ROUTING.md, routing.md, RULES.md, master-route.ps1) is freely reusable. Caution is needed only with the GPL subproject CTF-Sandbox-Orchestrator/ — it must not be pulled into a closed-source product.
The tool is intended for authorized testing. The authorization gate in RULES.md is Markdown, not a technical control. The real risk is not in file permissions (all 100644, nothing executable) but in the instructions: 402 Markdown files tell the agent to run .sh and .ps1. What must be controlled is what you allow the agent to execute.
File sizes and modes come from the Git Trees API (/repos/zhaoxuya520/reverse-skill/git/trees/main?recursive=1, 764 entries, response not truncated — truncated: false). Lines and characters were counted over the raw files from raw.githubusercontent.com. Token estimate is a heuristic: 1 token per CJK character, 1 token per 4 Latin characters. License zones were determined by the actual location of LICENSE files (the nearest enclosing LICENSE governs its subtree).
Data captured 5 August 2026.
concept:forensics-and-reverse-engineeringorg:collider-club Sign in to join the discussion.