MIT Open Source · Python 3.10+ · Zero dependencies

Analyze your Agent Skill,
generate an explorable interactive report

Deterministic scan + evidence-based semantic analysis + strict quality validation. Outputs a 3-panel HTML report with dynamic SVG flowchart, capability catalog, input routing matrix, and live file preview.

ming-skill-analyzer — self-analysis workflow (from real analysis report)
ROUTING PREPARATION ANALYSIS VALIDATION DELIVERY OUTCOME Trigger matched Single target match Out of scope Dir confirmed Zero errors Fix & retry START Discover Skill ACTION Read Request DECISION · 100/100 Route Match ACTION Load SKILL.md ACTION Select Resources GATE Output Dir Gate TOOL Scan + Analyze VALIDATION Strict Validate TERMINAL Render + Serve & Deliver From real analysis report Main path (verified) Boundary / retry 9 nodes · 10 edges routing_score: 100/100
🔁
Self-verified
This page's content comes from the tool analyzing itself
The flowchart, capability descriptions, and findings above come directly from running ming-skill-analyzer on itself and reading the resulting analysis.json. The tool passed its own strict quality validation.
1.0
evidence_coverage
All nodes have source file evidence
0
unsupported_claims
Zero analysis claims without evidence
0.0
inference_ratio
All conclusions trace to source files
From the analysis report's capability catalog

Five pipelines, each backed by evidence

The descriptions below are taken directly from analysis.json, each traced to source file line numbers.

Single-target routing control
Determines whether the request is for understanding, teaching, or visualizing one Skill. Rejects multi-Skill aggregation, execution, modification, and plain code review. Self-analysis is valid: the description explicitly allows "comparisons between that skill's own capabilities or execution branches."
single-target-routing
Deterministic read-only scan
scan_skill.py produces a file inventory, redacted text, and reference signals (with confidence + detection_rule fields; CSS/@media and in-code-block false positives filtered) plus a schema v2 analysis template, without executing any target scripts.
deterministic-scan
Evidence-based semantic analysis
Uses facts.json as the sole factual authority to complete analysis.json: capabilities / input_classes / lanes / nodes / edges / findings. Every conclusion must have a source excerpt; inferences must be explicitly marked.
evidence-analysis
Strict quality validation
validate_analysis.py --strict checks structural integrity, evidence excerpt accuracy, and outputs three quality metrics: evidence_coverage (≥90% required), inference_ratio, and unsupported_claims (must be 0). Only accepts schema_version: 2.
strict-quality-validation
Live report rendering & serving
render_report.py --live generates a 3-panel HTML app (file tree + analysis report + dynamic SVG flowchart). serve_report.py provides a file API on 127.0.0.1, rejecting path traversal, symlinks, and sensitive files.
live-report-delivery
How it works

Three steps to a report, each independently rerunnable

Each step has clear inputs and outputs. Existing analysis runs are never overwritten (enforced in three places: SKILL.md, scan_skill.py, and render_report.py).

1
Deterministic scan
Reads all Skill directory files, redacts secrets, extracts reference signals, and produces facts.json (fact layer) and analysis-template.json (schema v2 template). No target code is executed.
python3 scripts/scan_skill.py ./my-skill --output-dir ./out --prompt "analyze trigger boundaries"
2
Semantic analysis + strict validation
Complete analysis.json (capabilities / input_classes / lanes / nodes / edges) based on facts.json, then run --strict validation to ensure zero unsupported claims and ≥90% evidence coverage.
python3 scripts/validate_analysis.py out/facts.json out/analysis.json --strict
3
Render + live serve
Generate a 3-panel HTML app. The loopback server provides a live file API so the browser reads current disk files on demand. Verify the URL before handing off.
python3 scripts/render_report.py out/facts.json out/analysis.json --output out/report.html --live --lang en python3 scripts/serve_report.py --report out/report.html --skill-root ./my-skill --open
Findings from the self-analysis

The tool produced these four findings about itself

Content taken directly from the findings[] array — not hand-written copy.

strength · high
Never-overwrite policy enforced in three places
"Never overwrite an existing analysis run" is declared in SKILL.md; scan_skill.py exits with code 4 when existing files are detected; render_report.py raises SystemExit when the report path already exists. Three-layer enforcement eliminates any "approved overwrite" ambiguity.
strength · high
Documentation and implementation fully aligned (v2 migration complete)
After three fix iterations: analysis-schema.md and analysis-guidance.md both describe v2 format; validate_analysis.py and render_report.py only accept v2; scan_skill.py constants clearly named (FACTS_SCHEMA_VERSION / ANALYSIS_SCHEMA_VERSION).
teaching-value · info
Self-analysis is a natural consistency test for the Skill
This analysis passed the tool's own --strict validation (evidence_coverage 1.0, unsupported_claims 0), forming a closed loop — the Skill's own quality standard holds for itself. This kind of bootstrapping test is rare in Skill development.
boundary · low
Server process terminates when the Agent session ends (host capability boundary)
SKILL.md says "Do not stop the live server unless the user asks", but serve_report.py is terminated when the Agent session ends in practice. This is a host capability boundary, not a Skill defect. Agents should include the restart command in the handoff.
Install & run

Zero external dependencies — Python stdlib only

Scripts use only Python 3.10+ standard library. No pip install required.

Quick start

# 1. Clone git clone https://github.com/LiMingHuaGit/ming-skill-analyzer # 2. Scan your Skill python3 scripts/scan_skill.py ./my-skill \ --output-dir ./my-analysis \ --prompt "analyze trigger boundaries" # 3. Complete analysis.json, then validate and render python3 scripts/validate_analysis.py my-analysis/facts.json my-analysis/analysis.json --strict python3 scripts/render_report.py my-analysis/facts.json my-analysis/analysis.json \ --output my-analysis/report.html --live --lang en # 4. Open in browser (127.0.0.1 only) python3 scripts/serve_report.py \ --report my-analysis/report.html \ --skill-root ./my-skill --open
Python 3.10+ stdlib macOS · Linux · Windows Modern browser
🔓

MIT license, fully open

Developed by 李明华 (Minghua Li) and open-sourced as part of a personal toolchain. Designed for Agent Skill authors, teams using Claude Code / Codex, and developers who want to understand how Skills work.