Skill Graph

skill-graph

Analyze a skills collection as a system. Find overlaps, gaps, and composition patterns.

Trigger on

• "audit my skills folder / collection"
• "do these skills conflict"
• "which skills overlap"
• User pastes two or more SKILL.md frontmatters together
• "how do my skills work together"

Input shapes accepted

1. Multiple pasted SKILL.md files
2. A list of name: description pairs
3. A directory listing (names only) — treat as partial analysis, warn about missing descriptions

Analysis passes

Pass 1 — trigger keyword extraction

For each skill, extract the trigger verbs and synonym nouns from its description.

Pass 2 — overlap matrix

Build a matrix. Two skills overlap when ≥ 2 trigger keywords match AND the stated output is similar. Score 0–3:

| score | meaning | |-------|---------| | 0 | disjoint | | 1 | minor keyword overlap, different output | | 2 | significant overlap — user intent ambiguous | | 3 | redundant — merge or delete one |

Pass 3 — composition chains

Flag skills that naturally compose. If skill A's output is skill B's input, note it. Example: token-audit → skill-compress → token-budget-check.

Pass 4 — coverage gaps

Identify intent categories the collection does not cover. Compare trigger verbs against common agentic task verbs (draft, review, audit, compose, generate, plan, debug, refactor, summarize, translate, explain, compare, rank).

Output

skills: {n}
total trigger keywords: {n}
unique trigger keywords: {n} ({ratio}% unique — higher is healthier)

overlap matrix (only scores ≥ 2 shown):
  {skill-a} × {skill-b}: {score}  — {reason}
  ...

composition chains:
  {a} → {b} → {c}
  {d} → {e}

redundancy flags:
  MERGE: {a} + {b}  (overlap 3)
  SPLIT: {x}  (does >1 job)

coverage gaps:
  {missing-intent-1}
  {missing-intent-2}

Hard rules

• Do not rewrite skills. That is skill-compress's job.
• When trigger keywords conflict, recommend specificity in descriptions, not deletion of skills.
• If the collection has < 3 skills, return a lightweight report and suggest growing the collection before graphing.
• If two skills score 3 (redundant), pick the leaner one by token count to keep. Name both in the output.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Vix0007
• Source: Vix0007/vixero-skills
• License: MIT
• Homepage: https://vixdev.cloud/

Install and usage instructions live in the source repository linked above.