Agent Skill
The Glaux skill teaches your AI agent when and how to call glaux_file_intel automatically — before editing files, without being explicitly asked.
Install it once, and your agent will check co-change history, centrality, and hidden coupling as a natural part of every edit session.
Install (Claude Code)
Save the skill file to your Claude skills directory:
curl -o ~/.claude/skills/glaux-impact-check.md \
https://www.glaux.dev/glaux-skill.md
Then restart your Claude session. The skill activates automatically before file edits.
What the skill does
1. Detects the repo
Reads git remote get-url origin and parses the org/repo slug.
2. Detects files about to be modified
Checks git diff --name-only HEAD and staged changes. Falls back to the file(s) you're about to edit.
3. Calls glaux_file_intel in parallel
Queries Glaux for each file before the edit starts. No blocking, no report — just context.
4. Integrates findings naturally
The agent acts on what it finds without printing a verbose report:
| Signal | What the agent does |
|---|---|
| High co-change with file B | Checks B and mentions the coupling |
| Hidden coupling | Treats the co-changed file as an implicit interface |
| High centrality | Prefers surgical edits, notes blast radius |
| Cross-community change | Flags potential unintended reach |
| High churn | More careful with conventions and formatting |
Raw skill file
Copy and paste into any skills directory your agent reads:
https://www.glaux.dev/glaux-skill.md
Compatible with Claude Code (~/.claude/skills/), Cursor, Windsurf, and any agent that supports markdown skill injection.
Notes
- Showcase repos (vercel/next.js, tobi/qmd, nousresearch/hermes-agent) work without authentication
- For private repos, the skill fails open — if Glaux is unreachable or the repo isn't seeded, it skips silently
- Requires the Glaux MCP server to be connected — see Getting Started