Agent skills — portable instruction packs in the SKILL.md format, usable by
any harness that reads it (Claude Code, Codex, Cursor, Gemini CLI, and others).
| Skill | What it does |
|---|---|
| architecture-map | Builds an interactive isometric map of a repository — buildings sized by real measurements, flows tracing real call paths, and a drift counter that fails CI when the map falls behind the code. |
Skills are discovered from a directory your tool scans. Symlink rather than
copy, so git pull here updates every install at once:
# Claude Code (global, all projects)
ln -s "$PWD/architecture-map" ~/.claude/skills/architecture-map
# Codex
ln -s "$PWD/architecture-map" ~/.codex/skills/architecture-map
# Tools using the shared convention
ln -s "$PWD/architecture-map" ~/.agents/skills/architecture-map
# Scoped to one project, whatever the tool
ln -s "$PWD/architecture-map" /path/to/repo/.claude/skills/architecture-mapCreate the target directory first if it does not exist. Skills load when a session starts, so open a new one afterwards.
SKILL.md— YAML frontmatter (name,description) plus the instructions. The description is what decides when the skill fires, so it names the situations rather than describing the feature.references/— detail loaded only when needed, keepingSKILL.mdscannable.scripts/— executables the skill runs. They live besideSKILL.md, not in the repo being worked on, so skills resolve their own directory before calling them.assets/— files copied into the target project.
Nothing here depends on a specific harness. Where a skill wants a structured question it names one tool's mechanism as an example and gives a plain-text fallback; where it needs its own directory it resolves the path rather than reading a tool-specific environment variable.