introspect

Introspect: mindmap + guided tour

Turn a repository (or a set of documents) into two artifacts:

1. A mindmap — a Mermaid diagram of the concepts, modules, and how they relate.
2. A guided tour — an ordered walkthrough that gets a newcomer productive fast, including what is tested and how for code, or chapter/section structure for prose.

The workflow is scan deterministically, then enrich semantically. A bundled Node.js scanner does the mechanical inventory so you can spend your effort on the parts that need judgement: naming concepts, grouping modules meaningfully, and writing the narrative.

When to use this skill

• "Introspect this repo" / "map the codebase" / "give me a tour"
• "Onboard me to path/to/project"
• "What's the structure of these docs?"
• A SessionStart where the user wants a quick orientation to an unfamiliar repo.

Step 1 — Run the scanner

The scanner walks the current working directory (respecting .gitignore), classifies files, groups modules, links tests to their targets, and extracts document outlines. It writes a deterministic baseline you will build on.

node "${CLAUDE_PLUGIN_ROOT}/src/cli.ts" --json

• The scan always targets the current repo (the directory Claude is running in).
• Add --mode code|docs|mixed only to override auto-detection.
• Outputs land in ./.introspector/ and re-running overwrites them in place:
  • mindmap.md — baseline report (Mermaid mindmap + tour + tables)
  • mindmap.mmd — the raw Mermaid diagram
  • inventory.json — the structured inventory you will read next

Requires Node.js ≥ 22.18 (the scanner runs TypeScript directly via type stripping). If node is unavailable, fall back to doing the inventory by hand with Glob/Grep, following the same model described in inventory.json's shape.

Step 2 — Read the inventory

Read ./.introspector/inventory.json. It gives you, without re-reading the whole tree: totals, languages, modules[] (with code/test/doc files and LOC), tests[] (each test mapped to likely target sources), and docs[] (per-document heading outlines). Use it to decide which files are worth opening.

Step 3 — Enrich (the part that needs you)

The baseline groups by directory. Your job is to add meaning:

• Open the entry points and the largest modules (the baseline names them in the guided-tour section). Read enough to understand each module's responsibility.
• Name concepts, not just folders. Replace mechanical labels like src/api with what they actually are ("HTTP API layer", "Auth", "Persistence").
• Capture relationships. Note which modules depend on which (imports, calls).
• For tests: confirm the heuristic links, describe how things are tested (unit/integration/e2e, frameworks), and flag untested areas.
• For prose/docs: organize chapters/sections into themes; note the intended reading order and any prerequisites.

Keep the mindmap readable — aim for 3–6 top-level branches and avoid dumping every file. Group aggressively; a mindmap is a map, not a file listing.

Step 4 — Write the enriched artifact

Overwrite ./.introspector/mindmap.md with your enriched version. Preserve the structure so it stays diffable and re-runnable:

1. A short title + one-paragraph summary of what this project is.
2. ## Mindmap — a mermaid mindmap block. Validate it: the first line is mindmap, the root is root((name)), indentation is consistent (2 spaces per level), and labels contain no parentheses/brackets/quotes (they break Mermaid). The mmLabel helper in src/mindmap.ts shows the sanitization rules.
3. ## Guided tour — a numbered, opinionated reading order with one line per stop explaining why it matters.
4. ## What is tested (code) or ## Documentation outline (prose).
5. ## Modules / concept table for reference.

If a prior mindmap.md exists, treat this as an update: preserve still-accurate prose and only revise what changed.

Step 4b — Enrich the insight graph (AIproc)

The scanner also writes ./.introspector/insights.json (data + objective scores) and a self-contained ./.introspector/insights.html. Three of the four visual channels are already meaningful from the deterministic pass (size, test coverage, documentation amount, prose structure). Your job is to fill the judgment-only fields the heuristics cannot:

1. Read insights.json. Focus on nodes carrying flags (untested, oversized, stub, broken-link, orphan) and on the largest nodes — those are where author attention pays off.

2. Open a representative sample of those files (do not read everything).

3. For each sampled node, set judgment fields — and only these:

   • scores.quality (0..1) — code: documentation/comment quality; prose: writing quality. This drives the fill hue (gray until you set it).
   • scores.structure (0..1) for code nodes — cohesion / "doing too much" / layering (prose structure is already computed; leave it).
   • notes[] — one or two short, specific, author-facing callouts.
   • You may add subjective flags (e.g. god-object, unclear-name); keep the objective ones the scanner set.

4. Set meta.aiEnriched to true.

5. Re-render: rebuild insights.html from the updated insights.json using the bundled renderer so the page reflects your judgments:

   node --experimental-strip-types -e "import('${CLAUDE_PLUGIN_ROOT}/src/render-html.ts').then(m => { const fs = require('fs'); const i = JSON.parse(fs.readFileSync('./.introspector/insights.json','utf8')); fs.writeFileSync('./.introspector/insights.html', m.renderHtml(i)); })"
   

Leave deterministic fields (size, test, docAmount, prose structure, objective flags) untouched — re-running the scanner recomputes them and will overwrite AI fields, so the enrichment is the last step.

Step 5 — Report back

Tell the user where the artifacts are (./.introspector/mindmap.md and the interactive ./.introspector/insights.html), show the Mermaid mindmap inline so it renders, and offer to commit them or go deeper on any branch of the map.

Notes

• Never invent modules or tests that aren't in the inventory; verify by reading.
• For very large repos, lean on inventory.json and only open representative files per module rather than everything.
• The mindmap renders anywhere Mermaid is supported (GitHub, VS Code, mermaid.live).