token — a Claude Code skill that stops re-reading your whole project

token gives Claude Code a code map of your project and a rule to ask the map first, so it stops grepping and reading file after file every time you ask "where does X happen?".

The map is built locally by graphify from the AST (tree-sitter, 30 languages). No LLM calls, no API key, nothing leaves your machine. Building a map takes seconds.

/token

map: 63 nodes · 176 edges · 8 communities · graph.json 48 KB
tokens: GRAPH_REPORT.md ≈ 651 · raw source ≈ 21551 · query output typically 0.5–4k
freshness: up to date
CLAUDE.md: token rule written → ./CLAUDE.md

Why

On a code-navigation question, Claude usually explores: grep, open a file, open three more, repeat. On a mid-sized project that can easily be 100k+ tokens of raw source before it answers. A graph query returns the relevant symbols, files and line numbers in a few hundred to a few thousand tokens, and Claude then opens only those lines.

Measured example (one "how does this flow work" question on a real project): ~169k tokens of exploration → ~2.9k tokens with the map.

Savings depend on the task. They are large for "where / what calls this / how does the flow go" questions and small for a one-line fix in a file you already have open.

What it does

Running /token in a project:

1. Installs graphify if it's missing: uv tool install, then pipx, then a private venv in ~/.local/share/token-skill/venv. No sudo.
2. Builds or updates the map in graphify-out/ (AST only). Files with a code extension but an XML body (Tiled tilesets etc.) are added to .graphifyignore first, because they break parsing.
3. Writes a rule to the project's CLAUDE.md, in a marked, idempotent block: for code questions, run graphify query before the first grep/read, never cat the whole graph.json, and update the map after changing code.
4. Reports real numbers: nodes/edges, the token size of the report vs. the raw source, and whether the map is stale.

Optional --strict mode installs a Claude Code PreToolUse hook that blocks the first raw file read in a session until a graph query has run.

Install

Requirements: macOS, Linux or Windows (Git Bash or WSL), bash, Python 3.10+, internet for the first install.

git clone https://github.com/WebitroHQ/token-skill.git
bash token-skill/token/scripts/install.sh

The install copies the skill to ~/.claude/skills/token/ and sets up graphify. The last line should be ready: type /token in Claude Code. Then restart Claude Code, because skills load at session start.

Check:

~/.local/bin/graphify --version && ls ~/.claude/skills/token/SKILL.md ~/.claude/skills/graphify/SKILL.md

If graphify is not found, add export PATH="$HOME/.local/bin:$PATH" to your shell profile.

Usage

In Claude Code, inside a project:

You say	What happens
/token	install/verify, build or update the map, write the CLAUDE.md rule, report
/token from scratch	full rebuild (--force)
/token strict	also install the blocking hook

After that you just ask code questions as usual; the CLAUDE.md rule makes Claude query the map first.

The script also works on its own:

T=~/.claude/skills/token/scripts/token.sh
bash $T status  [path]            # exists? fresh? token summary   (exit 2 = no map, 4 = stale)
bash $T build   [path] [--force]  # build/update the map
bash $T query   "how is auth checked" [path]
bash $T install-project [path] [--strict]

What it writes

• ~/.claude/skills/token/: this skill
• ~/.claude/skills/graphify/ and a 4-line # graphify block at the end of ~/.claude/CLAUDE.md (added by graphify's own installer)
• ~/.local/bin/graphify (+ graphify-mcp), and ~/.local/share/token-skill/venv if the venv route was used
• Per project: graphify-out/ (commit it or add it to .gitignore), .graphifyignore, and the marked block in CLAUDE.md

Limits

• Code only. Docs, PDFs and images need an LLM to map; this skill doesn't run one.
• Call edges depend on the language. Tested with graphify 0.9.87: Python and JS/TS get full call edges. Kotlin gets calls inside the same class but misses calls into other objects (e.g. queue.advance()). Dart (Flutter) gets definitions and imports only, no call edges at all, so "what happens when X runs" questions don't benefit there. The map still tells Claude which file a class or function lives in.
• Some files yield no symbols (plain PHP scripts, config files like vite.config.ts). graphify warns about them, and they still need a raw read.
• The map goes stale when code changes. token.sh query checks this before every query and updates the map first if any code file is newer (seconds on small projects). Tested: rename a function and move its caller to a new file without rebuilding. Before the fix, graphify query answered confidently with the old name and old file:line. With the check, the answer comes back with the new name and correct locations. Calling graphify query directly skips this check.
• Windows needs Git Bash or WSL; native PowerShell isn't supported.

Uninstall

rm -rf ~/.claude/skills/token ~/.local/share/token-skill
uv tool uninstall graphifyy   # or: pipx uninstall graphifyy

Then remove the # >>> token-skill block from any project CLAUDE.md and delete graphify-out/.

Credits

The map itself is graphify (Apache-2.0). This repo is a thin, deterministic wrapper around it for Claude Code, made by Webitro.

Made by Webitro

Webitro is a custom software studio. We build AI agents, chatbots, automation and e-commerce systems, and we design custom skills and command sets for teams using Claude Code. If you want a skill tailored to your own workflow, get in touch.

License

MIT