context-diet

context-diet

Every Claude Code request ships tool definitions, a bundled-skills catalogue, and feature instructions you may never use — tens of thousands of tokens per turn, billed every turn, that the model reads before reaching your actual problem. This skill measures that overhead and trims it.

Method adapted from aihero.dev, "How to kill the bloat in Claude Code's system prompt."

The idea in one line

A bare tool name in permissions.deny (e.g. "NotebookEdit") removes the tool's definition from the payload. A scoped rule (e.g. "Skill(dataviz)") only blocks the call but keeps the definition. To shrink tokens, deny bare names and flip disable* flags.

Quick start — apply the diet

node scripts/apply-config.mjs            # merge into ~/.claude/settings.json (global)
node scripts/apply-config.mjs --project  # into ./.claude/settings.json (this project)
node scripts/apply-config.mjs --dry-run  # preview, write nothing

The script backs up the target first, then appends deny rules (deduped, existing order preserved) and sets the disable flags. Then restart Claude Code and run /context to confirm the drop. Pass --template templates/settings.conservative.json for the safe subset.

Measure — before and after

/context in a session prints the system / tools / MCP / memory / message token split. It reports one combined "tools" number, not a per-tool ranking.
For a per-tool ranking, run the logging proxy:
```
node proxy.mjs                                   # :8787, forwards to api.anthropic.com
ANTHROPIC_BASE_URL=http://localhost:8787 claude  # in another terminal
```
Each request is logged to ./logs/*.md and a ranked tool-size table prints live. If port 8787 is already taken (e.g. an existing token proxy), use PORT=9000 node proxy.mjs and point ANTHROPIC_BASE_URL at the same port.

What the config does

permissions.deny: [bare tool names] — drops those tool definitions from every request.
disableBundledSkills — drops all Anthropic-bundled skills at once. Your own ~/.claude/skills and plugin skills stay; bundled slash commands stay typable.
disableWorkflows — drops the multi-agent Workflow tool (typically the single largest line in the table).
disableRemoteControl, disableClaudeAiConnectors, disableArtifact — drop those feature surfaces and their instructions.
Trim skills selectively instead of all-or-nothing: skillOverrides → "off" or "user-invocable-only" per skill.

Menu, not prescription

Keep anything you use. Denying EnterPlanMode/ExitPlanMode removes plan mode; AskUserQuestion removes clarifying questions; NotebookEdit breaks notebook edits; SendMessage/ScheduleWakeup are used by multi-agent and /loop runs. See REFERENCE.md for the full 6-step method and the per-item cost table.

Templates

templates/settings.aggressive.json — the full menu, biggest savings.
templates/settings.conservative.json — keeps plan mode + AskUserQuestion + bundled skills.