Claude Code Context Diet

claude-code-context-diet

Every request Claude Code sends carries infrastructure you didn't type: system-tool definitions, the full skill catalog, MCP tool schemas, and feature clusters. It is resident in context on every turn, and you pay for it whether or not you use it. On a heavily-extended setup (many plugins, a big ~/.claude/skills folder, several MCP servers) this can be tens of thousands of tokens before you say a word.

This skill is the trim procedure: measure → rank → cut → re-measure. Every change is a settings.json edit, reversible in one line.

Activate: "my Claude Code is slow / expensive to start," "what's eating my context window," "trim the system prompt," "I have too many skills/tools loaded," "reduce per-request tokens in the CLI," or after installing a batch of plugins/skills.

Not to be confused with

• agent-context-economy — optimizes an agent you build and serve (prompt-cache the tool prefix, compress tool results). That's your product's input economy.
• eco-mode — trims Claude's output verbosity.

This one is about the Claude Code client's own resident payload — the bytes the CLI adds to your requests as a user of it. Different layer, different levers.

Step 1 — Measure first

Run /context. It breaks the current window down by category (system prompt, system tools, MCP tools, memory files, custom agents, and — crucially — skills). Write the numbers down; this is your baseline. Never trim on a hunch — trim on a number.

Want per-tool granularity? Point Claude Code at a logging proxy (set the API base URL to a small local script that logs each request body, then forwards it) and rank the tool/skill definitions by serialized size. The biggest single entries are your first targets.

Step 2 — Read the breakdown honestly

Typical fat, largest-first on an extended setup:

• A skill pack you rarely fire. e.g. a compliance bundle (fedramp/gdpr/hipaa/iso/soc2/pci), a cloud-vendor bundle, a language-specific bundle. Each skill's name + long trigger description sits in context every request. 80 skills × a paragraph each is real weight.
• Bundled skills shipped with the CLI you never invoke.
• Feature clusters — workflow/multi-agent tooling, connectors, artifacts, remote control — each contributes tool definitions.
• MCP servers. Many tools each. (If your client already lazy-loads MCP tool schemas on demand, these are cheap already — confirm before cutting.)

Step 3 — Cut, in order of leverage

All edits go in ~/.claude/settings.json (global) or a project .claude/settings.json. Setting keys evolve between CLI versions — confirm the current names via /config or the Claude Code settings docs before pasting.

1. Disable whole feature clusters you don't use. One flag drops a bundle of tools: bundled-skills, workflows/multi-agent, cloud connectors, artifacts, remote control. Each is a boolean in settings. Biggest bang, least surgery.

2. Deny individual tools by bare name. A permissions.deny entry with a bare tool name removes that tool's definition from the payload — not just its execution. Use for a handful of specific heavy tools you never call.

3. Demote skills to typable-but-not-auto-loaded. A skill override set to user-invocable-only keeps /skill-name working when you want it, but drops the skill from the list Claude auto-loads and reasons over — so its description stops costing you every request. This is the surgical move for a big skill folder: keep them installed, demote the rarely-used majority.

4. Uninstall / move truly-dead skills and plugins. If you'll never type it, remove the plugin or move the skill dir out of ~/.claude/skills. Zero cost, zero ambiguity.

5. Prune MCP servers. Disconnect servers you don't use in this workspace; prefer a client mode that discovers MCP tool schemas on demand over one that resides them all.

Step 4 — Re-measure and keep the receipt

Run /context again. Compare to the Step-1 baseline and record the delta (e.g. "skills 41k → 6k, total start 78k → 34k"). If a number didn't move, the flag was wrong — most keys are version-specific, so verify against your CLI version.

Caveats — don't amputate what you use

• Keep clusters you actually rely on. Workflow/multi-agent tooling and background jobs look bloated but are load-bearing if you run them. Cut by usage, not by size.
• Demote, don't delete, when unsure. user-invocable-only is reversible and keeps the skill a keystroke away — prefer it over uninstalling.
• Watch prompt-cache warmth. Shrinking the stable resident prefix is pure win; churning settings between every session busts the cache. Set it once, leave it.
• Everything here is one line to undo. Flip the flag back, re-run /context, confirm.

The one-paragraph version

/context to get the baseline → find the fattest category (usually an unused skill pack) → disable feature clusters you don't use, deny specific heavy tools by name, and demote the rarely-used skill majority to user-invocable-only (typable, not auto-loaded) → /context again and keep the before/after. Trim by measured usage, never by vibe; every change is one reversible line.