context-budget

Overview

Audit what's eating my Claude Code context — skills, MCP servers, the CLAUDE.md chain, and my growing memory — and produce a prioritized list of what to trim. The point is keeping the whole setup lean, not just the pack. This is skill #11, a conscious break of the original 10-cap (since raised again, 11→12, for exploit-hunt), justified by token economy being a recurring concern.

When to use

• Context feels heavy / output quality slips, or /context shows a big always-on chunk.
• I've added skills, agents, or MCP servers, or MEMORY.md keeps growing.
• Before adding more — to know if there's headroom.

The key model: always-on vs on-demand

• Always loaded (every turn): skill description lines, MCP tool schemas, the CLAUDE.md

chain, and `MEMORY.md` (the memory index). This is the real budget — trim here first.

• On-demand (only when invoked): a skill's SKILL.md body, an agent's prompt. Cheap — a

long body costs nothing until it runs, so don't optimize bodies.

How it works

1. Inventory

• Skills: make details (or claude plugin details <plugin>) for the pack's projected

token cost; note each description length (always-on).

• MCP servers: count servers and tools (~500 tokens of schema per tool, always-on). Flag

servers with many tools, or that just wrap a CLI I already have (gh, git, supabase).

• CLAUDE.md + memory: size of the CLAUDE.md chain and MEMORY.md + memory/*.md. The

memory index is always-on, so its growth is a direct context tax.

• Agents: description length (loads on every spawn) and body size.

2. Classify each component

• Always needed → keep.
• Sometimes needed (domain-specific, rarely triggered) → lazy-load / disable until used.
• Rarely needed / redundant / overlapping → remove.

3. Detect the usual bloat

• Skill/agent description over ~30 words (an always-on tax on every turn).
• MCP over-subscription (tools I never call, or CLI-wrapper servers).
• MEMORY.md entries that are stale, duplicated, or no longer true → prune them.
• CLAUDE.md verbosity that should be a rule or a memory instead.

4. Report

Total always-on overhead, a breakdown by component (skills / MCP / CLAUDE.md / memory / agents), the headroom left, and a prioritized trim list — biggest always-on saving first.

Output

A short report plus the top 3 trims by always-on tokens saved. Act on the always-on items first; on-demand bodies are not the problem.