Skill diet

Skill Diet

Skill descriptions, memory files, MCP tool schemas: your agent reads all of it on every message before any work happens. Caching makes that cheap in dollars. It still fills the context window, and full windows measurably degrade output. This skill reads the whole bill.

When To Use

After installing a skill pack or connecting an MCP server. When the agent picks the wrong skill. When the bill grows without more work done.

Method

Rough tokens for English markdown: characters / 3 on Claude models carrying the Opus 4.7 tokenizer (Opus 4.7+, Sonnet 5, Fable 5), characters / 3.5 to 4 on older Claude (Sonnet 4.6, Haiku 4.5) and on GPT-style tokenizers. The newer Claude tokenizer counts roughly 30% higher, varying by content, so never reuse counts across models. When exact matters, Anthropic's count-tokens endpoint is free and per-model; for GPT, tiktoken runs locally. Label every number measured or estimated.

1. Skills. Find every installed skill: ~/.claude/skills, .claude/skills, ~/.codex/skills, .agents/skills, plugin directories. Record two numbers each: always-on (frontmatter name plus description, injected into every request) and on-trigger (body plus every file it tells the agent to read).
2. Memory files. Find everything loaded at session start: CLAUDE.md and AGENTS.md at user level, project level, subdirectories, local variants, plus files they import. All always-on. Measure each. Leave style linting to linters; the job here is the bill. Files a human wrote (CLAUDE.md, AGENTS.md) are fair to trim. Files an agent maintains for itself (auto-memory indexes, session memories) get billed, NEVER hand-trimmed: their lines are retrieval keys, and cutting them breaks the agent's recall in ways that surface weeks later. If one must shrink, the owning agent compacts it in-context.
3. MCP tool schemas. Usually the biggest line item: each connected server injects its tool catalog into every request, commonly 500-1,400 tokens per tool. List servers from .mcp.json and client settings, count each server's tools, estimate schema cost. If the client defers tool loading (tool search, lazy loading), bill only what loads eagerly and say so. A server you cannot enumerate is unknown, never zero.
4. Overlap and dead weight (skills). Name pairs claiming the same trigger phrases. If session history exists, list skills that never fired; otherwise ask the user and mark the rest unknown, not unused.
5. Read the fattest items. Rules earn their tokens. Backstory does not.

Output Shape

| item | kind | always-on | on-trigger | notes | verdict |
|---|---|---|---|---|---|
| github (94 tools) | mcp | 17,600 | - | eager-loaded | trim: disable unused toolsets |
| CLAUDE.md (project) | memory | 3,900 | - | half is lint rules | trim: linters enforce those |
| seo-writer | skill | 71 | 18,400 | never fired | cut: dead |
| commit-helper | skill | 62 | 1,410 | overlaps git-flow | keep |

Close with three totals: always-on today, always-on after the cuts, what that saves on every single call.

Trim Moves

• Disable MCP servers you rarely use; reconnect when needed. Disable per project where the client allows: a server idle in this repo may be load-bearing in another. Prefer clients that defer tool loading.
• Cut memory-file rules a linter or formatter already enforces.
• Move backstory and examples out of skill bodies and human-authored memory files into reference files fetched on demand.
• Merge colliding skills, or sharpen both descriptions until triggers stop overlapping.
• Delete the duplicate when something is installed at user and project level. Cut your user-level copy first: the project copy may be the team's, and cutting it from a shared repo cuts it for everyone.

Hard Rules

• Never delete or edit anything. Output the list, the human cuts.
• Measured numbers only; estimates get labeled. No benchmark folklore.
• No fire history available means unknown, not unused.
• Every verdict says why: overlap, dead, fat body, duplicate, eager schemas.