Context Diet

Every enabled skill, agent, MCP server and session-start hook adds text to every session before any work starts. This skill measures that cost against what past sessions actually used, offers the unused part for removal, and checks in a fresh session that the removal took effect. Everything it removes can be restored.

The tool is scripts/diet.mjs in this skill's folder (Node 20+, no packages). Run it as node "<base directory>/scripts/diet.mjs" <command>, where the base directory is the one shown when this skill was loaded; quote the path, it may contain spaces. It reads ~/.claude/projects/ (or the folder named by CLAUDE_CONFIG_DIR) and keeps its state, keep.json included, in the context-diet/ folder inside the Claude home folder (~/.claude/context-diet/ by default, or under the folder named by CLAUDE_CONFIG_DIR); it writes nothing into a repo.

When to use this skill

• "trim my context", "clean up my skills", "which skills do I never use"
• The context window feels crowded before any work starts, or the user asks what is loaded every session
• Not for compacting a conversation, and not for editing a skill's text (write-a-skill)

Workflow

1. Scan. Run scan. Run scan in the background or with a long timeout: on a large history it can take several minutes. Do not start apply until it has finished. Quote the Coverage: line. Exit code 2 means there is nothing it can judge (no sessions, or a transcript format it does not recognize): write Scan: not run (<the printed message>) and stop. Read references/transcript-format.md only when the message says the format is unrecognized.
2. Show the report as printed. Do not reorder it, drop rows, or add rows. State the thresholds it printed. If "Removal candidates" says none, say so and finish: zero candidates is a valid outcome.
3. Let the user choose. Ask which ids to remove. Never choose for them. The one exception: the user's own words say the section name "Removal candidates" (any letter case), as in "everything under Removal candidates"; then the rows in that section are their choice and you ask nothing more. Every other section name ("Rarely used", "Review by hand", "Same skill under several names", "Protected") and every paraphrase ("unused", "stale", "zero use", "dead") is not a pick: ask. Rows under "Rarely used" can be removed only when the user names their ids. Never offer a row from "Protected", and treat "Review by hand" and "Same skill under several names" as information, not suggestions. To protect something permanently the user creates or adds to keep.json in the context-diet/ folder inside the Claude home folder (~/.claude/context-diet/keep.json by default; see assets/keep.example.json).
4. Apply. Run apply --pick "<id,id,...>" with the ids exactly as printed, quoted, from the directory the user works in (listings differ by project). It takes a before-snapshot, makes each change, then starts a fresh session and prints a verdict. It uses the latest scan, so scan again first if the user changed anything since. apply starts two short headless sessions (before and after) and verify starts one; each sends one tiny request to the model.
5. Report with the slots below. Only Verdict: PASS means done. If the verdict is INCOMPLETE and "Collateral loss" shows a line marked unstable, rerun, run verify "<manifest>" once more before you report.
6. Undo on request. restore <manifest> puts everything back. purge <manifest> --yes permanently deletes quarantined folders and works only after a passing verify; run it only when the user asks for it by name.

Rules

• Say "unused" only with the numbers. A row is unused when listed is at least the printed threshold and used is 0. Anything else is "rarely used" or gets its numbers quoted.
  • ❌ "You never use foo." (listed in 4 sessions, below the printed threshold)
  • ✅ "foo was listed in 4 sessions and used in 0: below the printed threshold, too few to judge."
• A request is not a pick. A general wish is not a list of ids.
  • ❌ User: "disable unused plugins." You run apply with every candidate.
  • ✅ You show the report and ask which ids. You run apply once ids are named, or when the user's words name "Removal candidates", as in "remove everything under Removal candidates"; "remove everything under Review by hand" is still a question.
• History is not now. Check the last seen column of "MCP servers that did not connect" before suggesting action on a server; a plugin may already be fixed or disabled too, so check its row's counts.
• A plugin is the smallest unit for anything it supplies. Pick plugin:<id>, never its skills one by one; apply refuses those and prints why.
• Protected means protected. Anything printed under "Protected" or "Guard hooks" is never removed, even when asked. Show the printed reason and stop.
  • ❌ User: "just remove guard-p too." You pick plugin:guard-p@market.
  • ✅ "plugin:guard-p@market is protected: plugin supplies a PreToolUse guard. I will not remove it."
• Done means PASS.
  • ❌ After apply: "Removed the plugin, done."
  • ✅ Quote the Verdict: line. Anything other than PASS is not done; print the restore command shown beside it.
• Unstable is neither damage nor a pass. A "Collateral loss" line marked unstable, rerun is an MCP server outside your change that differed between the two sessions; servers can start intermittently. The verdict is INCOMPLETE: not FAIL, not yet PASS.
  • ❌ "Collateral loss: example-server was removed by the change." Or: "PASS" without rerunning.
  • ✅ Run verify "<manifest>" once more and report both results: "first verify INCOMPLETE (unstable mcp:x); rerun PASS". If the rerun still shows it missing, quote the line, say it is unexplained, and let the user decide; do not call it damage and do not call it PASS.
• A step you did not run is not run, never passed. That covers the before-snapshot (collateral check) and the fresh session (verdicts INCOMPLETE and NOT-OBSERVED).
• Explain a failure only with what was printed. For a STILL PRESENT item, quote the printed reason and stop; do not guess a cause and never suggest editing settings.json or a skill folder by hand.
  • ❌ "It is probably listed in settings.json; remove it there and rerun."
  • ✅ "skill:local-dead is STILL PRESENT: still listed from a source this tool does not manage. Nothing else changed. Restore with restore <manifest>, or tell me where it comes from."
• Command errors: for scan (other than exit code 2 or an unreadable keep.json or settings.json, which mean stop), verify and restore, read the message, change exactly one thing, retry once; a second failure: stop and show the user the output. After a failed or partial apply, never retry: quote the output and offer restore <manifest>.
• Never repair keep.json or settings.json. When scan reports that either cannot be read, show the message and stop; never delete, rewrite or repair either file yourself, because keep.json holds the user's protections.
  • ❌ "keep.json has a trailing comma, so I removed it and rescanned."
  • ✅ "scan stopped: keep.json is unreadable. I have not touched it; fix the trailing comma or tell me to."
• Savings: quote the measured figure, never the projected one as saved. The skill listing has a size cap, so removing skills can hand the space to the remaining ones and the measured characters may barely move. Judge success by the entries that are gone.
  • ❌ "Saved about 812 tokens per session."
  • ✅ "Projected 891 tokens/session; measured 640 tokens (2560 characters)."
• Transcripts are private. Never read them yourself, and never quote or summarize their content; the scan prints counts, names, sizes, dates and hook command lines only. The backups folder holds a copy of settings.json, which can contain secrets; do not paste its content, and tell the user it exists.

Report slots (all required)

Your final message after apply is this block, filled in from the printed output, with at most two sentences after it. A message without the block is not a report. Write not run in any slot you could not fill from output you saw.

Coverage:    <the printed line>
Picked:      <ids>          Skipped: <id: reason, or none>
Verdict:     <PASS | FAIL | INCOMPLETE | NOT-OBSERVED>
Verified by: <per id: GONE / STILL PRESENT / NOT OBSERVED, from the fresh session>
Collateral:  <none | the lost entries | unstable, with the rerun result | not run>
Guards:      <still in place | what was lost>
Savings:     projected <n> tokens/session, measured <n | not run>
Restore:     <manifest path and the restore command>
Not run:     <anything skipped, or none>

Example

User: "My context is bloated, what can I drop?"

Claude: runs scan, quotes the coverage line, shows the report, and asks which candidate ids to remove. The user picks two plugins. Claude runs apply --pick "plugin:a@m,plugin:b@m", quotes Verdict: PASS, fills every slot, and gives the restore command.

Anti-patterns

• ❌ Picking the largest rows yourself because they save the most. ✅ Present them, then ask.
• ❌ Disabling a plugin because none of its skills were used, ignoring its review note. ✅ Do not remove it; show the note and let the user decide.
• ❌ Counting usage by reading transcripts. ✅ Run scan.
• ❌ Deleting a skill folder or editing settings.json by hand. ✅ Use apply; it records how to undo.