harness-diet — put your always-loaded context back under budget

A harness gains weight automatically (every incident note, every boss directive, every fix lands as another paragraph in an always-loaded rule) but only loses weight when someone runs a diet. Field measurement on a mature harness: growth-to-reduction commit ratio of ~60:1, natural growth of ~70KB/month, and adherence degrading as the pile grows. Anthropic's guidance is blunt: "target under 200 lines per CLAUDE.md file. Longer files consume more context and reduce adherence", and rules without paths frontmatter load "with the same priority as CLAUDE.md" — every session, every token (memory docs).

Result you should expect: one field run took a harness from 146KB always-loaded (143% of budget) to 100.6KB (98%) with zero governance loss — every killswitch, decision table, trigger keyword, and canonical command verified present after the diet.

Budgets (defaults — override via flags)

Metric	Budget	Why
CLAUDE.md	< 200 lines	Anthropic official guidance
Per always-loaded rule file	8,192 B	Keeps any single rule scannable; forces narrative out
Total always-loaded rules	102,400 B	Attention-budget ceiling; beyond this, adherence drops

Bundled script paths: commands below use $HOME/.claude/skills/harness-diet/… (the install.sh layout). For marketplace (/plugin install) installs, resolve under ${CLAUDE_PLUGIN_ROOT}/skills/harness-diet/… instead. The scripts are stdlib-only Python; if unavailable, perform the step manually from the described contract.

Phase 0 — Measure

DIET="$HOME/.claude/skills/harness-diet"   # or "${CLAUDE_PLUGIN_ROOT}/skills/harness-diet"
python3 "$DIET/scripts/harness_diet_audit.py"            # human summary
python3 "$DIET/scripts/harness_diet_audit.py" --json     # machine output
python3 "$DIET/scripts/harness_diet_audit.py" --strict   # exit 2 if over budget (CI/ratchet)

Auto-discovers ./.claude/rules, ~/.claude/rules, and CLAUDE.md files at project/user level. A rule is always-loaded iff its YAML frontmatter has no paths: key.

Phase 1 — Classify every always-loaded rule

Walk the list from largest to smallest and pick one bucket per file:

1. Keep always-loaded — cross-cutting behavior needed in every session (tone, security boundaries, routing indexes). Diet the body (Phase 2) but keep the load class.
2. Path-scope — the rule only matters when specific files are touched. Add paths: globs. ⚠ Path scoping fires on file access, never on spoken keywords. A rule triggered by what the user says cannot be path-scoped. ⚠ No literal brackets in globs ([locale] parses as a character class and silently never matches).
3. Convert to skill / hook-injected — task-specific procedure? Official guidance: "use skills instead". Best case: the rule's real trigger is already a deterministic hook — have the hook inject "Read first" into context and drop the always-load.
4. Migrate narrative to reference — the default for oversized keepers (Phase 2).

Phase 2 — Migrate (block-level, never prose compression)

For each oversized file:

1. Extract a preservation manifest first (before touching anything):

   python3 "$DIET/scripts/preservation_check.py" extract rules/big-rule.md > /tmp/big-rule.manifest.json
   

2. Append the outgoing blocks to references/<name>-ref.md (or your repo's reference dir), verbatim under a dated section header. Append before rewriting the rule body, and commit both together — a slimmed body pushed without its reference content is a content hole for every mirror that pulls in between.
3. Rewrite the rule body: keep the verdict skeleton (what to do / never do), leave a one-line pointer to the reference for the "why" and the war stories.

Must stay in the body (never migrate): CRITICAL/IMPORTANT banners and their verdicts, decision tables, trigger keywords and agent names (they are a routing index, not prose), killswitch env vars, canonical command lines, IDs and threshold values.

Migrate aggressively: incident narratives, measurement history, duplicated explanations of the same point, long example blocks, anything already present in the reference file.

Phase 3 — Verify (two lenses, fresh checker)

1. Loss lens (deterministic):

   python3 "$DIET/scripts/preservation_check.py" verify rules/big-rule.md /tmp/big-rule.manifest.json
   

   Exit 2 = something load-bearing got migrated. Restore it before proceeding. Review the manifest by hand first — prune entries that were intended to move.
2. Efficacy lens: re-run the audit; recompute the total; confirm the savings are real (reference files are lazily Read, not @-imported — check the rule bodies contain no @path imports of the reference, or the "savings" still load at launch).
3. Fresh checker: hand the diff to an independent reviewer that didn't write it (pairs with the review-loop skill). Diets fail quietly: a plausible-looking slim body missing one qualifier is exactly what the maker cannot see.

Phase 4 — Guard (keep it off)

A diet without a guard regrows. Ship-with options in $DIET/hooks/:

• rules-budget-guard.sh — PostToolUse (Edit|Write) advisory: the moment a rule edit pushes a file past budget, the editing session gets a context note telling it to migrate, not append. Non-blocking by design — a blocking pre-commit gate silently stalls auto-commit pipelines.
• Re-run --strict weekly (cron, session-start dispatcher, or CI) and surface the report only when over budget.

Anti-patterns (each one cost a real harness a failed round)

Anti-pattern	Why it fails
Prose compression ("tighten the wording")	Trigger keywords, guru names, stage names are the routing index. A compression pass that saved 12% destroyed 70% of routing signals and was fully reverted. Move blocks; don't rewrite sentences.
Dedup toward a private file	If rules are distributed to a team but ~/.claude/CLAUDE.md is one person's, "delete from rules, it's already in CLAUDE.md" deletes it from every teammate's context. Dedup toward the shared surface.
Slim body committed, reference append not	Mirrors pull a body full of dangling pointers. Commit both in one change.
Trusting the maker's "everything preserved" claim	Extract the manifest before the edit and verify after. Field run: the maker claimed "6 bullets kept verbatim"; the deterministic check found two qualifiers gone.
Treating one diet as the fix	Growth is automatic; reduction is manual. Without the guard, one measured harness regrew from 145KB toward its 220KB peak within weeks.

References

• references/methodology.md — budget rationale, growth mechanics, byte-accounting worksheet, verification lens design, worked field results.