Raze

The law this skill runs on: the best part is no part; the best process is no process. Occam's razor in code — entities must not be multiplied beyond necessity. Where /simplify tidies what stays and /code-review hunts bugs in what stays, raze attacks the prior question: should this part, this requirement, this step EXIST at all? It runs Musk's Algorithm in strict order — question requirements, DELETE, and only then simplify/accelerate/automate — because optimizing a thing that should not exist is the most expensive mistake there is. Deletion is the judgment ("generation is free, judgment isn't"); raze makes the cut the product.

Three hard invariants:

• Delete before optimize — never the reverse. The steps run in order; you may not simplify/accelerate/automate a part until it has survived the delete pass. Refusing to polish a deletion candidate is itself the core safety rule.
• Propose-on-approval, never auto-delete. Default mode emits a kill-list and writes nothing. Cuts land only in apply, only on individually-approved entries (no bulk "yes to all" — one wrong bulk delete is unrecoverable), one kill per commit, tests green between each. Branch-only, reversible by git.
• The ≥80 genuinely-unused gate. No part is proposed for deletion below 80/100 unused-confidence, and 80 demands real evidence: the import/usage graph traced from REAL entrypoints (never a stale audit), zero live callers, and a trial cut that leaves the build + tests green. Zero test coverage = HIGH-RISK, manual proof required, never auto-trusted.

/raze <target> (default — scan + propose)

Run the Algorithm as ordered evidence passes (fan the reads to scout subagents; only the candidate list returns).

1. Pin + budget. Restate the target as an inventory of PARTS (functions, files, flags, config keys, deps, abstraction layers, doc sections, CI/build steps, pipeline stages, even whole skills). Set the add-back budget (default ~10%). Negative control: if nothing is a deletion candidate, you were handed lean code — stop and redirect to /simplify.
2. Question the requirements (make them less dumb). For each part, trace WHY it exists to a named owner (git blame, the introducing PR/ticket, the constraint it cites). When the reason is NOT self-evident in the diff/commit, this trace runs through /moonwalk <part> (raze's mandatory pre-pass): a part may be proposed for KILL only on a KNOWN-DEAD verdict (reason recovered AND no longer applies); a KNOWN-LIVE or UNKNOWN verdict blocks the cut (KEEP / WATCH — the Chesterton's-fence rule below). Flag every requirement whose only authority is a confident assertion with no live consumer — those are the most dangerous because the least questioned. Compute the idiot index (part cost ÷ irreducible essential cost) where measurable.
3. Delete (the core act). Hardest-to-justify first, decide KILL or KEEP. KILL: no entrypoint reaches it (proven by import graph), its requirement died in step 2, speculative flexibility nobody asked for, redundant layers/flags. Build the KILL-LIST ordered by (idiot index × deletion-safety); each entry: the part · why it dies · the /moonwalk verdict (must be KNOWN-DEAD) · the import-graph/call-site proof it's unused · UNUSED-CONFIDENCE 0-100 (floor 80 to propose) · blast radius (and is it tested?) · the exact cut (diff) · an explicit ADD-BACK BET.
4. Hand off the survivors — do NOT re-roll. Only now is optimization allowed, and it is not raze's lane: route surviving code to /simplify (reuse/altitude/efficiency) and /code-review (correctness on the new edges each deletion created). Token-fat → /burn. Name the handoffs; don't re-implement them.
5. Accelerate, then automate — last, survivors only. For surviving PROCESS targets, recommend cycle-time cuts before automation (automating waste just makes it faster); surface as recommendations, never auto-install.
6. Report + route. Emit the kill-list (proposed ≥80 / watching <80 / refused), the idiot-index leaderboard, the unanswered "why does this exist?" requirements, and the lane handoffs. Honest-empty: if nothing clears the 80 gate, say so and cut nothing — a manufactured deletion is worse than none.

/raze audit

Read-only. Produce the kill-list + idiot-index leaderboard + the requirements ledger (every part whose only authority is an unbacked assertion), and propose/write NOTHING. The "what would I cut, and why" preview — mirrors /hone audit.

/raze apply

Execute ONLY pre-approved kills from a prior scan. One kill = one commit; run the target's pinned check (tests/build/the app per /run) with real output BEFORE the next cut. A red check halts immediately, reverts that single cut, and reports — never proceeds through a broken state. A deletion that breaks the check is ADDED BACK and logged as a restore (paid-for tuition: record WHY, so a future raze does not re-kill it). Record each landed kill + its add-back bet to ~/.claude/raze/cuts.md so the realized add-back rate can be graded later (closes the loop the way /hone closes grade→improvement).

Rules

• Order is law: question → delete → simplify → accelerate → automate. Never optimize a part that has not survived deletion.
• Refuse-list (never cut, regardless of confidence): security/allowlist/auth gates (never trust the model to self-limit), the logging/observability net, the tests themselves (deleting the test that would catch a bad delete is the cardinal sin), secrets handling, anything outward-facing or irreversible outside git. Chesterton's fence: a part whose purpose cannot be traced (a /moonwalk UNKNOWN verdict) is WATCHED, not cut — unknown reason is a reason to keep.
• The add-back rate is the two-sided grade: restores ÷ kills. Well above ~10% = you cut wrong (tighten the gate); near 0% = you were timid (you became /simplify); a small non-zero rate = aggressive enough yet safe. Net parts/lines deleted is the product — a session that nets +lines has not razed.
• Lane discipline: raze deletes; /simplify optimizes the survivors; /code-review checks the new edges; /burn cuts tokens. Hand off, don't duplicate.
• Shared-tree safety: branch-only, explicit git add <path> (never -A), no history rewrite on a shared branch; an adversarial /code-review pass must try to refute each cut before any PR.

Alias: /elonize runs this skill. raze is the razor (delete to the simplest sufficient form — the self-documenting verb); elonize is the broader basket of first-principles/Algorithm moves of which the razor is one. As more of those principles get encoded, they live here.