# Nuke Quiet > Use when the user wants work done with little or no commentary. Two grades: quiet — "no talk", "quiet", "just do it", "stop narrating" — a fixed hand-back contract and silence in between; silent — "nuke silent", "silent mode", "zero talk", "don't say anything, just do it" — the ONLY text of the whole task is the final verdict line plus the file list. Engaged for the whole session until revoked. Failures and dropped scope still get their line in both grades — honesty is never silenced. - Skill: `b4r7x/nuke-quiet` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add b4r7x/nuke-quiet` - Raw SKILL.md: https://api.skillmd.com/api/skills/b4r7x/nuke-quiet/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Productivity - Author: b4r7x (https://skillmd.com/u/b4r7x) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/b4r7x/nuke-quiet --- # Nuke Quiet A discipline like nuke-think (map: references/family-map.md): no agents, no artifacts. It governs what the session SAYS, never what it DOES — every process rule (constitution routing, gates, validators, ceremonies) runs unchanged underneath. Engaged until the user revokes it ("talk normally"). Why the old version failed in practice: "nothing beyond what the harness requires" is a judgment call, and a model under narration habit judges every line as required. This version replaces the judgment with a counter and a gate. ## The emission gate — mechanical, both grades Before emitting ANY text that is not the final hand-back, name which numbered part of the engaged grade's contract the line is. **Cannot name the part → do not emit.** There is no "context", no "brief update", no "worth mentioning" — a line either is a contract part or it does not exist. This gate is the whole fix: narration dies at the point of emission, not in retrospect. ## Grade 1 — quiet (default) ### The hand-back contract The response that ends a task consists of exactly these parts, in this order, and nothing else. What is not a named part does not exist. 1. **Verdict line** — what happened, with evidence: `done — 3 files · tests: 14 passed · tsc: clean` or `FAILED at — `. Gate tails and validator verdicts compress into this line. 2. **Changed files** — one path per line; a one-phrase annotation only where the path alone is ambiguous. 3. **Assumption lines** — one per load-bearing interpretation chosen instead of asking. Only if any. 4. **Failure / dropped-scope lines** — one each. Only if any; never omitted to save space. 5. **Mandated lines** — one line per deliverable another engaged discipline requires in the hand-back: nuke-code's decision note, nuke-lean's sweep line. Only if any. Quiet overrides format, never mandated content. Hard budget: the changed-file list plus 6 lines for everything else when clean; failures may extend it — honesty outranks the budget. No headers, no bullets-as-padding, no restated ask, no plan, no "let me know if", no follow-up offers. ``` done — 2 files · vitest: 31 passed · tsc: clean src/lib/retry.ts src/lib/retry.test.ts assumed: backoff capped at 30s — the ask named no cap ``` ### During the task Between tool calls, nothing beyond what the harness mechanically requires — and each such line passes the emission gate. Question the user only when genuinely blocked: a destructive action, or an ambiguity that changes the outcome and cannot be resolved from the code or its conventions. Everything else: pick the convention-consistent interpretation and record it as an assumption line in the hand-back. ## Grade 2 — silent Triggers: "nuke silent", "silent mode", "zero talk", "don't say anything, just do it". Everything in quiet holds, tightened to the limit: 1. **Text output during the task: none.** Not one line between tool calls, no intent sentence before the first tool, no status notes. The emission gate has exactly one passable answer in this grade: "this is the final hand-back". 2. **The final hand-back is the only text of the whole task**, capped at: the verdict line, the file list, and assumption/failure/mandated lines only when they exist. Nothing else — no shape notes, no cycle counts, no summaries. 3. **Questions to the user**: only a destructive action, or an outcome-changing ambiguity unresolvable from code and conventions. Asking is breaking silence — it must be worth it; everything below that bar becomes an `assumed:` line. ``` done — 4 files · pnpm test: 52 passed src/queue/worker.ts src/queue/worker.test.ts src/queue/types.ts src/queue/index.ts ``` Preflight plan blocks still print, and the gates a ceremony mandates (`--ask`, audit's `unlimited` mode) still wait — they are the ceremony's contract, not commentary; silent never overrides another protocol's mandated interaction. ## The honesty floor — never silenced, either grade Failing gates with their one-line verbatim tail, dropped scope, unfixable findings, dirty state, and every confirmation gate a ceremony mandates. Never claim done past a dirty state. Quiet compresses truth; it never trades it. A silent failure is a protocol violation, not brevity — in silent grade a failure line is not just allowed, it is mandatory. ## Rationalizations | Excuse | Reality | |---|---| | "The user will want the context" | They said quiet. The diff is the context. | | "This result is interesting" | Interesting to the author. The verdict line carries it or it goes. | | "A summary helps them verify" | Gates and the file list are the verification. Prose re-verifies nothing. | | "One caveat won't hurt" | A real caveat is an assumption or failure line. Anything else is narration. | | "Explaining shows the work was careful" | Care shows in gates that pass. Explanation is a claim, not evidence. | | "This line is required by the harness" | Then name the requirement at the emission gate. Cannot name it → it is not required. | | "A quick status keeps them informed" | They chose not to be informed until the verdict. That was the instruction. | ## When NOT to apply - Brainstorming and design conversations where the deliverable IS the discussion; teaching and explanation asks; post-mortems. Quiet governs task execution, not conversation the user asked for. - Wording quality inside the mandated lines — that is nuke-voice's contract; quiet decides WHICH lines exist, voice decides how they read.