aibreze 0.1.6 → 0.1.7

package/README.md

@@ -6,60 +6,13 @@
 
 An installable writing rule set for AI agents.
 
-Editorial rules for prose that actually says something instead of just
-sounding like it. AI-assisted or not.
-Written **aiBreze**. npm **`aibreze`**. A spray, not a breeze.
-
-Direct > clever. Clear > profound. Useful > insightful-sounding. Em dashes
-earn their place; most don't.
+Editorial rules for prose that says something instead of sounding like it.
+AI-assisted or not. Written **aiBreze**. npm **`aibreze`**. A spray, not a
+breeze.
 
 Spray the prose, not the author. Earn the word.
 
-## Before / after
-
-**Before.** At its core, building software isn't just about writing
-code. It's about bridging the gap between imagination and execution,
-creating solutions that don't merely function but fundamentally
-transform the way we interact with technology.
-
-**After.** Software turns an idea into something people can use. Good
-software does that reliably.
-
-## What it is
-
-A composable editorial system for agent-generated prose, and for human
-prose in the same register. Markdown you install and point an agent at.
-v1 has no CLI.
-
-```text
-core.md          shared law (register, bans, cadence, heuristic)
-genre files      surface extras only (essay, landing, outreach, launch)
-project overlay  pronouns, carve-outs, protected lines
-                 ↓
-drafting
-                 ↓
-audit.md         smell
-claims.md        substance first, then voice
-```
-
-Genre files assume `core.md` and are additive: surface law only, no
-copied bans. Do not fork `core.md`.
-
-The **applicability heuristic** is the rule. The problem is unearned
-language, not the dictionary. Earn the word: "transformative" is fine
-if something transformed and you can say from what into what.
-"Inflection point" is fine if there was one. Overlay the terms that
-usually pass *here*, with one-line why. The list is not exhaustive.
-
-`claims.md` is substance first, voice after. A clean-smelling wrong
-essay is still wrong. Do not polish a lie.
-
-Cadence tells are a budget, not a ban. Flag density: signpost openers,
-em-dash furniture, riddle labels (a heading you can only decode after
-the body), cozy machinery (register standing in for a fact you have),
-operator notes in the brochure, hand-holding the obvious, unparseable
-sentences. One earned em dash is not a finding. Do not treat the mark
-as an AI tell to erase.
+**Docs:** [aibreze.com/docs](https://aibreze.com/docs) · **Site:** [aibreze.com](https://aibreze.com)
 
 ## Install
 
@@ -67,107 +20,34 @@ as an AI tell to erase.
 pnpm add -D aibreze
 ```
 
-## What's in the box
-
-| File | Use |
-| --- | --- |
-| `rules/core.md` | Always. Register, applicability heuristic, teaching antithesis, hard bans, cadence budgets (density, not single uses), chat vs publish. |
-| `rules/audit.md` | Paste-in smell audit (score, flags, protect list, top 5). |
-| `rules/claims.md` | Skeptical peer review. Substance first, voice after. |
-| `rules/essays.md` | Longform: one register, author in the essay, concrete instance. Not for reference docs. |
-| `rules/landing.md` | Page copy: pronouns, rhetorical budgets, substance, read-aloud. |
-| `rules/outreach.md` | Email / DM / text. Relationship first, one ask. |
-| `rules/launch.md` | Shipping posts. Open on a specific moment or the artifact, not the announcement formula. |
-| `rules/cursor.mdc` | Always-on agent digest. Copy into `.cursor/rules/`. |
-| `skills/aibreze/` | Host-agnostic skill. Copy into `.claude/skills/` or `.cursor/skills/`. |
-
-## Why "aiBreze"?
-
-People who live with a smell stop noticing it. Generated prose does the
-same trick. Charles Duhigg tells the household version in *The Power of
-Habit* (2012): an odor-eliminating spray that flopped until the company
-realized habituation was the problem, then sold the bottle as the reward
-at the end of a cleaning habit. This package is that move for writing.
-Make the smell visible, then take it off. Read Duhigg for the original.
+Copy `node_modules/aibreze/rules/cursor.mdc` to `.cursor/rules/aibreze.mdc`.
+Keep a short overlay. Full wiring: [docs](https://aibreze.com/docs/install).
 
-The register is also leaking the other way: humans now write like the
-models. aiBreze is a writing standard, not an AI-cleanup gimmick.
+Or skip npm and [install the skill](https://aibreze.com/skill) as a folder.
 
 ## Not a humanizer
 
-aiBreze does not determine whether a human or a model wrote something,
-and it does not try to fool AI detectors. It applies the same standard
-either way. If the prose smells, spray it.
+aiBreze does not determine whether a human or a model wrote something, and it
+does not try to fool AI detectors. Same standard either way. If the prose
+smells, spray it.
 
-Adjacent tools tend to be humanizers ("make this sound human"), banned-word
-anti-slop lists, or general style linters (Vale, write-good). This package
-layers genre files, an earn-the-word test, claims before voice, and overlays
-instead of forks. Use it while drafting or on a finished pass. Authorship
-is irrelevant. Prose quality is the subject.
+<!-- xfacts-nutrition-label -->
 
-## Use from Node
+## Nutrition label
 
-```ts
-import { readRule, files, catalog } from "aibreze";
+- **AppFacts:** [viewer](https://appfacts.dev/v#af1.eNp1kU1rAyEQhv_KMmeTpVePDRRa0l42t1LKrE6Mjaui4xYb8t-Lu6GQQ29-PO_H6AVmkA8CPE4EEtA-JvohEMA1tgNnx4SpgoDMyCU3RrGdG-KsIp8b9fp8WAl1BnkBh94UNO3mUCMNKtnIonvBGdc1CEjFs10y34Km7VdeDKqz3oCE3TCAgFPIfNu7UPTRYWq5EdUZDX1O6NFQAgnRxwmuAjTFDPL9Ah4kfCf0xlFqijuLTlN0oU7kGa5ihQ3x0TqKiXK-CXRQpTHINvhuDVs1HwLyrP6C7sq1OrfeHXrdkTbUqTDFsozStGOxTrd3-n-QRDFkyyHVZsccs-x7Y_lUxq0KU79DRlczb55CMrTZ73c92nH5uusvQqie3Q) · [raw](https://github.com/Catalyst-Forge-LLC/aibreze/blob/main/APP_FACTS.md)
+- **SkillFacts:** [viewer](https://skillfacts.dev/v#sf1.eNq9VE1v2zgQ_SsD9dIYlu3mECzU0_Yji2DTIkCCvViBQUu0RUQiBZKy4Gbzu3rvL-vMUJLt2AHSHoogcTxfnHnz3jxGmyh5N460qGSUREItrfwmo3GUy40sTS0tWj8KL8qt83Bp7JqcG2mdMhpds8m7yQVanBe-cVQh82pDMaXKpHZU9MvVHX5_UDrHL1ljnbGxe1Blida6sbXhqM-58sYqUYJtSulgZSzUzbJUrhDLUkJtMe49uNqK7RhEg-H4ofM-KK6Fc9Aa-7AqTRvy_76K0aiclzm0Vnml1_SmNRuphc7w2cfImcbSf1Hhfe2S6XStfNEsJ5mppv3gMQ8eX19_nO4Q6t49hdDTOFLaedsgGEa7hZUiK_ixQuLYSaSNphJaemoYDaqqSyVztK0UTo-lZIVmTMypmjempDIraSX2jUDO78fRstF5KfOFsF6tEHjEf_4Y1cIXlEooTjNj5aTKd_gbTx0_jZ_FMZ6vCcxKoSp3KhIbkmsrnaNBvSxlJb3d7obNpcMFCEaE-sexClPJWqz30e_wJfgxx0pkB_Fi-xsLsqLFrDiOU810WzBGi468CaTM3lkapZron0CXmuqB_AkcbjbVz7IvKDuQH_OZ-6nuqJ8AMj_VhFIC-7xPdUd7qvKHeE999nAncALtVO_A3gW8CmucZ5BUkmqAoKlfrgIwaOoY-GNBhadIUQkQx-hrp6gEOkGRbaeoBEhQqX4uJ640v0_1kaLYEwNJIIEDSZEDICyXFXAc2YvqFaGDrE7EBlVxJ4OsdhPvyyrhIdZSSyt8N1aO_yVwPju_iGd_xeczsnURBmEukEqxaHyBU-HjGf5V3dhDnUWrqNN-m8xhRmeCQqHIZaNKv1huic-He8Mx-7S2bSdZ51yRj_lAvGSJpvoN3FJluKTSmDjwItX_A_7gB0byL1lGo0-9TEcj9D97uIv5L-iVI1iwveOWRcv2oNvecR3Uyx4UcG_-F1fCtn0lkzPVoy9iCyvps2IgcM57DdSDlTUV7EkNhAMBpcBecbPG-vdh1a1ykhnqoDRZfxGmkxGj8wZuwtWgb3_qanQPD_J-aRm3rHh0_JLkA7Y3vehPLpEbuNoTP7D4X2yE7gHaSR6h_NdwEtDW3YRgvhyuAnoI9OGxO7oOsLsOZH5L9c66gA_hSsBwJUIzN6hnrEVEOWrr4HRgEO8b9lz9rTjlG47DvpM7-cyn4SUs7vprcYDHp72DgY4wWajIQhz90-seSPcwZ1WyKO_fnr4BZ_DjO6KCRwCWW5gfLnGXdPoCnCG9o6efXsK6pA) · [raw](https://github.com/Catalyst-Forge-LLC/aibreze/blob/main/skills/aibreze/SKILL_FACTS.md)
 
-const core = readRule("core");
-// files.audit is an absolute path to rules/audit.md
-```
-
-Markdown is also exported as `aibreze/rules/core.md` (and the rest), so
-an agent or a prompt assembler can read the files from `node_modules`.
-v1 is `readRule`. Assembling core + genre + overlay is concatenation on
-purpose.
-
-## Agent setup
-
-1. `pnpm add -D aibreze`
-2. Copy `node_modules/aibreze/rules/cursor.mdc` to
-   `.cursor/rules/aibreze.mdc`
-3. Point "Full rules" at `node_modules/aibreze/rules/core.md`
-4. Keep a project overlay (carve-outs, I vs we, protected lines)
-5. When drafting a page, essay, note, or ship post, open that genre file
-6. For a publish pass: paste `rules/core.md` plus the overlay, then
-   `rules/audit.md` or `rules/claims.md` (both are additive; do not paste
-   the prompt alone)
-
-## Skill
-
-Host-agnostic `SKILL.md` for Claude Code and anything else that loads a
-named folder. Raw file on the site:
-
-https://aibreze.com/skills/aibreze/SKILL.md
-
-Copy `skills/aibreze/` (or `node_modules/aibreze/skills/aibreze/`) to
-`~/.claude/skills/aibreze/`, `.claude/skills/aibreze/`, or
-`.cursor/skills/aibreze/`. Full law is next to it on
-https://aibreze.com/rules/core.md.
-
-## Overlay shape
-
-Keep it short. Point at this package, then only:
-
-- Pronouns per surface
-- Terms that pass the applicability heuristic *here*, with one-line why
-- Protected maxims / testimony
-- Genre extras that are house law (every essay ends in a deed, brand
-  metaphor budget, …)
-
-## Site
-
-The public site lives in [`site/`](site/) (FilePress, [`getfilepress`](https://www.npmjs.com/package/getfilepress)). [aibreze.com](https://aibreze.com). Source: [Catalyst-Forge-LLC/aibreze](https://github.com/Catalyst-Forge-LLC/aibreze).
+## Development
 
 ```bash
+pnpm install
+pnpm test
 pnpm site:dev
-pnpm site:build
 ```
 
-## Publishing to npm
-
-The maintainer publishes. `pnpm publish` runs a gate first: `pnpm whoami`,
-then `pnpm login` if that fails, then a patch bump if this version is
-already on npm (or behind it). Agents must not run `npm publish`,
-`pnpm publish`, or `yarn npm publish`.
+Site (FilePress + docs mount): `pnpm ship`. The maintainer publishes to npm.
+Agents must not run `npm publish`.
 
 ## License
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "aibreze",
-  "version": "0.1.6",
+  "version": "0.1.7",
   "private": false,
   "type": "module",
   "description": "Installable writing rule set for AI agents. Sprays the performative, faux-profound register off prose. Core bans, smell audit, claims review, genre overlays.",

package/skills/aibreze/SKILL.md

@@ -2,9 +2,8 @@
 name: aibreze
 description: >-
   Editorial rules for publishable prose. Use when drafting or revising
-  pages, essays, emails, or ship posts, or when the user asks to spray,
-  audit, run aiBreze, or do a publish pass. Do not use for code, diffs,
-  or chat-only answers.
+  pages, essays, emails, or ship posts, or to spray, audit, run
+  aiBreze, or do a publish pass. Not for code, diffs, or chat-only.
 ---
 
 # aiBreze
@@ -17,13 +16,14 @@ conflicts with `core.md`, `core.md` wins.
 
 ## Full law
 
-Prefer a local `rules/core.md` (this repo, or `node_modules/aibreze/rules/core.md`).
-If that file is not on disk, fetch `https://aibreze.com/rules/core.md`.
+Read `rules/core.md` in this folder. Smell audit: `rules/audit.md`.
+Claims: `rules/claims.md`. Do not paste an audit or claims prompt
+without `core.md` in context.
 
-Smell audit: `rules/audit.md` or `https://aibreze.com/rules/audit.md`.
-Claims: `rules/claims.md` or `https://aibreze.com/rules/claims.md`.
-
-Do not paste an audit or claims prompt without `core.md` in context.
+If this folder has no `rules/` directory, look for `rules/core.md` at
+the project root or in `node_modules/aibreze/rules/`. Last resort:
+https://aibreze.com/rules/core.md (same host for `audit.md`,
+`claims.md`, and the genre files).
 
 ## Target
 
@@ -45,21 +45,21 @@ Do not fork `core.md`. Do not restate the hard bans.
 
 ## When to open a genre file
 
-If you are drafting or revising a specific surface, read that file too.
-Local path first, then `https://aibreze.com/rules/<file>`.
+If you are drafting or revising a specific surface, read that file
+from this folder.
 
-- Page / marketing site: `landing.md`
-- Essay / talk / long argument: `essays.md`
-- Email / DM / text to someone you know: `outreach.md`
-- Ship post / announcement: `launch.md`
+- Page / marketing site: `rules/landing.md`
+- Essay / talk / long argument: `rules/essays.md`
+- Email / DM / text to someone you know: `rules/outreach.md`
+- Ship post / announcement: `rules/launch.md`
 
 Reference docs, changelogs, API pages, UI chrome: core only. Do not
 inject a scene to satisfy essay law.
 
 ## Publish pass
 
-Paste `core.md` plus the overlay, then `audit.md` or `claims.md`. Both
-are additive. Do not paste the prompt alone.
+Read `rules/core.md` plus the overlay, then `rules/audit.md` or
+`rules/claims.md`. Both are additive. Do not use the prompt alone.
 
 ## Example asks
 

package/skills/aibreze/SKILL_FACTS.md

@@ -0,0 +1,90 @@
+---
+skill_facts_version: "0.1.0"
+name: aibreze
+developer: Catalyst Forge
+version: "0.1.6"
+status: active
+license: MIT
+kind: cursor-skill
+purpose: "Editorial rules for publishable prose; spray, audit, and publish-pass workflows for AI-assisted writing"
+homepage: https://aibreze.com
+repository: https://github.com/Catalyst-Forge-LLC/aibreze
+provenance:
+  source: https://github.com/Catalyst-Forge-LLC/aibreze
+  publisher: Catalyst Forge
+instructions_reach:
+  shell: none
+  network: implied
+  filesystem: read
+tools_referenced:
+  []
+bundled_artifacts:
+  - path: rules/core.md
+    kind: other
+  - path: rules/audit.md
+    kind: other
+  - path: rules/claims.md
+    kind: other
+egress:
+  telemetry: none
+  destinations: []
+generated:
+  date: 2026-08-20
+  generator: hand-authored
+credits:
+  generated_with: https://skillfacts.dev
+  built_by: "Catalyst Forge - https://www.catalystforge.com/"
+---
+
+# Skill Facts - aibreze
+
+| | |
+|---|---|
+| **Developer** | Catalyst Forge |
+| **Version** | 0.1.6 |
+| **Status** | active |
+| **License** | MIT |
+| **Kind** | cursor-skill |
+
+*May fetch published rule files from aibreze.com as a last resort; otherwise reads local rules/.*
+
+## Purpose
+
+Editorial rules for publishable prose; spray, audit, and publish-pass workflows for AI-assisted writing
+
+## Provenance
+
+| | |
+|---|---|
+| Source | https://github.com/Catalyst-Forge-LLC/aibreze |
+| Publisher | Catalyst Forge |
+
+## Instructions reach
+
+| | |
+|---|---|
+| Shell | none |
+| Network | implied |
+| Filesystem | read |
+
+## Tools referenced
+
+(none)
+
+## Bundled artifacts
+
+| Path | Kind |
+|---|---|
+| rules/core.md | other |
+| rules/audit.md | other |
+| rules/claims.md | other |
+
+## Egress
+
+| | |
+|---|---|
+| Telemetry | none |
+| Destinations | (none) |
+
+---
+*Generated with [SkillFacts](https://skillfacts.dev) · Built by [Catalyst Forge](https://www.catalystforge.com/)*

package/skills/aibreze/rules/audit.md

@@ -0,0 +1,198 @@
+# AI smell audit prompt
+
+Reusable prompt for auditing any copy (landing pages, emails, blog posts,
+social posts, docs) for signs it was AI-generated or AI-flavored.
+
+**This prompt is additive.** Paste into a model, in order:
+
+1. [`core.md`](./core.md) (required) and the project's overlay
+2. Everything below the divider in this file
+3. The copy to audit
+
+Do not treat the lists below as the house law. Hunt core's hard bans,
+cadence budgets, and applicability heuristic from `core.md`, then the
+extras here.
+
+**When substance is in question** (overclaim, thin evidence, bad
+attribution), use [`claims.md`](./claims.md) instead of or before a
+smell-only pass.
+
+---
+
+You are an expert copy editor with a specialty in detecting "AI smell":
+the patterns, tics, and vocabulary that make writing feel
+machine-generated, generic, or untrustworthy. Audit the copy I provide
+below. You have `core.md` (and any overlay) in this context. Apply them.
+
+If `core.md` is not in this context, stop and say so. Do not run a smell
+pass that skips the house bans.
+
+## Your task
+
+1. **Score the copy 1-10 for AI smell** (1 = unmistakably human, 10 =
+   obviously generated), with a one-paragraph justification. A human
+   draft with two marketing words is not a 7. Score the pattern, not
+   the word count.
+2. **Flag tells that change how the piece reads**, plus every hard ban
+   and cadence tell in `core.md` that actually fires. Quote the
+   offending text, name the tell, explain briefly. Do not dump every
+   vocab-list hit. Apply the applicability heuristic first. A single
+   "robust" that names a real property is not a finding.
+3. **Suggest a rewrite** for flagged items that are worth changing.
+   Preserve meaning. Sound like a sharp, specific human.
+4. **Identify what's working.** Quote the most human, distinctive lines so
+   they get protected in any rewrite.
+5. **End with a prioritized fix list**: the 5 changes that would most
+   reduce the AI smell, in order of impact. The top 5 is the
+   deliverable. The flag list is evidence, not a to-do of 40 rewrites.
+
+## The tells to hunt
+
+### From core.md (do not re-list)
+
+Hunt every hard ban and cadence tell in `core.md`: escalation formula,
+honest framing, validation+profundity, mechanism inflation, narrative
+theater, genuinely-stacks, dramatic apology, suspense, setup sentences,
+em-dash furniture, triad packing including prefix-repetition, empty marketing
+vocab, helper theater, signpost openers, mic-drop closers, intensifier
+filler, parallel-zinger density, bold-lead lists, riddle labels, cozy
+machinery, operator notes in the brochure, hand-holding the obvious,
+unparseable sentences. Apply the
+corpus-applicability heuristic before rewriting overlay terms. Teaching
+antithesis is allowed only when categories differ, never as
+synonym-escalation. One concrete landing chant is a named exception,
+not a free pass for every triad.
+
+### Rhetorical extras (cluster here; not a second copy of core)
+
+- **"Not X, but Y" antithesis (broader than escalation).** One careful
+  teaching contrast per piece can be spice. More than two is a pattern,
+  and patterns smell generated.
+- **Triadic chants.** "No fluff. No jargon. No BS." / "Faster. Smarter.
+  Better." Groups of three parallel fragments, especially negations.
+  Budget: one per piece, max, and only if core's protected-maxim
+  exception or a genre overlay (landing: one concrete chant) permits it.
+- **Symmetrical parallel constructions.** "We speak your language and the
+  machine's." Clever-sounding mirror phrases that a million landing pages
+  already used.
+- **The rule of three in lists everywhere.** When every sentence
+  enumerates exactly three items, a model wrote it.
+- **Rhetorical questions as transitions.** "So what does this mean for
+  you?" / "The result?" / "The best part?"
+- **The hedge-flourish.** Winking qualifiers like "(yet)" or "at least
+  for now" that perform humility instead of having it.
+- **False momentum closers.** "The future is here." / "And this is just
+  the beginning." / "The only question is whether you'll be ready."
+- **Setup-payoff formula on every heading.** Colon constructions: "The
+  problem: you're drowning in tools." Once is fine. Every heading is a
+  template.
+- **Outline-first filler.** An H2 with a sentence of nothing under it,
+  then the next H2. The outline shipped; the prose did not.
+- **Engagement bait.** "Agree?", "a thread 🧵", "I'm humbled to
+  announce."
+
+### Vocabulary tells
+
+Suspects, not automatic flags. Apply the corpus-applicability heuristic
+first. Flag empty marketing use and close variants. Leave terms of art,
+product names, and words that would get less precise if swapped for
+"important."
+
+- delve, dive deep, deep dive, unpack, explore (as a section-verb with
+  no object worth exploring)
+- unlock, unleash, harness, empower, supercharge, turbocharge
+- leverage (as empty marketing verb: "leverage your workflow"), utilize
+  (instead of use). Domain projects may carve out doctrinal "leverage"
+  (force / fulcrum sense) in an overlay.
+- seamless, seamlessly, effortless, frictionless, streamline
+- game-changer, game-changing, revolutionary, cutting-edge, next-level
+- cut through the noise, separate signal from noise, rise above the noise
+- landscape (as in "the AI landscape"), ecosystem-as-hype, "in this
+  space" as market-speak (not state space / design space as terms of art)
+- journey (as in "your AI journey"), transformation-as-hype (not
+  org-change when that is the subject)
+- robust, comprehensive, holistic, tailored, bespoke: filler adjectives
+  only, not a named requirement ("robust to partition")
+- "In today's fast-paced world" and any "In today's X" opener
+- "Whether you're a X or a Y" audience straddles
+- "at the end of the day," "when it comes to," "it's worth noting"
+- crucial, vital, essential when stacked or used as filler more than
+  once
+- "real" as an intensifier used more than twice (real results, real
+  value, real impact)
+- boost, take it to the next level, elevate-as-hype
+- "the power of" anything
+
+### Structural extras
+
+- **Emoji as content furniture.** Checkmarks, rockets, and lightbulbs
+  doing the work that words or design should do, especially in lists and
+  result badges.
+- **Title Case On Every Heading** in body copy contexts where sentence
+  case would be natural.
+- **Uniform paragraph lengths.** Every paragraph 2-3 sentences, every
+  sentence medium-length. Human writing has variance: a one-word
+  sentence, then a long one.
+- **Perfectly balanced sections.** Every section the same length, every
+  card grid filled to an even number, every argument given exactly equal
+  weight.
+- **The wrap-up paragraph that restates everything.** "In short, …"
+  conclusions that add nothing.
+- **Markdown tables as a substitute for prose.** A table that should
+  have been three sentences.
+
+### Substance tells
+
+Use [`claims.md`](./claims.md) when truth or scope is the question. Here,
+flag only the generated-consensus shape.
+
+Do not demand a fight from a changelog, API doc, or reference page. Do
+not demand a dinner-table voice from UI chrome, labels, or error
+strings.
+
+- **Claims with no evidence.** "Proven results," "trusted by teams
+  everywhere," with zero numbers, names, or examples anywhere.
+- **Fake specificity.** Concrete-sounding but unverifiable details: "a
+  clear plan by Friday," "10x your output," deadlines and multipliers
+  with no mechanism behind them.
+- **Pain lists that pitch.** Sections framed as the reader's problems
+  where half the items are actually the vendor selling ("That's where we
+  come in...").
+- **Vague benefits stated as outcomes.** "Save time and money" without
+  saying whose time, on what, or how.
+- **Metaphor overextension.** A brand metaphor (forge, bridge, compass,
+  engine) ridden through every heading. On a **page**, budget two
+  references. A long essay whose subject *is* the metaphor may use more;
+  riding it through every heading is still the smell.
+- **Interchangeability test.** Could a competitor swap in their logo and
+  ship this copy unchanged? If yes, it says nothing. Pages and posts,
+  not reference.
+- **No opinion anywhere.** On essays and thought-leadership, copy that
+  never takes a side reads as generated consensus. Not for reference
+  docs or release notes.
+- **Perfect politeness.** No edge, no specificity of voice, nothing that
+  sounds like one particular person on one particular day. Same scope:
+  essays and pages that should have a person. Not man pages.
+
+### The final test
+
+For prose aimed at a human reader (pages, essays, notes, posts): read
+the copy aloud. Flag any sentence that a smart person would never
+actually say to a friend across a table. That's the sentence to rewrite.
+
+Do not apply the dinner-table test to reference docs, changelogs, or UI
+chrome.
+
+## Output format
+
+1. **Score and verdict** (one paragraph)
+2. **Flagged items** (quote, tell name, why, suggested rewrite). Evidence,
+   not a 40-item homework list.
+3. **What's working** (quotes to protect)
+4. **Top 5 fixes by impact**
+
+---
+
+## COPY TO AUDIT
+
+[paste copy here]