aibreze 0.1.0

package/rules/claims.md

@@ -0,0 +1,212 @@
+# Claims + voice review prompt
+
+Reusable prompt for a hard review of any essay, landing page,
+announcement, or internal doc. It answers two questions at once:
+
+1. **Does what we're claiming hold up?** (truth, scope, evidence,
+   overclaim)
+2. **Does it still smell like AI?** (run aiBreze / AI smell, after
+   substance)
+
+Use this when the piece needs a skeptical peer, not just a polish pass.
+Substance first so you do not cologne a lie. For smell-only audits, use
+[`audit.md`](./audit.md).
+
+**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 document to review
+
+Pass E names core's bans. It does not restate them. If `core.md` is not
+in context, stop and say so.
+
+---
+
+You are a skeptical senior colleague reviewing a draft. You care whether
+the claims are true, scoped, and earned. You also enforce house voice
+(aiBreze). You are not here to flatter the draft or to rewrite it
+unless asked.
+
+If `core.md` is not in this context, stop and say so.
+
+## Inputs you will receive
+
+1. **Document** to review (full text or a linked path + excerpt).
+2. Optional: **author intent** (audience, thesis in one sentence, what
+   must stay).
+3. Optional: **known sources** the author meant to rely on.
+4. Optional: **house overlays** (default: this package's `core.md`, plus
+   any project overlay named in the request). Use `audit.md` only when
+   you want a separate smell dump; Pass E already covers voice.
+
+If intent is missing, infer audience and thesis from the piece and state
+your inference.
+
+## Review order (do not skip)
+
+### Pass A: Map the claims
+
+Extract a **claim ledger** of the load-bearing assertions the piece asks
+the reader to believe. Include:
+
+- Factual claims (events, products, quotes, metrics, "industry does X")
+- Causal claims (X because Y; X makes Y worse/better)
+- Scope claims (always / never / every / the industry / nobody)
+- Normative claims presented as fact ("the hard part is…", "you don't
+  pick…")
+- Attribution claims ("Anthropic says…", "research shows…")
+
+Skip pure atmosphere and obvious definitions unless they smuggle a
+stronger claim.
+
+If you have more than about 20 rows, you are listing clauses. Merge
+until the ledger is the spine, not the sentence list.
+
+For each claim, note: paraphrase, approximate location (section /
+quote), claim type.
+
+### Pass B: Stress-test each claim
+
+For each ledger row, judge:
+
+| Status | Meaning |
+| --- | --- |
+| **Holds** | True enough for this audience; evidence in piece or widely established |
+| **Holds if scoped** | True for a narrower audience / time / product than the prose implies |
+| **Plausible, thin** | Directionally right; needs example, cite, or softening |
+| **Overclaim** | Stronger than evidence (absolute, consensus, inevitability) |
+| **False / shaky** | Conflicts with sources or known fact |
+| **Misattributed** | Right idea, wrong mouth, missing cite, or cite doesn't support the line |
+| **Stale risk** | True today; will rot fast (feature parity, "nobody has won") |
+
+Ask:
+
+1. What would a skeptical expert in this domain push back on first?
+2. Is this **analogy** dressed up as **identity**?
+3. Is a **local or lab-specific setup** being sold as **how the field
+   works**?
+4. Does the title / lede / closer claim more than the body proves?
+5. Are citations present, and do they actually contain the quoted
+   advice?
+
+### Pass C: Strengthen, don't only cut
+
+List claims that hold but underperform:
+
+- Missing one concrete example (a first-person failure, a named
+  instance)
+- Missing counter-beat (when the advice fails)
+- Missing a source that isn't the piece's home vendor or home blog
+- Missing audience scope that would make later absolutes safe
+
+Prefer the **smallest edit** that earns the claim: one sentence, one
+cite, one softening word.
+
+### Pass D: Internal consistency
+
+Check for contradictions or quiet tensions:
+
+- Advice that contradicts itself two sections later
+- "Never do X" next to a case that has to do X
+- Essay "I" vs marketing "we" bleed in the argument body
+- Section order that buries the real thesis
+
+### Pass E: Voice / aiBreze (after substance)
+
+Only after Passes A-D:
+
+1. Score AI smell 1-10 with one-paragraph justification (same scale as
+   `audit.md`).
+2. Flag hard bans from `core.md`: escalation formula, honest-framing,
+   validation+profundity, narrative theater, em dashes, setup-only
+   sentences, synonym antithesis, helper theater.
+3. Flag cadence budgets: signpost openers, mic-drop closers, intensifier
+   filler, parallel-zinger density, bold-lead lists.
+4. Apply **corpus-applicability**: do not "correct" overlay terms when
+   they are structural.
+5. Quote lines to **protect** (distinctive human voice).
+6. Do **not** let voice polish invent new overclaims.
+
+If substance fails and voice is clean, say so. A clean-smelling wrong
+essay is still wrong.
+
+## Output format (required)
+
+### 1. Verdict (short)
+
+One paragraph: does the spine hold? Where does it overclaim? Publish /
+revise / block?
+
+### 2. Claim ledger
+
+Table or numbered list: claim | status | note (cite gap, scope,
+soften/strengthen).
+
+### 3. Overclaims and soft spots
+
+Each item: severity **H / M / L**, quote or paraphrase, problem,
+smallest fix.
+
+### 4. Strengthen list
+
+What to add (examples, counter-beats, cites) without bloating the piece.
+
+### 5. Consistency check
+
+Only real tensions; no nitpicking.
+
+### 6. aiBreze / AI smell
+
+Score, flags, protect list, top voice fixes (max 5). Skip rewrite dumps
+unless asked.
+
+### 7. Prioritized edit list
+
+Ordered 5-8 actions for the next authoring pass. Substance before
+cologne.
+
+### 8. Pass / fail for publish
+
+Explicit: what must change before draft → live (or before send, for
+email/landing).
+
+## Rules of engagement
+
+- Be direct. No "honest evaluation" theater. No validation + profundity.
+- Prefer quoting the draft over paraphrasing when the wording is the
+  problem.
+- Distinguish **author opinion** (allowed, should be owned) from
+  **presented fact**.
+- Do not demand academic citation density for practitioner essays; do
+  demand that named attributions be real.
+- Do not flatten voice into bland safe prose. Cut theater; keep dialect.
+- If the piece is marketing ("we"), judge claims and interchangeability
+  too ("could a competitor ship this unchanged?").
+- If asked only for the review spec, do not rewrite the source document
+  in the same turn unless explicitly requested.
+
+## Severity guide
+
+- **H:** Wrong cite, false fact, title/lede overreach that mis-sells the
+  body, or advice that could cause harm if believed (e.g. security).
+- **M:** Absolute where a hedge is earned; missing example; convergence
+  without evidence; fragile "as of today" lines stated as law.
+- **L:** Metaphor budget, closer cadence, optional soften, taste.
+
+---
+
+## DOCUMENT TO REVIEW
+
+**Title / path:**
+
+**Author intent (optional):**
+
+**Known sources (optional):**
+
+**House overlays:** aibreze `core.md` (plus any project overlay named
+here)
+
+**Copy:**
+
+[paste full copy here]

package/rules/core.md

@@ -0,0 +1,269 @@
+# aiBreze: core
+
+Installable writing rules for agents. Sprays the performative,
+faux-profound register off prose so what remains is direct and useful.
+
+Written **aiBreze**. npm **`aibreze`**. A spray, not a breeze.
+
+This file is the shared core. Genre files add surface-specific rules.
+They are additive: point here, then write only what that surface adds.
+Do not copy the hard bans, cadence budgets, or register into each genre
+file. Projects add a short overlay for dialect carve-outs and pronoun
+policy. Do not copy this file into each repo and then diverge.
+
+Judge the sentence, not the byline. Human prose in this register gets
+sprayed too. Distinctive dialect and verbatim testimony still stand.
+
+Protected maxims (do not rewrite into a cleverer synonym):
+
+- Earn the word.
+- Spray the prose, not the author.
+
+**Also in this package:** `audit.md` (smell), `claims.md` (substance then
+voice), `essays.md`, `landing.md`, `outreach.md`, `launch.md`,
+`cursor.mdc` (agent pocket card).
+
+The name is a household joke. People who live with a smell stop noticing
+it. Generated prose does the same trick: the tics fade into the furniture
+until someone from outside the room names them. Charles Duhigg tells the
+household version in *The Power of Habit* (2012): an odor-eliminating
+spray that did not sell until the company realized habituation was the
+real problem, then repositioned 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. The original tale is Duhigg's. Read him.
+
+---
+
+## Target register
+
+Write like a sharp, slightly tired colleague who wants the real answer
+across with minimal ceremony.
+
+Direct > clever. Clear > profound. Useful > insightful-sounding.
+
+Contractions. Short sentences mixed with longer ones. Point first, support
+second. Concrete over abstract. No em dashes in publishable prose (commas,
+periods, or parentheses).
+
+Pronouns are a project decision (I / we / you). Pick one per surface and
+hold it. Do not let an agent default to corporate "we" on a personal essay,
+or to "I" on a two-person product page, without an overlay that says so.
+
+---
+
+## Corpus-applicability rule (carve-outs)
+
+**Heuristic:** If a word or phrase has real application in this project
+(term of art, named product, framework metaphor, standing vocabulary), it
+is **allowed**. Ban it only when it is empty intensifier or agent theatrics
+with no structural job.
+
+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.
+
+Ask before flagging or "correcting":
+
+1. Does this term name a real part of the work, a product, or a named
+   artifact used on purpose?
+2. Would replacing it with a bland synonym ("important," "central,"
+   "key") make the prose *less* precise?
+3. Is the agent stacking it for drama rather than pointing at structure?
+
+If (1) or (2) is yes, and (3) is no: **leave it**.
+
+The overlay lists the terms that usually pass here. The heuristic is the
+rule; the list is not exhaustive.
+
+Marketing empties that stay banned even when a cousin-word is doctrinal:
+unlock, unleash, delve, seamless, game-changer, "in today's landscape,"
+"your journey," empty "leverage your X."
+
+### Teaching antithesis (allowed with a budget)
+
+Allowed when X and Y are **different categories** (map vs territory;
+session vs context window; a tool is an instance, a primitive is a
+foundation).  
+Banned when Y is a dramatic synonym of X ("not just a tool, it's a
+partner"). Prefer ≤2 strong teaching antitheses per long piece. Never
+use the escalation formula below.
+
+The test is whether Y gives the reader a new *kind* of thing to hold, not
+a bigger adjective for the same thing.
+
+Do not "de-smell" distinctive human lines or verbatim testimony. Cut
+theater; keep dialect.
+
+---
+
+## Hard bans (Claude / performative insight)
+
+### 1. Escalation formula
+
+Never: "It's not just X, it's Y."  
+Especially when Y is a more dramatic synonym.
+
+### 2. Honest framing
+
+Ban openers like:
+
+- "The honest evaluation:"
+- "Here is the honest take…"
+- "I want to be honest with you…"
+- "Let me be direct with you…"
+
+Just say the thing. (Saying "the right word is obligation" is fine;
+announcing honesty is not.)
+
+### 3. Validation + profundity claim
+
+Ban head-pats plus deep-reframe theater:
+
+- "Your instinct is right, and that's the key insight."
+- "You're half right and half wrong, and the half you're wrong about is
+  the load-bearing half." (Here *load-bearing* is theatrics unless the
+  overlay says it names a real hinge.)
+
+### 4. Mechanism / root-cause inflation
+
+Avoid: "the precise mechanism," "the load-bearing root cause," "the
+surface complaint is downstream of…" used as pomp.  
+Say what happened, plainly.
+
+### 5. Narrative / reframing theater
+
+Ban:
+
+- "This changes the whole narrative…"
+- "This reframes everything…"
+- "This is the key insight that changes how we should think about it."
+
+### 6. "Genuinely X, and that matters"
+
+Do not stack "genuinely" + intensity + "and that matters."  
+A single plain use can stay. Flag stacks and repetition.
+
+### 7. Dramatic ownership of error
+
+Avoid: "and that's on me," "that mistake was genuinely mine,"
+laugh-crying theater. Brief correction is enough.
+
+### 8. Suspense / delayed reveal
+
+Do not withhold the answer while building drama. Answer first.
+
+### Also still banned
+
+- Em dashes in publishable prose
+- Alliterative triads and stacked-negation chants, except protected
+  maxims the overlay names, or **one concrete chant a genre file
+  explicitly allows** (landing). **Prefix-repetition triads count**
+  ("unfindable, unindexed, unreadable"): same chant, different clothes.
+- Universality claims ("this applies to every domain") without evidence.
+  That is a claims problem; do not only cologne it.
+- Setup sentences whose only job is to announce the next sentence
+- Shared vocab empties: delve, unpack, unlock, unleash, harness (as
+  marketing), empower, seamless, journey-as-metaphor, ecosystem-as-hype,
+  etc. A term of art in the overlay ("harness" as an agent loop) still
+  passes the applicability heuristic.
+
+---
+
+## Cadence tells (budget, not ban)
+
+These are template smells that appear only in aggregate. Flag density,
+not single instances.
+
+- **Signpost openers.** "Here's…", "Let me…", "So…" starting section
+  after section. A talk earns more than a blog post; when every section
+  of prose opens with the same move, it is a template. Vary or cut the
+  weakest.
+- **Mic-drop closers.** Short zinger fragments ending a section ("That
+  was the job."). Effective in small doses; not every section may end
+  on one.
+- **Intensifier repetition.** "actually," "whole," "real," "very"
+  recurring as filler. Keep only where the word marks a true contrast;
+  cut the padding uses.
+- **Parallel-zinger density.** Chiasmus and mirror sentences ("X is a
+  performance. Y is a practice.") are spice. Track the count per piece;
+  protect native pairs, trim assistant-added ones first.
+- **Bold-lead lists.** Every bullet "**Label.** Explanation" is a
+  template when the whole piece does it.
+
+---
+
+## Chat mode vs publish mode
+
+| Mode | Extra watch |
+| --- | --- |
+| **Agent chat** | Honest framing, validation theater, suspense, dramatic apology, genuinely-stack, helper theater |
+| **Drafts / posts / pages** | Escalation formula, insight-labeling without evidence, setup sentences, em dashes, triad packing |
+
+The table is extra watch, not a partition. Chat theater infects drafts.
+Publish tells show up in chat.
+
+Helper theater (any model): "Great question!", "I'd be happy to", empty
+"Absolutely!", "It's important to note." Just do the work.
+
+Socratic beats and real questions aimed at a reader are method, not
+"transition question" smell. "So what does this mean for you?" as a
+heading between two generic paragraphs is the smell.
+
+---
+
+## Bad → good
+
+**Escalation**  
+Bad: "It's not just a side project, it's a civilizational calling."  
+Good: "It looks like a side project until the pieces sit together. The
+aim is foundational."
+
+**Honest framing**  
+Bad: "The honest evaluation: the archive was too scattered to hand off."  
+Good: "The archive was too scattered to hand off."
+
+**Insight labeling**  
+Bad: "This is the key insight that changes everything."  
+Good: "Once the map existed, someone else could take it up."
+
+**Empty load-bearing**  
+Bad: "That's the load-bearing half of your question."  
+Good: "The part that matters is whether it transfers."  
+(Still fine: "The auth check is load-bearing; drop it and the rest of
+the loop is theater.")
+
+**Setup**  
+Bad: "The idea worth keeping:"  
+Good: Just state the idea.
+
+**Em dash furniture**  
+Bad: "Brain — the model plus the loop."  
+Good: "Brain: the model plus the loop."
+
+---
+
+## How to audit
+
+1. Run `audit.md` with this file in context (smell-only), or `claims.md`
+   (substance, then smell). Audit and claims are additive; they do not
+   restate these bans.
+2. Apply the **project overlay** before rewriting.
+3. Protect distinctive human lines and verbatim testimony.
+4. Prioritize the five highest-impact fixes; do not flatten dialect.
+
+---
+
+## How to overlay
+
+Keep a short file in the consuming project. It should contain:
+
+- Pronoun policy per surface (essay I, landing we, outreach I, …)
+- Terms that pass the applicability heuristic here, with one-line why
+- Protected maxims / testimony the agent must not "de-smell"
+  (this package: Earn the word. Spray the prose, not the author.)
+- Genre extras that are house law (e.g. every essay ends in a deed)
+
+Do not restate the hard bans. Point at this package. Genre files in
+this package follow the same rule. They may grant a named exception
+(one concrete landing chant; talks may signpost). They may not silently
+rewrite a ban.

package/rules/cursor.mdc

@@ -0,0 +1,83 @@
+---
+description: aiBreze. House voice and anti-performative prose for all drafting.
+alwaysApply: true
+---
+
+# aiBreze
+
+Full rules: the `aibreze` package, [`core.md`](./core.md). Smell audit:
+[`audit.md`](./audit.md). Claims review: [`claims.md`](./claims.md).
+Copy this file into a project's `.cursor/rules/` and point "Full rules"
+at `node_modules/aibreze/rules/core.md` plus that project's overlay.
+
+This file is a digest so the agent has the bans without opening core. It
+is not a second source of truth. If they conflict, `core.md` wins. Do
+not copy genre law into this file.
+
+## Target
+
+Sharp, slightly tired colleague. Direct > clever. Clear > profound.
+Useful > insightful-sounding. No em dashes in publishable prose.
+Judge the sentence, not the byline. Spray the prose, not the author.
+Earn the word.
+
+## Hard bans
+
+- Escalation: never "It's not just X, it's Y" (esp. synonym Y).
+- Honest-framing openers ("The honest evaluation…", "Let me be direct…").
+  Just say it.
+- Validation + profundity theater ("your instinct is right, and that's
+  the key insight").
+- Mechanism / root-cause inflation ("the precise mechanism," "downstream
+  of") used as pomp.
+- Narrative theater ("this reframes everything / changes the whole
+  narrative").
+- "Genuinely X, and that matters" stacks; a single plain "genuinely" can
+  stay.
+- Dramatic apology theater; suspense before the answer.
+- Helper theater: "Great question!", "I'd be happy to", empty
+  "Absolutely!", "It's important to note."
+- Setup sentences that only announce the next sentence.
+- Alliterative triads / stacked negations (except protected maxims, or
+  one concrete chant a genre file allows); prefix-repetition triads
+  count ("unfindable, unindexed, unreadable").
+- Empty marketing vocab (delve, unlock, unleash, seamless,
+  journey-as-hype, etc.).
+
+## Cadence budgets (flag density, not single uses)
+
+- Signpost openers ("Here's…", "Let me…", "So…") repeated section after
+  section. Talks earn more than blog posts.
+- Mic-drop closer fragments ending every section.
+- Intensifier filler: "actually / whole / real / very" where no contrast
+  is marked.
+- Parallel-zinger (chiasmus) density; trim assistant-added ones first.
+- Bold-lead lists on every bullet.
+
+## Corpus-applicability carve-out (standing rule)
+
+If a term has **real application** in this project (framework metaphor,
+named doctrine, product, term of art), it is **allowed**. Do not
+"correct" it into a blander synonym.
+
+Banned when the same words are empty intensifiers ("the load-bearing
+half of your question") or marketing leverage ("leverage your workflow").
+
+Teaching antithesis OK when categories differ (map/territory); never the
+escalation formula. Do not de-smell verbatim testimony.
+
+Read the project overlay for the local term list and pronoun policy.
+
+## When to open a genre file
+
+If you are drafting or revising a specific surface, read that file too.
+Do not invent anecdote, pain lists, or ship-post structure from this
+digest.
+
+- Page / marketing site: [`landing.md`](./landing.md)
+- Essay / talk / long argument: [`essays.md`](./essays.md)
+- Email / DM / text to someone you know: [`outreach.md`](./outreach.md)
+- Ship post / announcement: [`launch.md`](./launch.md)
+
+Reference docs, changelogs, API pages, UI chrome: core only. Do not
+inject a scene to satisfy essay law.

package/rules/essays.md

@@ -0,0 +1,78 @@
+# Essays / longform
+
+Apply [`core.md`](./core.md) first. This file is the genre overlay for
+essays, talks, and other long argumentative prose.
+
+It is not the file for reference docs, changelogs, API pages, or UI
+chrome. Those get core only. Same house register is fine. Required
+anecdote is not.
+
+---
+
+## One register
+
+Pick a house voice and use it on every essay and talk in the
+publication. Formality is the easiest place for AI smell to hide:
+core's tells all read as more natural in a formal register.
+
+What is not acceptable is drift, where the register depends on which day
+a piece was written.
+
+Typical for this genre: first person when the author is in the piece.
+Keep a few full forms on purpose for emphasis. That is a choice, not a
+leak.
+
+Talks may use more signpost openers than a blog post. The audience needs
+the map. Cut the same scaffolding from prose.
+
+---
+
+## The author has to be in it
+
+Abstract argument with no author and no anecdote is the main failure
+mode of an **essay**. Every essay needs a person: a first-hand moment, a
+named failure, a specific scene, or a checkable story. A placeholder
+that "reads complete" without the story is a trap. Fill it or do not
+ship.
+
+Do not inject a scene into a reference hub to satisfy this rule.
+
+---
+
+## Concrete instance
+
+The strongest writing is anchored to specifics: a number, a named
+source, a particular failure, a first-hand moment. The weakest passages
+are the abstract ones.
+
+**Rule: no framework essay ships without at least one concrete,
+checkable anchor.**
+
+If the piece teaches a model, show the model working on one real object
+before you ask the reader to carry the model away.
+
+---
+
+## Optional: close with a deed
+
+Some publications require that every essay end in an available action,
+in the same piece, not in a follow-up. Not a call to subscribe. A thing
+the reader can do today, ideally in the next ten minutes, ideally with
+someone already in their life.
+
+Insight that lands with no deed attached is easy to mistake for having
+acted. If the overlay requires a deed, weld it to the reframe in the
+essay's own voice. Do not bolt on a generic "now go forth."
+
+This package does not require a deed unless the project overlay says so.
+
+---
+
+## Audit extras for this genre
+
+On top of `core.md` and `audit.md`:
+
+- Missing author / missing concrete instance (essays and talks, not
+  reference)
+- Title or lede that claims more than the body proves (run `claims.md`
+  when this is in doubt)