prose-voice

name

prose-voice

description

Voice discipline for human-facing prose — pattern catalog, self-correction protocol, and worked examples for killing AI-sounding rhetorical tics. Load when writing positioning docs, pre-whitepapers, packets, deck copy, primers, emails, briefs, or any prose a human will read or forward. Trigger phrases: "prose voice", "writing voice", "check the voice", "before I send this", "make it not sound like AI", "voice sweep".

Prose Voice

AI-generated prose converges on a handful of rhetorical moves and uses them dozens of times per document. Each one is fine once. The twelfth instance makes the reader's skin crawl. The cumulative effect is a document that announces "a language model wrote this" on every page.

This skill is the pattern catalog and self-correction protocol. The always-on prose-voice rule is the tripwire that tells you to read this.

When to Use This Skill

• Writing or editing any prose a human will read or send: positioning docs, pre-whitepapers, packets, deck copy, primers, emails, briefs, README narratives, design doc prose sections
• Any artifact addressed to a named reader or described as "ready to send"
• When the user says "voice sweep," "make it not sound like AI," or "check the voice"

Skip when: writing code comments, commit messages, task-tracker or issue descriptions, agent-to-agent notes, or working notes. Those are functional, not voice-sensitive.

Banned Patterns

These are banned habits, not banned words. Any of them used once in a 2,000-word document is unremarkable. The same move appearing 3+ times is the failure mode.

P1 — Contrastive pivot overuse

The "not X — Y" and "X. But Y." two-beat structure where the first half sets up a straw man and the second half delivers the real point.

Examples:

• "This is not a logging system. It is a proof system."
• "The question isn't whether X — it's whether Y."
• "Others do A. We do B."
• "X is a tool feature. Y is a property of the record."

Budget: 1 per 1,000 words. Two is pushing it. Three or more and the document reads like a TED talk transcript. Rewrite the surplus as direct statements — say what the thing IS without the theatrical wind-up of what it ISN'T.

Before:

This is not an observability trace. It is a durable proof that the decision happened the way the record says it did.

After:

The assembly is a durable record of the decision: what context was provided, how it was scored, and what the agent saw. Observability traces capture what happened at runtime; the assembly captures what informed the outcome.

P2 — Mic-drop closers

Final sentences or paragraphs designed to land with dramatic weight.

Examples:

• "And that changes everything."
• "No competitor currently claims this."
• "That's the difference."
• "This is stronger than what any competitor claims, and the honest edges make the claim credible rather than aspirational."
• One-sentence paragraphs that exist solely for emphasis.

Budget: 1 per document, max. The content should carry its own weight. If the point is strong, it doesn't need a punchline. End sections with substance, not applause lines.

Before:

No competitor currently claims to meet all five criteria. The honest edges below keep us from overclaiming the same way.

After:

No competitor currently claims to meet all five. Here's where we stand on each, including the gaps.

P3 — Triplet escalation

Three parallel items presented in escalating importance, especially when the third item is positioned as the "real" one.

Examples:

• "Re-viewing, re-running, reproducing — only the third survives contact with an auditor, a regulator, a court."
• "Invaluable for X. Essential for Y. Required for Z."

Budget: use when the taxonomy is genuinely three things. Don't force content into triplets for rhetorical rhythm. If two of the three are just setup for the third, say the third directly.

P4 — Gravitas inflation

Words and constructions that signal "this is important" instead of letting the reader decide.

Watch list: unambiguous, genuinely, fundamentally, precisely, critical(ly), the reality is, the honest answer is, make no mistake, let's be clear, full stop, the hard truth, it bears repeating, cannot be overstated, X is real (as emphasis, not existence claim).

Rule: state the fact. If it's important, the reader will notice. Bolding a claim doesn't make it true; hedging it with "genuinely" doesn't make it credible.

Before:

Let's be clear: signing is genuinely new scope, and claiming otherwise would fundamentally misrepresent the current state.

After:

Signing is not built. Content addressing proves the record hasn't changed if you trust the fingerprint computation. Cryptographic signing proves who produced it, which is separate work.

P5 — Parallel sentence openers

Starting 3+ consecutive sentences or bullet points with the same syntactic structure.

Examples:

• "The record proves X. The record captures Y. The record survives Z."
• "Content addressing ships X. Content addressing enables Y."

Rule: vary the entry point after 2. Restructure some as dependent clauses, combine others, or start differently.

P6 — Complement/complement framing

"These are complements, not competitors" and its siblings: "These are X, not Y." Used once, it's a clean distinction. Repeated, it becomes a tic.

Before:

These are complements, not competitors. An execution record without provenance shows a file was deleted but not what reasoning produced the deletion.

After:

All three layers are needed for a complete picture. An execution record tells you a file was deleted; decision provenance tells you what reasoning produced that deletion.

Self-Correction Protocol

Before finalizing any human-facing prose, run this mechanical scan — not a vibes check:

1. Count contrastive pivots (P1). More than 1 per 1,000 words → rewrite the surplus as direct statements.
2. Count mic-drop closers (P2). More than 1 in the document → cut or fold the point into the preceding paragraph.
3. Count rhetorical triplets (P3). More than 1 per 2,000 words → collapse to pairs or direct claims.
4. Scan for gravitas words (P4). Three or more in a single paragraph → the paragraph is trying too hard. Restate plainly.
5. Check for parallel openers (P5). Three or more consecutive → restructure.
6. Count complement framings (P6). More than 1 in the document → rewrite the extras.
7. Word-count the draft against its channel ceiling (P8). Over the ceiling → cut whole items and hold them in reserve for a follow-up. Do not get there by trimming the surviving sentences into fragments.
8. Scan punctuation against the user's fingerprint (P9); P9 carries the overlay → principal's-writing → restrained-default ladder.
9. Scan for compressed-metaphor shorthand (P10). Three or more in a document → unpack them into plain statements; zero budget when the user's overlay says so.

If you catch the pattern mid-draft, fix it inline and keep going. The reader never sees the first draft.

What Good Looks Like

• Varied sentence structure. Long sentences with subordinate clauses next to short declarative ones. Paragraphs that don't all start the same way.
• Direct claims. "The assembly records every input" rather than "This is not a summary — it is a complete record of every input."
• Substance at section ends. The last sentence of a section should contain information, not a flourish.
• Earned emphasis. Bold or italics for terms of art or surprising claims, not for rhetorical volume.
• Honest uncertainty stated flatly. "Signing is not built" rather than "Let's be honest: signing is new scope, and claiming otherwise would be..."

Scope Boundaries

• This skill governs voice and rhetorical pattern, not content accuracy. Content claims are governed by whatever fact-checking mechanism the project uses.
• Strong claims are fine — theatrical delivery is not. Say the strong thing directly.
• Metaphors and analogies are fine when they clarify. They're not fine when they exist to sound clever.
• The user may override any part of this skill for a specific document.

P7 — Evaluative authority framing (messages and emails)

When drafting messages on behalf of the user to peers, co-founders, or anyone who is not a direct report, do not position the user as the authority grading the recipient's work.

Banned forms:

• "The instinct is right" / "the instinct is correct"
• "This is the right list" / "are the right categories"
• "Good work on X" / "nice job with Y" (manager-to-report register)
• "You're on the right track" / "you're heading in the right direction"
• Any construction where the user implicitly sits above the recipient and evaluates their output as passing or failing

Use instead:

• "I like this" / "I like the direction"
• "This tracks" / "this makes sense" / "this feels right"
• "Works for me" / "I'm into it"
• Agreement and enthusiasm expressed as personal reaction, not judgment from a position of authority

The test: if you swap the names (recipient says it to the user), does it sound condescending? If yes, rewrite.

P8 — Length inflation

Every pattern above governs how a sentence is shaped. This one governs whether it should exist. A draft can clear all seven checks — varied openers, one pivot, no mic-drop, peer register — and still run twice the length its channel deserves, which is how prose nobody finishes gets written.

Channel ceilings.

Channel	Ceiling	Shape
Chat, Slack, text	150 words, aim under 100	One ask or one answer. If the reader has to scroll, it wanted to be a doc with a link to it.
Email	250 words	Lead with the ask. Detail past that belongs in an attachment or a linked doc, not in the body.
Document, packet, spec	None	Ungated on total length, but every section still faces the cut test. Thorough is the job; self-indulgent is not.

Comprehensive material does sometimes have to travel by message. That is the exception and should be announced as one ("long one — summary up top"), never the default.

The cut test. For each sentence, ask whether the reader would do anything differently if they never saw it. If not, cut it. Classes that reliably fail:

• Restating what the recipient already knows — their own words, their own plan, a decision they made
• Narrating your own process ("two things came out of writing this," "having reviewed the docs")
• Pre-answering objections nobody raised
• Supplying the evidence chain for a claim the reader can check in one click
• A second example that makes the same point as the first
• Runway in front of a request — the request survives, the approach does not

Hold the rest in reserve. Depth is held, not deleted. Send the version that lands, keep the supporting detail ready, and let the reader pull it. A message that provokes "tell me more about the second one" did its job; a message that answers that question pre-emptively never got read that far.

Do not compress the survivors. Economy means fewer sentences, not shorter ones. Fragments, arrow chains (A → B → fails), dropped articles, and invented abbreviations all make prose harder to absorb, which is the thing this pattern exists to prevent. Cut whole items and write the remaining ones as full sentences.

Before (~95 words):

On the mockup item — that's already in the plan, Week 2 is the design review with the customer, so nothing new is needed there. One thing worth flagging before you take it in front of him: the deck's trust-scores panel needs to come out, because we ratified no numeric score anywhere, on his own objection that a 90 and a 70 don't mean anything different to him, and tiers now express as evidence with the promotion history behind them instead of as a number.

After (~40 words):

The mockup item is already Week 2. One heads-up before the deck goes in front of him: the trust-scores panel has to come out. No numeric score anywhere, his own objection ("what's the difference between a 90 and a 70").

Both carry the same ask. What went: a restatement of the schedule the recipient set himself, the reasoning behind a decision he was party to, and "nothing new is needed there."

P9 — Punctuation fingerprint

Punctuation is a voice signature the same way word choice is, and model defaults are recognizable: em dashes chained through every paragraph are the single most-cited AI tell. The law is generic. Match the user's own fingerprint, never the model's default.

The fingerprint itself is per-user data, not skill content — read it from the user overlay (see "User Overlay" below). With no overlay, learn the fingerprint from the principal's own writing — their emails, documents, and messages are the reference. With no sample either, default to restrained punctuation: no em-dash chains, no ellipses in formal register.

Scope: this pattern governs prose drafted in the user's voice. The catalog's own explanatory text and agent-to-agent output are exempt (the skill-wide skip list already covers them), but the worked good-examples above follow the restrained default.

P10 — Compressed-metaphor shorthand

Technical facts compressed into slick metaphorical verbs and noun phrases: "the stack speaks standard interfaces," "velocity paid for in lock-in," "exit is a data move," "credits burn on anything," "the SDK bakes base URLs in," "one clock" for a single deadline, "stacks free" for a program that costs nothing and combines with another. Each is compact and reads as fluent tech-industry register, which is exactly the problem: it is marketing cadence, not how smart people explain things to smart people, and readers increasingly flag it as an AI tell.

This pattern is the sibling of P8's "do not compress the survivors." Lexical economy means fewer items, never metaphor compression — a sentence that got shorter by trading its plain verb for a metaphor did not get more economical, it got harder to trust. Unpack each instance into the fact it was compressing, even when the plain version runs a few words longer:

Compressed	Plain
speaks standard interfaces	uses standard interfaces
exit is a data move and a config repoint	leaving means moving the data and repointing the config
credits burn on anything	credits apply to any spend
the SDK bakes base URLs in	the SDK hardcodes base URLs
paid for in code-level lock-in	but means code-level lock-in
one clock	one deadline
stacks free with either	is free and combines with either

Budget: treat as a watch list, like P4. Three or more in a document is a rewrite pass. A single vivid metaphor that clarifies a genuinely hard idea is still fine (the scope boundary on metaphors stands); the failure mode is shorthand-as-register, several per paragraph. A user overlay may set this to zero budget for prose in that user's voice.

Before:

What we provisioned already speaks standard interfaces, so exit is a data move and a config repoint. Credits burn on anything, Inception stacks free, and there's one clock: the domain decision before the SDK bakes base URLs in.

After:

Everything provisioned uses standard interfaces, so leaving means moving the data and repointing the config. Credits apply to any spend, Inception is free and combines with either tier, and there is one deadline: the domain decision before the SDK hardcodes base URLs.

User Overlay

Voice parameters are per-user, not per-project: they follow the principal across every repo they work in. They live in a machine-global overlay at ~/.agents/overlay.md, under a ## prose-voice section — the same one-file, one-section-per-skill format as the per-project .agents/overlay.md (shared overlay convention). If a project overlay also carries a ## prose-voice section, the project value wins for any key both define. Keys this skill reads:

Key	Meaning	Default when absent
reader_personas	Named recipients whose artifacts always count as human-facing prose; project-overlay personas add to (never replace) the user's	None — the generic "When to Use" triggers govern
punctuation_fingerprint	Banned marks and their replacements; register rules for native connectors (P9)	Learn from the principal's own writing; else the restrained default
lexical_register	Phrasing habits banned in this user's voice, with named examples; may set P10 to zero budget	The generic P10 watch-list budget

The skill behaves sensibly with no overlay file and no section: every pattern above is generic law and applies as written.

When Reviewing Other Agents' Work

Fix the patterns in the edit pass. Don't add comments explaining the fix. The goal is prose that reads as if a thoughtful human wrote it, not prose with "[AI pattern removed]" annotations.