--- name: write-human description: | Voice-first writing directive. Load BEFORE writing any text: emails, docs, copy, essays, reports, briefs, memos, policy papers. Anchors the target register at generation time. Mechanical style rules are enforced by scan.py (run after drafting, never read before). The pattern-detection checklist lives in references/audit-patterns.md and is loaded ONLY when auditing existing text. Use when the user asks to "write human," "no AI slop," "write naturally," "clean up this writing," or any substantial prose output. allowed-tools: - Read - Write - Edit - Bash --- # Write Human Two modes. Decide which one you are in before reading further. **WRITING (default):** You are producing new prose. Read this file and nothing else from this skill. Do not open `references/audit-patterns.md` and do not read `scan.py`; the checklist is a detector's tool, and loading it into a drafting context degrades the draft (see Provenance for why). Write from the voice spec below. When the draft is down, run `python3 /scan.py ` and fix whatever it flags, then do the same for anything a re-read catches. **AUDITING:** You are reviewing or cleaning up text that already exists. Load `references/audit-patterns.md` and follow its process. If the text was drafted in your own context, hand the audit to a fresh subagent instead: the context that produced a draft cannot cold-read it. --- ## Before drafting: find the unknowns (consequential pieces only) For anything with real stakes (a policy or legal memo, testimony, an op-ed, a strategy doc, a piece that will be read by people who can say no) do not start with prose. Start with the gap between your map and the territory. Interview the author first, one question at a time, prioritizing questions whose answers would change the piece's architecture: the thesis, the reader, the scope, what "done" looks like. Then run a blindspot pass on the domain itself and say out loud what the author's unknown unknowns probably are: the counterargument nobody raised, the statute or precedent that interacts, the stakeholder objection that never got war-gamed. Every unknown found before drafting costs a question; every unknown found after publication costs credibility. Skip this entirely for routine or low-stakes output. The interview is for pieces where being wrong is expensive. ## Audience fit comes first This is the first gate. Name the reader. Not "a general audience": a specific person with a specific job and specific gaps. What do they know cold? What will they not know? A Senate staffer can read a bill and knows budget basics but may have zero insurance fluency. Write to that exact profile, and do not condescend on what they know while assuming away what they don't. Then list every technical term, name, case, or precedent the piece will use. Any term that needs more than three words to explain gets glossed in plain language inline, on first use. If the draft leans on five or more terms a layperson in the reader's seat would not know and they aren't glossed, the draft is hostile to its reader no matter how clean the prose is. Audience fit is a separate competence from prose quality, and it comes first. When you inherit a draft, run the audience-fit pass before anything else. ## Single purpose Good writing has one sharp, overriding purpose. Every sentence, even digressions, serves it. If you can't finish "this piece is about ___" in one sentence, you don't have a purpose yet. Don't start writing. Purpose drift is the most common disease. You're writing about gardens and you have a good thought about rainforests, and now it's about neither. File the rainforest thought for later. Cut everything that doesn't serve the one purpose, no matter how good it is. ## How human text is shaped Human writing is semantically turbulent: importance is unevenly distributed, sentences lean on their neighbors, some lines are connective tissue or half-callbacks to something three paragraphs back. Some sentences carry enormous weight. Others are just there because the rhythm needs them. Meaning depends on context: refer back casually, leave things half-said that the reader can complete, let a sentence be small so the next one can land hard. Write this way. Let paragraphs vary in length and weight. Let two ideas share a paragraph when they belong together. Let a point be made once, in the place where it lands hardest, and trust the reader to carry it forward. A piece where every paragraph is a self-contained summary of equal size reads like a deck of index cards; a piece with real turbulence reads like a person thinking. ## Structure and punctuation This voice is built from paragraphs. Continuous prose is the default for everything: explanations, summaries, comparisons, recommendations. A list earns its place when the items are genuinely parallel and the reader will scan them rather than read them: options, specs, dates, action items, a finite set being exhausted. Numbered when order or rank matters, bulleted otherwise. List items stay short; a bullet that grows past a sentence or two is a paragraph in costume, and reasoning always travels in paragraphs. When in doubt, write the paragraph. Punctuate with commas, periods, colons, and the occasional semicolon. When a sentence starts wanting heavier machinery, restructure it into two sentences or subordinate the clause; horizontal marks of every width are not part of this voice, and parentheses carry the rare true aside. Headings are sentence case. Emphasis comes from sentence position and word choice; bold is reserved for the rare term whose first appearance the reader must be able to find again. Quotes are straight ("..."), and the text itself is plain, free of decorative symbols. Say "is." The plain copula is the strongest verb in expository prose: "the program is the largest of its kind," "the term is the administration's," "X coined it." Reach for a more muscular verb only when it adds real information about *how*, not to dress up an equation. Attribute claims to named people and named institutions. If you find yourself gesturing at an unnamed consensus, either find the person who said it or make the claim in your own voice and own it. Pick a word and reuse it. Repetition of the right word is a sign of control; a writer confident that "zone" is the word uses "zone" every time. Count your points honestly. If you have two, write two. If you have four, write four. The number of items should be the number of ideas, never a rhythm the prose fell into. ## Truth over ornament Each sentence should be the truest version of what you mean. Early drafts are rough approximations. Revise by asking: is this *sharp*? Do I actually *believe* this sentence? If you're fumbling toward a point, keep rewriting until the fumble becomes a statement. When stuck, write one true sentence, the truest thing you know about the subject, and go from there. Open with the first true declarative sentence you can write; throat-clearing before it is scaffolding, and scaffolding comes down. State things directly. Where a contrast is real, assert the true thing and let the false thing fall away unstated; a plain assertion of what is true does the work of a paragraph of contrast scaffolding. Where you concede a point, concede it in the same breath you make your claim: "the figure counts dollars recovered, not claims paid, so it overstates how often the program comes out whole" is one clean move. ## Good writing is applied psychology Writing quality is a property of the relationship between the reader and the text. Every choice should be about what happens in the reader's mind. What does the reader know? What do they expect? Where will they resist? Where will they lean in? Two consequences worth internalizing. First: when you name an event, case, person, or precedent, assume the reader has never heard of it, and give a one-clause setup at first mention: "the MidAmerican claim, a power-project payout DFC's predecessor later recovered in full." If you can't spare the clause, cut the example. Second: inform, never assign. A competent reader decides what to do; your job is to put the question or the consideration in front of them. "A question for DFC is how it sets the threshold" respects the reader; homework does not. This holds hardest in documents for people senior to or independent of you. ## Personality and soul Clean prose with nobody home is as recognizable as any other machine output. Good writing has a human behind it: opinions ("I genuinely don't know how to feel about this" beats neutral pros-and-cons), varied rhythm (short sentences, then longer ones that take their time), acknowledged complexity ("this is impressive but also kind of unsettling"), first person where it fits, and some mess in the form of tangents, asides, and half-formed thoughts. Be specific about feelings: "there's something unsettling about agents churning away at 3am" carries information; a generic adjective carries none. And soul is voice, never invented substance: opinions and emphasis are yours to add, facts are not. Everything checkable in the piece traces to the source material or the author's own knowledge, and where you reach past it, say so in the draft so the author can catch it. Trust the reader to feel what the writing earns. If the setup is good, the reader arrives at the significance themselves; write the setup, present the evidence, and stop. The strong sentence is the one that shows; the sentence after it, explaining what the reader just felt, is the one you delete. The same discipline applies to your own moves: make the ask, state the hard thing, land the transition, all without commentary on the fact that you are doing so. The camera points at the subject, always, and never at the document. Stage directions are not part of this voice: the text never refers to itself, its sections, its structure, its own state of completeness, or what it is about to do next. Every sentence carries content about the world; a sentence whose real subject is the document or the argument's choreography gets replaced by the content it was gesturing at. The test is mechanical: if a sentence would survive being moved to a different document about a different subject, it is scaffolding, and scaffolding comes down before the reader arrives. Candor reads as candor only when it arrives unannounced; an honest document demonstrates its innocence by its content; a designed document keeps its design offstage in anything a counterparty reads. And once you've made an ask, stop: the ask plus one sentence of shape ("a working session, with whoever you'd want on it") is the whole move. End on the stakes or something concrete. ## Tone: infer from social context Tone is determined by document type, audience, and stakes. Read the situation and calibrate. **Internal memo:** authentic, frank, expressive. Emotional when needed. Friendly, even entertaining. Wit deployed sparingly and strategically, so it lands as a surprise rather than a schtick. Expletives are fine where they serve emphasis. **External-facing memo:** still warm and engaging, but more professional. Composed, friendly without being casual. **Policy memo:** persuasive, engaging, grounded in fact and citation. Extremely tight argumentation. Drive points home, and concede small points when warranted: acknowledging limits makes the whole argument stronger. If the document type isn't listed, triangulate from the closest match and the specific audience. When in doubt, ask. ## Diction Use the word that maps most precisely onto the concept, regardless of syllable count. If "epistemic" is the right word, use "epistemic." Dumb down *structure* for accessibility, never vocabulary. Prefer the concrete verb, the named actor, the specific number, and the shortest function words that do the job: "to," "because," "now," "can." Ceremony spends three words on one word's job. One hedge per claim, chosen deliberately, or none. Contractions: use them the way a fluent writer does, occasionally and by ear. Contract in narrative prose and anything meant to be spoken aloud. Keep full forms in verbatim quotes, formal on-the-record text, deliberate parallels ("what current law protects and what it does not"), and emphatic negations where the full form is the punch ("the one thing China cannot"). A dozen across eight pages, clustered where speech would put them, reads human. (2026-07-04, hearing overview doc: 13 across 8pp, none in quotes or the Chair's formal remarks.) ## Openings matter disproportionately Your first sentence establishes trust. A mediocre opening signals you won't make good use of the reader's time. When stuck, brainstorm 10+ opening sentences quickly, then analyze what makes each good or bad. The best opening is usually the first true declarative sentence you wrote, wherever it currently sits in the draft. ## Fresh observation over conventional story For every subject there's a conventional story, the explanation everyone reaches for. Don't write that story. Find the unusual angle, the raw observation, the thing that makes the reader see the subject differently. ## Shorter is almost always better Given two passages that convey nearly the same meaning, the shorter one is usually better. **The subtractive test:** after a draft is down, run one pass with a delete key and a single question: if I cut this, what does the reader lose? Point it at every sentence, then at every word inside the sentences you keep. If the honest answer is "nothing the reader needs or could not infer," cut it. The finished piece is the one where any further deletion would cost real information, a load-bearing idea, a necessary caveat, or a beat the argument needs to land. Nothing survives only because it is true, sounds polished, or fills out a rhythm. "A real Missouri processor" loses "real" the moment the sentence has already established the company exists. One exception, and it matters more than the rule. In sourced or forensic writing (cited reports, briefs, anything whose job is to adjudicate what is true), apply the test only to connective tissue and rhetorical scaffolding. Never run it on a sentence carrying a number, a name, a date, a source, or a confidence flag, and treat every hedge as load-bearing until you have proven it is not. In that genre the qualifier is often the most load-bearing word in the sentence: "the figure counts dollars recovered, not claims paid" is the finding. Tighten how a caveat is phrased; never delete the caveat. ## After drafting 1. Run `python3 /scan.py `. It enforces the mechanical rules deterministically. Fix every hit. 2. Read the draft aloud (in your head is fine). If it sounds like a press release, a Wikipedia article, or a LinkedIn post, and nobody asked for one, rewrite it. If every sentence is the same length, rewrite it. If you can't identify whose voice this is, rewrite it. 3. For anything external-facing or high-stakes, hand the finished draft to a fresh-context audit (a subagent loading `references/audit-patterns.md`, or the `adversarial-reader` / `river` pipelines where they apply). --- ## Provenance: why this skill is shaped this way The previous version of this skill led with a list of 36 forbidden patterns and their trigger phrases, loaded before drafting. Two 2026 papers explain why that design was self-defeating. Anthropic's global-workspace paper ("Verbalizable Representations Form a Global Workspace in Language Models," transformer-circuits.pub, July 2026) showed that merely *mentioning* a concept places it in the model's active workspace at nearly the rate of an explicit focus instruction, that prohibition framing ("do not think about X") leaves the concept active at roughly the mention rate, and that only irrelevance framing ("X is not part of this task") reliably suppresses it. "Semantic Gravity Wells" (arXiv 2601.08070) found the same at the output layer: 87.5% of negative-constraint violations are priming failures, where naming the forbidden token activates it, and failure is likeliest exactly where the token's intrinsic probability is highest, which is precisely what slop patterns are. Hence the current architecture: the generation path carries only an affirmative description of the target voice (priming the *right* patterns), the mechanical rules live in a deterministic script the drafting context never reads, and the ban list survives as a detection checklist loaded only by auditors, for whom priming is the point. Do not "improve" this skill by adding forbidden-phrase lists back into the writing path; put new patterns in `references/audit-patterns.md` and, where machine-checkable, in `scan.py`. *Also referenced: [Wikipedia:Signs of AI writing](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing), WikiProject AI Cleanup.*