Prose — plain text that communicates¶

Text is the one artifact every project ships, on surfaces nobody designs: a store listing truncated after three lines, release notes read on a phone, an error string, a report a person reads once. Until now Muriel judged prose only negatively, through the anti-slop list in SKILL.md. The list says what not to write. This channel says how to write the thing.

Part of the muriel skill. The invocation gate still applies: Muriel joins for copy whose structure or register is unsettled, or for an audit of shipped text, not for a one-word label change.

When to use¶

Store listings, product pages and promotional text, where the surface truncates and the reader is deciding whether to install.
Release notes, changelogs read by users, announcements.
UI strings: labels, empty states, errors, permission prompts, onboarding.
Explainers and technical briefs, README prose, documentation pages.
A report, handoff or summary an agent writes for a person.

Not for: scientific manuscripts (read science.md, whose register carve-out wins), the visual composition of a README (readme.md) or an editorial page (web.md), or an affordance problem dressed as a copy problem (see Affordances before instructions in SKILL.md: fix the control, then delete the coaching).

Purpose in the solution¶

Before writing, tell the calling agent:

Purpose: what the reader must know or do after reading, in one sentence, and the one decision the text serves.
Integration: the surface and its limits. Character or byte caps, where it truncates, which lines are visible before "more", whether headers and bullets render, which facts the surrounding product already states.
Proof: the counts, the first-screen test, the Stranger pass and the linter run that will verify it.

The reader contract¶

A person reads plain text to answer three questions, in this order:

What is this?
What do I do with it?
What does it cost or require?

Answer them in the part of the text the surface shows before any effort: the first three lines of a store description, the first sentence of a release note, the subject line, the first screenful of a README. Everything after that is for the reader who has already decided to keep reading, and it is organized for scanning, not for reading in order.

Default workflow¶

Name the reader and the one decision. Write both down. They decide every cut.
List the facts first, each with its source: a verified build, a document, a measurement. A claim without a source does not enter the draft.
Choose the register from the table below. Registers have different rules for tense, caveats and history, and mixing them is the usual way copy goes wrong.
Write the first three lines against the reader contract. Rewrite them until a stranger could answer the three questions from those lines alone.
Structure the body for scanning: one idea per header, one list of parallel items, caveats once.
Cut. Apply the deletion test to every sentence: if removing it loses no information, remove it.
Measure: counts against the surface's limits, the first-screen test, the linter, a Stranger pass.
Hand over the text with its counts and the facts it rests on.

Structure rules¶

Promise, definition, action. Line one names the benefit in under about 50 characters. Line two says what the thing is and where it runs. Line three says what to do first. Social proof, when it exists, is sentence two of the second paragraph. When it does not exist, the slot stays empty.
One idea per header. A header is a benefit phrase of two to five words, set in capitals where the surface has no type hierarchy. Under it, one or two paragraphs of at most about 45 words each.
One list of parallel items. Features, steps, options or files go in a single bulleted list of six to twelve items, each under twelve words, each starting the same way (a noun or a verb, not a mix). Two lists in a row are a sign the structure is wrong.
Price and permissions as lists, not prose. What is free, what the purchase adds, what the subscription is, in bullets. One sentence after the list states the model: one-time or recurring.
Caveats once. Requirements, limitations and privacy statements go in one block near the end, one line each. A caveat repeated in three sections reads as anxiety and costs the reader three times.
Never lead with a negative. A header that says what the product lacks ("X: not available here") belongs in the caveats block. The same fact can open a section about where X is available.
Current truth, not history. Reference text describes how the thing works now, in the present tense. A fixed defect gets one sentence stating what is true now, with at most a clause of history. Before-and-after tables, mechanisms and "why our tests missed it" belong in the changelog or commit.
A maker line, when the reader cares who made it. One sentence, no adjectives: who, where, how many.
Stop when the content stops. No closing summary, no restatement, no offer.

Sentence rules¶

One idea per sentence, about twenty words, carried by a verb. Start a new sentence rather than joining clauses with a semicolon or an em-dash.
Put the topic in the subject position. When an inanimate subject takes a punchy transitive verb ("every export embeds the list", "the data tells us"), flip it: copula or passive with the topic as subject ("the list is embedded in every export"). Keep the active form only where the thing literally performs the action. This is the opposite of the usual "prefer active voice" advice, which is why it needs stating. Williams' Style (characters as subjects, actions as verbs) is the grounding.
Name the actor. "The tests pass" names nobody, and "the analysis recommends" names a thing that cannot recommend. When an AI wrote the code and a person is the author of the document, the person's "I" does not claim the code. Write what the code does.
Concrete nouns over category nouns. "Kaleidoscopes and flowing patterns" beats "visual experiences". A number beats an intensifier, and every number carries its unit and label, on its own line or in a table when it changes what the reader does.
No flag-planting. "The part worth noting", "the key insight", "here's the thing" and their relatives instruct the reader how to weight the next sentence instead of earning it. They are always deletable. Delete them and move the claim to first position.
No breathless adjectives and no false profundity. "Remarkably", "surprisingly", "stunning", "immersive", "seamless", "powerful" describe the writer's feeling. State the fact, explain why it matters, move on.
Gloss every in-house noun on first use. A preset family, a paid tier or a panel name means nothing to a stranger until the sentence says what kind of thing it is: "17 Looks, ready-made visual styles", "Touch, the paid add-on". A Stranger pass on a store description found this the only defect left after the structure rules were applied.
Product names exactly as the project's naming document spells them, and platform or trademark names only in referential phrases ("for macOS", never "ProductMacOS").
Register rules beat these rules. Scientific methods take the passive, and a hedge there is precision.

Registers¶

Register	Reader and moment	What changes
Store listing, product page	Deciding whether to install. Sees three lines, then scans	Reader contract in the first three lines. One feature list. Price and permissions as lists. Caveats once. No keyword stuffing, no subscription boilerplate for a one-time purchase
Release notes	Already a user. Wants to know what changed for them	One sentence of promise, then bullets under twelve words, benefit first. A fixed defect is one sentence of current truth
UI string	Mid-task. Reads a few words	Identify content, state, consequence or recovery. Never narrate a basic operation. Fix the affordance instead. Errors say what happened and what to do next
Explainer, technical brief	Learning how it works now	Present tense, current state, no "Update (v2.5)" blockquotes. Rewrite the section when the system changes
Blog post	Following a story	Keeps the thinking at the time. Changes go in a dated note at the end
Scientific	Reviewer	Passive in methods, hedging required, numbers with CI and effect size. The anti-slop defaults yield to the genre
Agent report to a person	Did not watch the work. Reads once	Lead with the outcome, and with what could not be verified. No narration of the reasoning, no names invented during the session, code and commands in fenced blocks, numbers in a table, stop when done

Verification¶

Counts. Measure every field against its surface with a script, never by eye: characters for most text fields, bytes for fields platforms meter that way (App Store keywords), lines for subjects. Recount after every edit.
The first-screen test. Cover everything below the fold the surface shows (three lines for a store description, the subject line for mail, the first screenful for a README) and answer the reader contract from what is left.
The deletion test. Remove each sentence in turn. If nothing is lost, it stays out.
Read it aloud. A sentence that needs a breath in the middle is two sentences.
The Stranger pass. Seat the Stranger from references/jury.md on the text alone, brief withheld, and compare its three answers with the reader contract. The Pedant covers labels, units and numeric claims. The seat shares the workspace's memory, so its recognizing a product name is not evidence that a stranger would. Read its disclosure and weight it by its Prior strength. The Squinter, Thumbnail and Loupe do not apply to text.
The linter, then the read. python3 -m muriel.aiism <file.md> catches shape tells (antithesis stacks, tricolons, anaphora, em-dash density). It skips single words by design and reports a flag-planting sentence as clean. Where the science-agent toolchain is installed, science-agent prose-audit <file> --severity=info adds the phrase rules. Canary any linter with a known slop sentence before trusting a clean report, then read the draft anyway: throat-clearing, vague declaratives and false agency are judgements about meaning, and no detector will make them. This channel quotes the tells it bans, so the phrase rules flag its own examples. A quoted tell is the one warning to read past.

Anti-patterns¶

Symptom	What it costs	Fix
A description with no bullets	The scanning reader finds nothing to land on and leaves	One feature list, one price list
The same caveat in three sections	Reads as anxiety and triples the reading cost	One caveats block near the end
A header stating what is missing	The reader meets the product's limit before its benefit	Positive header about where it exists. The limit moves to the caveats
Price and permissions in a paragraph	The reader cannot find what is free	Two lists and one sentence on the model
Adjective stacks, "experience" nouns	Says the writer's feeling, not the product	Concrete nouns, a number with its unit
Keyword stuffing	Reads as a search page, not a product	Keywords in the keyword field, prose for people
Fixed-bug post-mortem in reader-facing text	The defect becomes the project's story	One sentence of current truth
Flag-planting sentence	Asserts importance instead of earning it	Delete it. The claim moves to first position
Inanimate subject, transitive verb	Reads as machine-default and buries the topic	Topic as subject, with a copula or the passive
A closing summary or offer	Restates, and ends on the writer	Stop when the content stops
Subscription boilerplate on a one-time purchase	Legal reflex, wrong product	One sentence: one-time, not a subscription
Narrated operations in UI ("click to explore")	Coaching instead of affordance	Fix the control, then delete the line

Prior art¶

Nielsen, How Users Read on the Web (1997): most readers scan, and text written for scanning measured better in his usability tests. The reader contract and the single-list rule follow from it.
Krug, Don't Make Me Think (2000): omit needless words, which is the deletion test.
Williams, Style: Toward Clarity and Grace (1990): characters as subjects and actions as verbs. The inanimate-subject rule is its application.
Orwell, Politics and the English Language (1946): the six rules, including the one this channel keeps breaking on purpose ("break any of these rules sooner than say anything outright barbarous").
U.S. plain-language guidelines (plainlanguage.gov): audience first, short sentences, headers that carry meaning, one idea per paragraph.
hardikpandya/stop-slop (MIT): the source of the anti-slop list in SKILL.md, which this channel complements rather than repeats.