---
name: "standard-writing"
type: standard
description: "The quality bar for every written deliverable: plain English, a point of view, concrete detail over claims, and no model-sounding prose. Read before writing or reviewing any copy, deck text, article, case study, proposal, email or message, and cite it from any skill that writes. Includes the plain-English and AI-tone check that is run on a draft before it is shown."
license: CC BY-NC-SA 4.0
---

# Writing standard

The bar a piece of writing has to clear before the Owner or a client sees it. It is a standard, not a convention: more than one choice is possible, one is better, and a piece passes or fails by judgement. It scopes to **all written deliverables** (deck copy, site and product copy, articles, case studies, proposals, outbound, email, documents, messages), whichever agent writes them. Draft holds it (`agent-draft`); any skill that writes cites it and runs the check below, so each citing skill carries one line and no rubric of its own.

## What passes

- **Plain English.** Short words, short sentences, the verb doing the work. A reader outside the field follows it on the first read.
- **A point of view.** The piece says something a reader could disagree with. A line that describes what a product does and takes no position fails.
- **Concrete over claimed.** A named detail, a number, a quote or an example beats an adjective. If the evidence is not in the brief, flag the gap; never fill it from inference.
- **A regulated-status claim comes from a named source that names the entity.** Authorisation, registration, licensing, membership, accreditation, insurance, a guarantee: write it only where a named source says it of **this** entity — the regulator's own register, the entity's own filings, or a document the client has supplied. It is never inferred from the sector the entity works in, from what a competitor says of itself, or from what the claim would plausibly be. An unconfirmed one is **marked in the draft and raised**, never softened: a hedge still reads as a claim, because a reader takes a regulated status as a fact however it is qualified, and the cost of getting one wrong lands on the client rather than on the draft. This is the one case where the general rule above is stated in terms, because it is the case where inference feels safest. (Worked case: a regulatory authorisation was written into the first draft of an ad set on 2026-10-27 from the sector alone and removed only after the Owner questioned it — **reported**, APP-2717.)
- **Matter-of-fact confidence.** Plain speaking and precision carry the tone. No hype, no urgency that is not real, no inflation.
- **House mechanics.** Sentence case, British spelling, no exclamation marks, outcome before adjective, unless the brand's voice reference says otherwise.
- **Right voice.** The piece reads as the brand it is for. The brand's voice is read from its `<stem>.voice-reference` document (`convention-brand`); a piece that ignores it fails even if it reads well.
- **On brief.** It does what the brief asked for, for the audience the brief named.

## The plain-English and AI-tone check

Run it as a separate pass on a finished draft, after the writing and before anything is shown. Read the draft cold, against the brief and the voice reference, and mark each of these:

- "X, not Y" and its variants ("it's not just X, it's Y"), which set up a contrast nobody drew. The form stays only when it is the best way to land the point: both sides are concrete things the reader recognises, and the draft says in one line why the contrast beats a direct statement. Used as a default, it is a mark to fix. Used with that justification, it passes.
- Lines that fit any product or any brand: swap the name and nothing changes.
- Stock words and phrases that signal generated prose (unlock, seamless, elevate, leverage, in today's fast-moving world, game-changing).
- Three-item lists and paired adjectives used for rhythm instead of meaning.
- Openers that announce the point and closers that restate it.
- Hedges stacked on a claim, or a claim with no evidence behind it.
- Sentences all the same length, or paragraphs all the same shape.

Each mark is fixed or defended in writing. A draft with unresolved marks does not go to the Owner. The check is judgement against the bar above; the list is where to look, and it does not define the bar.

**The check tests how a line sounds, not whether it carries anything. Add the summary test.** Every mark above is a mark of generated prose: a contrast nobody drew, a stock word, a rhythm list, a stacked hedge, a uniform sentence. A draft can carry none of them and still say nothing — plain, varied, unhyped and empty — and a reader meets that draft as well written and comes away with nothing. The list cannot catch it, because each mark asks how a line reads and none asks what it carries. Run this as the last mark of the same pass, on the same finished draft: **summarise the draft from the draft alone**, in a tenth of its length or less, covering what it asserts and what the reader is meant to do or believe; then read the draft against that summary, line by line. **Any line the summary drops with no loss is padding** — cut it, or defend it in writing as any other mark is defended, and a defence names what the line carries that the summary could not: a figure, a name, a date, a concession, a transition the argument needs, or a line of voice the brand's reference asks for. The pass is not a ratio and not a word count: the draft passes this mark when no dropped line is left undefended. Write the summary before re-reading the draft and do not edit it to fit what you then read — a summary written alongside the draft becomes a description of it and drops nothing. **This mark belongs to the whole-document read and fires with it**, on the rendered artefact before the export or the send, and it resolves through the gate already set below: an undefended dropped line is one more unresolved mark, not a second gate. (Worked case: an AI summary of a site read better than the site it summarised, and no mark in the check as written would have been raised against a line of it — **reported**, APP-2884.)

**The whole-document read is a gate on export or send, not a stage in drafting.** The check above already states the pass and the point in drafting at which it runs, so the rule is not the gap. The **trigger** is. Nothing binds the pass to the moment the artefact leaves, so a run that has validated every edit believes it has validated the document. Bind it here: **the whole-document read fires on the rendered artefact, in the form the reader will receive it** — the exported PDF or DOCX, the published page, the deck as it will be presented, the message as it will be sent — and it fires **before** the export or the send, never on the way out and never after. **Edit-level and structural checks do not discharge it.** Each edit can be correct and each sync can fit its frame with pixels to spare, and a pile of correct edits composes into a document nobody has read. The defects that survive edit-level validation are exactly the ones only a continuous read catches: a figure restated inconsistently across sections, a repetition invisible while the sections were checked apart, a section in the wrong order, a misspelled company or client name, a typo in a line that was never the line under review, and a non-content layer — a spacer, a placeholder, a guide — that looked like nothing on screen and renders in the export. A document validated edit by edit and never read end to end has not been checked; it has been assembled. One instance warrants this because of where it lands: the gate is the last place a defect is cheap, and the first reader past it is the Owner or the client. (Worked case: the 2026-10-22 CV session — a document plus two Figma versions — was validated per edit and per structure throughout, and the first whole-document read happened on the exported PDFs in front of the Owner. Six defects surfaced at once: placeholder spacer blocks exporting as black boxes, "customers needs", "Over 8+ years", a wrong agency name, Education placed above Selected impact, and one figure repeated four times. Several were already visible in screenshots taken earlier in the session and were not raised. **Reported** — APP-2629.)

## Where examples and rules live

Good examples, rejected examples and rules learned are kept per brand in the brand's voice reference, not here. A rule belongs in this standard only when it holds for every brand.