copywriting

copywriting — write it in the product's own voice

Part of super-ux — see system-map.md. After changes, resolve and run the brand linter as described below.

First action, every time: read the brand pack. docs/brand/voice.md, terminology.md, and the channels.md record for the surface being written.

No pack? Stop and hand off to `brand-voice`. Do not improvise a voice. Guessing it and being wrong costs more than the pause, and a product whose copy was invented surface by surface is exactly the drift this layer exists to remove.

The opt-out is spoken, never assumed: "no brand" / «без бренда» — or "draft it" / «черновиком» — is the operator declining this route: write directly and say the pack was skipped on request, never skip it silently.

Ownership is by FILE, not by directory. brand-voice owns the pack's sources of truth — voice.md, the terminology dictionary, facts.md, the channel playbooks and the locale policy — and this skill never writes those: a term missing from the dictionary, or a number with no row in facts.md, is reported, never invented to finish the sentence — adding it is brand-voice's decision. This skill DOES own `strings.md` (the interface- string registry, written with Status: proposed) and the product text itself. Both live under docs/brand/, but the directory is not the owner — the file is. Writing a strings.md row is this skill's job and needs no extra approval for its LOCATION; a coordinated run claims the file before editing, and voice and facts stay untouched in the same pass.

Triggers

The phrases that route to this skill, in English and Russian. The description keeps only the strongest of them so the skill listing stays inside its budget (2026-10-20); the family routing hook matches this list.

"write copy" / "напиши текст", "App Store description" / "описание для App Store", "store listing" / "описание в сторе", "rewrite this" / "перепиши", "headline" / "заголовок", "CTA" / "кнопка", "microcopy" / "микрокопия", "error text" / "текст ошибки", "this sounds like AI" / "звучит как нейросеть", "post for X" / "пост в твиттер", "reel script" / "сценарий рилса", "video hook" / "хук ролика", "build a landing page" / "сделай лендинг" (the copy for it; the visual layer is sheleg-design's).

References

Modes

The humanization pass runs by default, in every mode that produces text. It is not a mode you enter, it is the last step of Write, Edit and Adapt, under the guards in ai-tells.md which are not optional. The reason it is a default rather than an option: a draft nobody swept carries the markers that file grades, and a reader registers them before they can name why.

voice.md records the state in two fields that answer different questions. `Humanization: on | off` is whether the pass runs, and it defaults to on when the field is absent. `Humanization pass:` names which implementation runs, and absent it is own, the only one that reads this pack's registers and canonical facts. Turning it off is a legitimate decision that outlives whoever made it, so off carries a Humanization declined: line with the reason and the date. B064 checks all three states.

Every delivery of copy states what happened, in one line, before the copy or immediately after it:

Humanization: on (own pass): 7 markers at 2 densities, 4 addressed, 11% changed
Humanization: off (declined 2026-08-30): wording fixed by counsel

The line is not decoration. A pass that runs invisibly is indistinguishable from one that did not run, and the reader of the copy is usually not the person who chose the setting.

Write

1. Name the surface. It must exist in channels.md; if it does not, that is

a brand-voice decision, not an improvisation here. For marketing work only, read marketing-copy.md → Before writing and seed/update the current brief from templates/brand/copy-brief.md. Keep unresolved inputs unknown; no extra brief load for interface-only work.

1. Apply the register: the axes from voice.md, plus that surface's deltas.

Deltas move axes; they never cross the invariants.

1. Write. Every claim traces to facts.md. Every product term comes from

terminology.md.

1. Deliver the copy first, then the reasoning. For headlines and CTAs give

two or three options with what each trades away.

1. For interface strings, add or update the strings.md row — key,

file:line, scenario, Status: proposed.

1. Run the humanization pass under the Humanize guards below, then print

the status line. On new text the change rate is measured against the draft you just wrote, so the 50% ceiling does not apply the way it does to an edit; report the rate regardless.

Edit

The seven sweeps from marketing-copy.md, in order, looping back after each: clarity, voice and tone, so-what, prove-it, specificity, emotion, zero risk. Deliver sweep by sweep, prioritised by impact, not by reading order.

The humanization pass runs last, after the seven, and then the status line. Last because the sweeps rewrite whole sentences and a humanization pass run before them is measured against text that no longer exists; and because the semantic-preservation checklist is cheapest to apply to a version nobody is about to rewrite again.

Adapt

One piece across several surfaces. Re-write per surface — never paste one register into another. What survives adaptation is the claim and the proof; what changes is length, structure, CTA policy and register.

State plainly what each version drops. A thread is not a page with line breaks.

The pass runs per surface, not once on the source. Each version is different text in a different register, and a marker density that is fine in a blog post is not fine in a landing hero. One status line per surface.

Humanize

The standalone mode, for auditing text that already exists without writing any: an inherited page, a competitor's copy, a draft somebody else wrote. The same guards govern the pass wherever it runs, and they are not optional:

• Above ~10 markers per 500 words, say so and rewrite from the argument — a

patch produces the same patterns with better words.

• Above a 50% change rate, do not ship. Report the rate and ask. That is

no longer an edit.

• Run the semantic-preservation checklist before output: numbers, dates and

proper nouns intact; causal direction unchanged; no negation inverted; quotations untouched; the core claim the same.

• Text that already reads naturally is left alone. Editing what is fine to

prove the pass ran is this mode's failure.

• A marker count is not a verdict and never gates anything. Say which markers

are present at what density; never say a text was AI-written. A number of markers does not prove authorship — the false positives fall hardest on people writing in a second language, and a writer is not a defect to be edited into fluency they did not ask for. The humanization pass is ADVISORY throughout: B060 warns, it does not error. ai-tells.md carries the measurement and what it binds. The one exception is the em dash (AT-06): it is a hard ban on text you write, never a count, and it is listed under Non-negotiables.

• A quote, a registered term, and an explicit `off` never force a rewrite. A

marker inside a quotation is the source's word; a term in terminology.md is meant to recur; and Humanization: off (with its reason) preserves the text exactly. Brand bans are a separate user policy — a forbidden word is the brand's own choice in terminology.md, not a machine-drafting tell, and it is that check's business, not this pass's.

• Read `voice.md`'s two fields first. Humanization: is whether the pass

runs and defaults to on; Humanization pass: names which implementation, and absent it is own. Neither absence stops work: run the default, print the status line saying it was a default, and offer to record it once. A value naming a tool that is not installed falls back to own and says so — a missing optional tool must not stop copy being written.

• Other implementations exist and two are worth knowing —

npx sshlg-skills humanizers lists what this machine has. Reach for one for an audit with no rewrite, for long-form prose, or when the writer has a sample of their own writing to match. Stay here for product copy: the brand pack's registers and canonical facts are the constraint, and a general-purpose humanizer does not read them.

Non-negotiables

• No em dash (U+2014) in text you write, in any language (owner decision,

2026-10-18). Not a frequency target: one is an error to fix before delivery, and no substitute stands in for it (--, an en dash or a spaced hyphen doing a dash's job). Use a full stop, a comma, a colon or brackets, or rewrite the sentence. The rare exception is a dash Russian grammar needs where no rewrite reads naturally. Quotations, code, commands, paths and historical records keep theirs. B062 and B067 check what a machine can prove; the rest is ai-tells.md AT-06.

• No decorative full stop in headings or labels, including nested HTML,

formatted Markdown and staccato fragments such as “Your agents. Your tools.” Read the rendered role, not merely a flat copy export. Keep real questions, ellipses, abbreviations, versions, URLs and normal paragraph punctuation. A project may extend this rule to hero summaries and captions; record that scope in its brand policy and test the generated page too. When restoring historical copy, reapply the current project punctuation policy; an old exact-text snapshot is not an intentional exception.

• No fabricated facts, statistics, quotes or experts. Not for a deadline,

not for a benchmark, not because a placeholder would look better. Refuse, say why, offer to find a real one.

• **No humor on error, destructive confirm, billing and receipts or

paywall and upgrade** — in any pack. The user is losing something there.

• One action, one name. Before naming an action, search strings.md for

it. A second name for an existing action is a defect, not a synonym.

• Never quote a number that is not in `facts.md`.

Definition of done

• Every string or section traces to a surface record and a voice.
• The humanization status line is printed, whichever state it reports.
• New interface strings are in strings.md with location and scenario.
• Resolve the linter before running it: use docs/brand/lint.py if present;

otherwise locate the installed super-ux linter brand_lint.py in its scripts directory and pass the project's docs/brand path. Do not invent a project-local script or assume the shell exports a plugin-root variable. If neither exists, name the missing tool and the manual coverage; never call it a lint pass.

• Confirm Sources: resolves to the touched surfaces. HTML heading roles must

survive a copy projection; flat paragraphs do not prove heading coverage. Run the resolved script with --fail-on B063 when title punctuation is a required gate (or --strict for all warnings). An older script that rejects the flag needs an update, not a retry that quietly drops the gate.

• A tone-of-voice checker sweep, beside the linter, never instead of it.

Gate on the tool in the session, not on config: when a voice-check MCP tool is present (one exposing *.check and *.voices.list), run it on each public text you wrote with the text's locale, the project's saved voice as voiceRef when its voice list has one, and the register and format its schema offers. The target is register: web with format: web-page for a page; until the schema lists them, use register: marketing (ad or store on those surfaces), omit format, and say the web-page pack was not available. Report the report id and address or reject each finding: they are advice, like B060. With no tool, print voice checker: not connected, brand_lint only and continue.

• Report scanned surfaces, remaining warnings and the exact command/exit code.

Exit zero with advisory warnings is not a clean review. Inspect the rendered copy and generators; a source-only check cannot see external CSS or runtime text.

• Anything reported as missing — a term, a fact, a surface — is named

explicitly, not worked around.

What else is in `references/`. This skill names 13 contract(s) directly; references/ holds 15, because a contract this skill names links others and the whole closure ships inside the skill — the skills CLI installs one directory and a sibling's file would arrive dangling. Open the ones named here; the rest are reached from them, by name, when a contract sends you.