brand-voice — define and hold the verbal identity
Part of
super-ux — see
system-map.md for the
whole pipeline. After changes, run
python3 docs/brand/lint.py
.
decides
what the product does. This skill decides
how it
speaks, and owns
in the target project.
Format contract: brand-contract.md
(brand-contract v1). Read it before writing or editing anything under
. Never deviate from its file names, field names or statuses —
the linter and the audit key off them.
The pack library: voice-packs.md — six
archetypes, each with a declared failure mode. The register model:
surface-registers.md. Locales:
localization.md.
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.
"tone of voice" / "тон оф войс", "brand voice" / "голос бренда", "brandbook" / "брендбук", "our terminology" / "наша терминология", "how should this sound" / "как это должно звучать", "the copy is inconsistent" / "текстовка разная везде", starting any marketing surface, onboarding a product into docs/brand/.
Invoked with no task → status, then one next action
- Report the state: is a pack recorded, what says its
is, how many facts carry or no source, locale parity, and the
linter's error and warning counts.
- Name every unresolved fact. Never invent one to close a gap — an
unknown is reported, not filled.
- Propose exactly one next action, usually the top open fact or the
surface the user actually came for.
Where the voice comes from
The pack is derived from
— personas, jobs, and the
deciding input:
what the user loses when the product fails. A voice that
charms when the stakes are a playlist is intolerable when the stakes are
payroll.
The dependency runs one way. A persona never changes because a tone was
appealing. With no foundation, work in degraded mode: stamp
, say plainly that the WHY layer should be built, and
continue rather than blocking.
The route itself has a spoken opt-out: an operator saying "no brand" /
«без бренда» declines it — proceed without the pack and say so, rather
than dropping it silently.
Choosing a workflow
| Situation | Workflow |
|---|
| No | Init |
| Pack chosen, product specifics missing | Calibrate |
| Positioning, personas or facts changed | Update |
| Consistency questioned, or before a launch | Validate |
Announce which one you are running.
Init
- Read the foundation. Personas, JTBD, journeys, stakes. Where it is
absent, interview: who is this for, what are they accountable for, what
does failure cost them.
- Read what already exists. Interface strings, landing copy, README,
store listing, recent posts. The current voice is data even when nobody
chose it.
- Propose one pack with reasoning, plus one alternative. Name why the
runner-up loses. Let the user pick.
- Seed from this skill's own , and fill the
block first — nothing else can be checked until the linter
knows where the project keeps its text.
- Inventory sweep for : routes, screens, buttons, states,
errors, empty states. Every string gets a row with its and the
scenario it serves. For a large codebase, use available read-only subagents
in parallel, one area each; without delegation, inspect the same areas
sequentially and keep the same evidence requirements.
- Ask which humanization pass this project wants, and write the answer to
's field. is the default and the only one
that reads this pack's registers and facts;
npx sshlg-skills humanizers
shows
what else is installed and how to install what is not. Ask once — the field
exists so nobody is asked twice, and it selects a pass, never a verdict.
- Everything starts .
Calibrate
The pack is a starting position; this is where it becomes the product's own.
- Terminology. Which generic words does this product replace, and with
what. Seed the banned table from weak verbs, hedging chains and
ai-tells.md, then add what is specific here.
- Facts. Every number the product quotes, with a source, a checked date
and a review date. Mark internal figures .
- Channels. One record per surface the product actually has. Register as
deltas; always carries both the physics and the brand half.
- Ready lines. Six to ten, written for this product, not copied from the
pack.
- Failure mode. Copy the pack's into so the audit can look for
it by name.
- The em dash invariant. Seed 's invariants with "no em dash
(U+2014) in new text, in any language" (owner decision, 2026-10-18). It is
not a voice choice a pack can relax: every pack inherits it, and only the
grammatical Russian dash where no rewrite reads naturally survives
(ai-tells.md AT-06, , ).
Present section by section for approval. Approved moves
→
.
Update
- Identify what changed — positioning, a persona, a price, a tier name, a
new surface.
- Apply the edits. A changed entity name propagates to ,
and every locale in the same change.
- Changed files drop to until re-approved.
- Re-stamp .
Validate
- Integrity — every field the contract requires is present; statuses are
legal; resolves against the foundation.
- Coverage — every surface the product ships has a record; every locale
declared has a file; every quoted number has a fact.
- Consistency — run
python3 docs/brand/lint.py
and fold its findings
in. It proves the mechanical half; report what it cannot see as questions
for .
- Staleness — foundation newer than the last calibration, facts past
review, physics rules past their checked date.
Definition of done
- Contract honored; the marker on every file.
- The user has seen and approved new or changed sections.
python3 docs/brand/lint.py
exits 0, or every remaining finding is
reported with a reason.
- Unknowns stated as unknowns. Nothing invented to make a table look full.
What else is in . This skill names 6 contract(s) directly;
holds 13, 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.