agentsclimarketplace

Voicekeeper

Skill knowledgeco/airworthiness/skills/voicekeeper

Fidelity and granularity controls for AI coding agents — the judgment a seasoned engineering team would apply, made explicit. Execution control + surface control. Model-agnostic. MIT.

Install
npx -y skills add knowledgeco/airworthiness --skill voicekeeper

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • 27 days oldThe repository was created 27 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Preserve product identity through interface language across screens, states, models, and implementation sessions. Use when creating or changing navigation labels, buttons, forms, validation, errors, warnings, confirmations, empty states, onboarding, background-processing or offline messages, permissions, destructive actions, assistant output, notifications, public release copy, or canonical product terms; when naming a new concept; or when auditing generic SaaS language and terminology drift. Discover the repository's voice constitution and term inventory, protect exact operator language, apply context-specific voice rules, keep state and error copy operational, detect prohibited or competing terms, and leave tone approval to the operator rather than claiming a model-generated voice is approved.

SKILL.md

5.1 KB, as published. Nobody here has run it

Voicekeeper

Product voice is an interaction system, not a late wording pass. Preserve the same relationship, terminology, certainty, and rhythm wherever the product speaks—including when it is blocked, degraded, or wrong.

Why this exists

Interface copy is often generated one screen at a time. Each string can be grammatical while the product as a whole shifts between teacher, assistant, institution, and generic SaaS template. Terms fork, errors lose the product's posture, and deliberate operator language gets “corrected” into house style. The visual system may remain recognizable while the product's relationship to the user dissolves.

Voicekeeper makes language constraints durable and testable without inventing a universal tone.

The core move: inherit the voice, then write the state

  1. Discover local authority. Read the repository's voice constitution, term inventory, surface constitution, content source, and relevant state manifest. Follow an established local format rather than replacing it with the portable one.
  2. Resolve the speaking context. Identify the surface, user state, consequence, technical depth, and whether language is protected, operator-authored, generated, or still pending.
  3. Use canonical terms. Apply preferred names, capitalization, pluralization, definitions, and migration rules. Do not create synonyms for variety when one concept already has an owned name.
  4. Write operationally. Make labels predictable and errors recoverable. Where applicable, state what happened, what remains safe, what the user can do, and what the system will do next. Adapt that structure when local voice or space constraints require it.
  5. Protect exact language. Preserve declared capitalization, cultural or spiritual terminology, coined names, punctuation, and syntax exactly. Do not normalize intentional language against generic style rules.
  6. Audit the inventory. Search interface copy for prohibited words, unowned alternatives, inconsistent casing, unsupported promises, accidental anthropomorphism, and tone drift across normal and degraded states.
  7. Route judgment honestly. A model may propose alternatives and explain tradeoffs. Only the operator may record a tone, naming, or cultural-language verdict as approved.

Use references/voice-constitution-format.md for the portable constitution, references/interface-copy-inventory.md for collection and drift checks, and references/errors-states-and-naming.md for state copy and naming mechanics.

Hard boundaries

  • Do not optimize for variety over term ownership. Repetition is correct when it preserves one product concept.
  • Do not fabricate approval. Leave operator tone, naming, and exact-language decisions pending until supplied.
  • Do not flatten protected language. Generic grammar and capitalization conventions never override an explicit exact rule.
  • Do not hide uncertainty. Match promises to measured system behavior; generated copy may not imply guarantees the implementation cannot support.
  • Do not separate error voice from product voice. Stress states are part of the interaction system, not exceptions written by infrastructure.

Output contract

Create or update the smallest local artifacts that preserve the voice: a constitution, canonical term inventory, copy inventory, pending decisions, and targeted findings. Use schemas/voice-constitution.schema.json where the repository adopts the portable schema. Do not duplicate strings already governed in a source catalog; reference them by stable key.

Optional composition

Voicekeeper is standalone. SurfaceCharter can supply content posture, Statewright can enumerate the states that need language, SpecimenFirst can compare writing directions, SurfaceWitness can audit copy substitution, and HandOffTracker can carry a consequential naming fork. None is required.

Adapting across agents

Keep the constitution and term inventory in the repository, not in one model's memory. Any agent can apply the same constraints if it receives those artifacts; automatic triggering is the only harness-specific feature.

A note on this file

Keep universal voice discipline here and detailed formats in references/. Every addition should prevent term drift, a false promise, lost exact language, or a fabricated operator verdict.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.