| name | ad-voice |
| description | Apply a confirmed machine-local personal voice profile when drafting, rewriting, or translating text. Use when the user asks "write like me", "make this sound like me", "use my voice", "adjust this for Slack/Discord/docs/email", "translate this but keep my personality", or invokes /ad-voice. Preserves facts first, carries owner communicative identity through functional language equivalents, applies separately attributed team/community accommodation, and returns one usable draft. Read-only: never learns from the request or changes the profile. Routes profile creation and corrections to ad-voice-tune.
|
| summary | Draft, rewrite, or translate through a confirmed local personal-voice profile. Preserves facts; layers owner identity, community accommodation, and context; never changes the profile. Falls back honestly when evidence is missing. |
| allowed-tools | Read, Bash |
/ad-voice
Apply a confirmed personal voice without changing it. The profile is machine-local; the draft may come from the conversation or an explicitly named file.
Preserve meaning before style. Keep every claim, name, number, date, citation, URL, identifier, commitment, uncertainty, and requested action. Never invent a detail to make the result sound personal. This skill is read-only: ordinary use and user reactions do not update the profile.
Step 1: Frame the draft
Determine the source text or brief, destination, audience, relationship, language, intent, and any hard length or format limit. Infer values already clear from context. Ask only when a missing value would materially change the draft and cannot be grounded.
Build a silent invariant ledger before rewriting. The ledger is for self-checking and does not appear in normal output.
Step 2: Resolve and validate the profile
Run from the current working directory:
node <skill-base-dir>/scripts/voice-profile.mjs resolve [--profile <path>]
node <skill-base-dir>/scripts/voice-profile.mjs validate [--profile <path>]
Resolution order is an explicit --profile path, $AGENTIC_VOICE_DIR/profile.md, then ~/.agentic/voice/profile.md. A path inside any Git repository is invalid. Read references/profile-contract.md before selecting profile rules.
If the profile is absent or invalid, do not load it partially. Apply the generic baseline in references/human-writing-baseline.md, state briefly that personal matching was unavailable when that limitation matters, and route calibration to /ad-voice-tune. Never claim the result matches the user.
Step 3: Select the active slice
Read references/application.md. Load only patterns whose scope and provenance match the request.
Apply in this order:
- factual and safety invariants;
- destination shape and hard constraints;
- owner communicative identity;
- target-community accommodation;
- audience, relationship, and channel context;
- target-language surface conventions.
A later layer cannot erase an earlier one. Team or community evidence may shape vocabulary, formality, brevity, and politeness, but never supplies identity. Cross-language identity uses a functional equivalent, not literal transfer of fillers, syntax, idioms, or learner errors.
A matching rule does not force a rewrite. When the input already satisfies the rule, leave it alone. When personalization has no demonstrated benefit, keep the simpler generic result.
Step 4: Draft and verify
Use the generic baseline to remove model-writing artifacts, then apply the selected profile slice. Compare the result to the silent invariant ledger. Restore any changed or omitted invariant before returning.
Normal output is one usable draft and nothing else. Show analysis, active rules, confidence, or a Portuguese pragmatic gloss only when the user asks or when a material limitation prevents an honest result.
Boundary
ad-voice applies confirmed state and never writes it.ad-voice-tune negotiates sources, proposes corrections, and owns every profile write.- A request to imitate another person is outside this skill. A team corpus informs accommodation only.
