FerroxLabs consume-voice-profile
Cross-role mode skill. Loadable by any specialist drafting in the user's own voice — sales copy, course modules, pitch narrative, brand-voice rules, long-form essays, the user's newsletter or author voice.
When to use
Load this skill when you are about to draft anything user-voiced AND a Voiceprint file exists. Skip the skill if no voice file is in the workspace — work from raw samples instead, or recommend the user run Voiceprint first. If the asset is in a customer's voice (testimonials, persona-driven sales pages quoting buyers), this is the wrong skill — go to research-customer-voice.
Procedure
1. Look for the voice file. Scan the workspace root for *-voice.md or voiceprint*.md. If you find one, read it. If you find none, ask the user for the file path or to paste contents. If they have nothing, surface the gap once: "No voice profile loaded — I can draft from samples or you can run Voiceprint first." Do not pretend.
2. Parse the six sections. Every Voiceprint file ships in the same shape:
- Voice Fingerprint (5–8 distinctive bullets — descriptive, not generative)
- Audience & Purpose
- DO (concrete moves, phrasings, structural habits)
- DON'T (refusals, banned phrases, tics, AI tells)
- Reference Examples (3–5 short excerpts with one-line notes — this is the operational anchor)
- Calibration Notes (when to dial casual up or down, register edges)
3. Frame the draft. Before writing, prepend this load-bearing instruction to your working context: "Use this voice profile when drafting. Match the do's, avoid the don'ts. If I'm asking for something the profile doesn't cover, ask before guessing." That single line keeps you from drifting into AI defaults the second the request gets ambiguous.
4. Draft. Match concrete moves from DO. Avoid every item in DON'T. Treat Reference Examples as the rhythm target — when you finish a paragraph, ask: does this sound like one of those excerpts or like a polite blog post?
5. Read it back. Aloud, or simulated aloud. Reference Examples are your tuning fork. If your draft sounds noticeably smoother, blander, or more upbeat than the samples, revise the lines that drift before you ship.
6. Capture drift. If the user pushes back on a line ("I would never write that"), log the corrected version to voice-notes.md so Voiceprint's next refresh can absorb it. The maintenance loop only works if specialists feed it deltas.
Decision rules
- Use the voice file when the asset is supposed to sound like the user — their newsletter, their LinkedIn post, their author voice in a book, their personal landing page.
- Do NOT use the voice file on customer-voiced assets (testimonial pages, persona ad copy, review-mined headlines). Customer voice comes from research, not Voiceprint.
- When DO and DON'T conflict (rare), DON'T wins. A refusal beats a positive move every time.
- When Reference Examples contradict the user's stated DOs, the samples win. The Voiceprint methodology already enforces this — trust the file, do not relitigate it in chat.
- One voice per asset. If two profiles are loaded (the user + a co-author), name whose voice dominates and which sections belong to whom before drafting.
Anti-patterns
- Loading the user's voice file into a customer-voiced asset — puts the founder's cadence on the buyer's testimonial, which reads as fake.
- Treating Voice Fingerprint bullets as a generation rulebook. They describe the voice; Reference Examples generate it. Drafting from bullets produces an impression of the voice, not the voice itself.
- Mixing two profiles without declaring which dominates. The result is a third voice that belongs to no one.
- Citing the file in chat ("per your voice profile, section 3…") instead of just writing in the voice. Tell-don't-show is the AI tell the user already hates.
- Ignoring the DON'T list. Every banned item in there is a paper cut the user has already complained about — re-introducing one is a regression, not a fresh draft.
Before / after
Brief: "Write the closing line of a LinkedIn post about a product launch."
Without the voice file loaded:
Thrilled to share our latest release — we cannot wait for you to dive deep and see how it's moving the needle.
Three AI tells in fourteen words. Generic, bland, hype-coded.
With the voice file loaded (DON'T flagged thrilled, dive deep, moving the needle as banned; DO listed vulnerable, concrete, anti-hype openers):
Shipped the thing. Here's the part I'm still nervous about.
Same job, opposite shape. Matches the user's actual cadence because the Reference Examples were the rhythm target — not a checklist of adjectives the model invented on the way down.