| name | voice-audit |
| description | Audits an existing voice profile skill against recent writing samples, folders, or fresh client/brand evidence to detect drift, stale claims, missing patterns, weak tone modes, generic checklist rules, and unsupported profile assertions. Use when the user asks whether a voice profile is current, good enough to write with, or ready for client/brand publishing. |
Voice Audit
You audit a voice profile against recent writing.
The goal is to catch drift before it shows up in public work. A profile that
worked in February may stop fitting in May after new topics, new positioning,
new clients, or a changed register.
Inputs
Ask for or locate:
- Existing voice profile skill
- Recent writing samples:
- ideal: last 5 published pieces
- acceptable: 3-7 recent samples
- acceptable: folder path containing current writing
- Time period represented by the samples
- Intended use case:
- newsletter
- ghostwriting
- executive posts
- brand content
- long-form essays
- landing pages
- client publishing
- Any known changes:
- new positioning
- new audience
- new topic area
- new client conventions
- more/less formal register
- more/less opinionated stance
- recent feedback from the voice owner or editor
If the profile or samples are not provided, ask for them or search the
workspace before proceeding.
Process
Step 1 - Parse the profile
Extract explicit claims from the profile:
- identity claims
- background and status claims
- default register
- audience assumptions
- core stance
- reader relationship
- signature moves
- tone modes
- evidence style
- opening patterns
- transition patterns
- closing patterns
- vocabulary and syntax rules
- never-do rules
- calibration examples
- format-specific instructions
Turn each claim into something testable.
Bad claim:
The writer is concise.
Testable claim:
The writer typically uses 1-3 sentence paragraphs, opens with a direct thesis,
and avoids throat-clearing.
Step 2 - Parse recent samples independently
Analyze the recent samples before comparing them to the profile.
Extract current patterns:
- recurring openings
- recurring argument structures
- sentence and paragraph rhythm
- default register
- confidence/hedging level
- proof style
- topic vocabulary
- abstraction level
- reader relationship
- emotional temperature
- current closing moves
- new patterns absent from the old profile
Do not let the profile bias this pass. Treat the current samples as the
measurement surface.
Step 3 - Compare profile vs. current writing
Classify each profile element:
- Strong - clearly supported by recent samples
- Thin - partially supported but under-specified
- Stale - used to be true but no longer appears
- Missing - appears in recent samples but is absent from profile
- Unsupported - claimed by profile but not evidenced
- Overfit - too narrowly tied to old examples, slogans, or catchphrases
- Contradictory - profile says one thing, recent samples do another
- Too generic - true of many writers and not predictive of this voice
Pay special attention to whether the profile is a persona or just a checklist.
A profile can be technically correct and still produce mechanical prose if it
is mostly rules.
Step 4 - Detect drift
Look for drift in:
- topic domain
- audience sophistication
- confidence level
- emotional temperature
- proof density
- abstraction level
- commercial positioning
- personal vs. institutional voice
- expert vs. peer relationship
- tactical vs. strategic orientation
- client/brand conventions
- format mix
Also flag drift caused by growth:
- the writer has published many new pieces
- the writer now covers new topics
- the client has revised positioning
- the voice has become more confident or more restrained
- calibration examples are no longer representative
Step 5 - Decide update urgency
Use this decision rule:
- No update needed - profile remains predictive and current.
- Minor patch - small gaps, stale examples, or thin sections.
- Medium revision - several weak sections or new topic/register patterns.
- Major revision - profile no longer represents the current voice.
- Full rebuild - profile is mostly generic, unsupported, or based on the
wrong source material.
If the user publishes actively, recommend audit + update every 6-8 weeks. For
steadier publishing, recommend quarterly.
Output
Produce this audit report:
# Voice Audit Report: [Target Name]
## Verdict
- No update needed
- Minor patch
- Medium revision
- Major revision
- Full rebuild
## Executive Diagnosis
## What Still Works
## What Is Thin
## What Is Missing
## What Is Stale
## Unsupported or Overfit Claims
## Generic Checklist Risk
## New Patterns in Recent Writing
## Recommended Updates
## Priority Patch List
| Priority | Section | Problem | Fix | Evidence |
|---|---|---|---|---|
| High | Moves | Missing opener | Add move | [sample] |
| Medium | Modes | Thin walkthrough | Add rhythm | [sample] |
| Low | Vocabulary | Old topic terms stale | Remove or demote | [sample/date] |
## Evidence Notes
## Suggested Next Step
For each major finding, cite the profile section and the relevant sample.
Explain the mismatch or supporting pattern.
Quality Bar
The audit must be evidence-based and useful for voice-update.
Do not merely say the profile is good or bad. Report what is strong, what is
thin, what is missing, and what to fix.
Prefer sharp diagnosis over politeness. If the profile would produce generic AI
output, say so and identify which sections cause the failure.
