voice-review
Runs parallel prose and craft review agents against a voice profile. Use when checking generated content for AI patterns and voice drift before publishing.
How do I install this agent skill?
npx skills add https://github.com/athola/claude-night-market --skill voice-reviewIs this agent skill safe to install?
- Gen Agent Trust Hubpass
The voice-review skill analyzes text for quality and voice consistency by passing content to specialized review agents. It handles external text which introduces a risk of indirect prompt injection, where malicious instructions within the text could attempt to subvert the review process. It also maintains local snapshots of reviewed text in the user's home directory for its learning feature.
- Socketpass
No alerts
- Snykpass
Risk: LOW · No issues
What does this agent skill do?
Voice Review Skill
Dispatch dual review agents and present unified findings.
When NOT To Use
- Producing the text (use
scribe:voice-generate) - A generic AI-pattern scan with no voice profile (use
scribe:slop-detector)
Method: Parallel Dual-Gate Review
Two agents run in parallel on the generated text:
- Prose reviewer: AI patterns, banned phrases, voice drift
- Craft reviewer: Naming, destinations, dwelling, devices, anchoring
Hard failures (banned phrases, em dashes) are auto-fixed. Everything else returns as advisory tables for user decision.
Required Progress Items
voice-review:text-loaded- Generated text readvoice-review:register-loaded- Voice register loadedvoice-review:agents-dispatched- Both reviewers launchedvoice-review:hard-fails-fixed- Auto-corrections appliedvoice-review:advisories-presented- Tables shown to uservoice-review:findings-verified- Citations confirmed by verifier
Step 1: Load Context
Read:
- The generated text (from file or clipboard)
- The active voice register
- The banned phrases list
Step 2: Dispatch Review Agents
Launch both agents in parallel:
Agent(prose-reviewer):
- text: {generated_text}
- register: {register_content}
- banned_phrases: {banned_list}
Agent(craft-reviewer):
- text: {generated_text}
- register: {register_content}
Step 3: Process Results
Hard Failures
Apply all auto-fixes from prose reviewer silently:
- Remove/replace banned phrases
- Replace em dashes with appropriate punctuation
- Rewrite negation-correction patterns
Report: "Fixed N hard failures (X banned phrases, Y em dashes, Z patterns)"
Advisory Tables
Present both tables to the user:
Prose Review Advisories:
| # | Line | Anchor | Pattern | Current | Proposed fix |
|---|
Craft Review:
| Dimension | Rating | Notes | Proposed improvement |
|---|
Step 4: User Decision
For each advisory row, user can:
- Accept (a): Apply the proposed fix
- Reject (r): Keep the current text
- Rewrite (w): Apply a custom fix
Present as:
[1] Prose: Frictionless transition at "Furthermore, the..."
Proposed: Cut transition, start mid-thought
[a]ccept / [r]eject / re[w]rite?
Step 5: Apply Decisions
- Apply accepted fixes to the text
- Skip rejected items
- For rewrites, incorporate user's version
- Save final text
Step 6: Snapshot (if learning active)
If the user has learning mode enabled:
- Save "post-review" snapshot (text after hard-fail fixes, before user decisions on advisories)
- Save "post-fixes" snapshot (text after user decisions)
- Both go to
~/.claude/voice-profiles/{name}/learning/snapshots/
Integration with voice-generate
When dispatched from voice-generate, the flow is:
- voice-generate produces text
- voice-generate calls voice-review
- voice-review dispatches agents, processes results
- User makes decisions on advisories
- If learning mode: snapshots saved for later comparison
Standalone Usage
Can also be run on any existing text:
/voice-review path/to/file.md --profile myvoice --register casual
Verify Findings Are Grounded (voice-review:findings-verified)
Write advisory rows to .review/findings.json, run the citation verifier
(Skill(imbue:review-core) Step 5), and drop or label UNVERIFIED any
the verifier rejects.
Verification
After the review completes, validate these conditions:
- Both review agents returned results (no timeouts)
- Hard failures auto-fixed and diff shown to user
- Advisory tables presented with accept/reject/rewrite options
- User decisions applied to the final text
- Final text saved to disk
- Snapshots saved (if learning mode active)
Exit Criteria
- Both review agents returned results without timeout
- Hard failures auto-fixed and diff shown to user
- Advisory tables presented with accept/reject/rewrite options
- User decisions applied to the final text
- Final text saved to disk
- Every advisory row carries a
Line(file:line) and verbatimAnchor;citation_verifier.pyconfirmed all citations (exit0) or unverified rows are dropped/labeledUNVERIFIED
Test Spec
The test suite (test_voice_review.py) validates:
- Skill file exists and references parallel dispatch
- Hard failure vs advisory separation is documented
- Prose reviewer agent exists with hard-failure patterns
- Craft reviewer agent exists with five-dimension ratings
- Both agents produce tabular output for downstream merging
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/athola/claude-night-market/voice-review">View voice-review on skillZs</a>