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:

1. Prose reviewer: AI patterns, banned phrases, voice drift
2. 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 TodoWrite Items

1. voice-review:text-loaded - Generated text read
2. voice-review:register-loaded - Voice register loaded
3. voice-review:agents-dispatched - Both reviewers launched
4. voice-review:hard-fails-fixed - Auto-corrections applied
5. voice-review:advisories-presented - Tables shown to user
6. voice-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:

1. voice-generate produces text
2. voice-generate calls voice-review
3. voice-review dispatches agents, processes results
4. User makes decisions on advisories
5. 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)

Every advisory row must cite a real line and a verbatim anchor. Write findings to .review/findings.json and confirm each citation resolves:

bash
python plugins/imbue/scripts/citation_verifier.py \
  --findings .review/findings.json --repo-root .

Drop or label UNVERIFIED any finding the verifier fails (exit 1); only verified findings enter the advisory tables. See Skill(imbue:review-core) Step 5 and Skill(imbue:structured-output) for the schema.

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 verbatim Anchor;

citation_verifier.py confirmed all citations (exit 0) or unverified rows are dropped/labeled UNVERIFIED

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