/voice-learn

name: voice-learn
description: Improves a voice profile by learning from manual edits. Use after editing generated text to refine registers and close voice drift over time.
globs: "**/*.{md,txt}"
alwaysApply: false
category: writing-quality
tags:
- voice
- learning
- improvement
- feedback
- iteration
tools: []
complexity: high
estimated_tokens: 2000
progressive_loading: true
modules:
- modules/snapshot-management.md
- modules/pattern-analysis.md
dependencies:
- scribe:voice-extract
- scribe:voice-generate
- scribe:voice-review

Voice Learning Skill

Learn from user edits to improve the voice profile over time.

When NOT To Use

• Building the first profile (use `scribe:voice-extract`)
• Reviewing text without changing the profile (use

`scribe:voice-review`)

Method: Three-Stage Comparison

Every piece flows through three stages: 1. **Pre-review**: Raw generation output (before review agents) 2. **Post-review**: After user accepts/rejects advisory fixes 3. **Post-edit**: User's manually edited final version

The learning agent compares stages 2 and 3 (post-review vs post-edit) to identify patterns in what the user changed. These patterns inform register and rule updates.

Core Rules

1. **Sharpen, don't add**: Modify existing rules to cover new patterns. Rule bloat degrades output. 2. **Tag specificity**: Register-specific patterns go to registers. Universal patterns go to craft rules or agents. 3. **Flag contradictions**: Opposite patterns across pieces require user resolution. 4. **Evidence threshold**: Patterns need 3+ instances (or 1-2 matching existing accumulator entries) before becoming rules. 5. **Detection surface**: Structural changes increase AI detectability. Craft-level changes are neutral. Prefer craft-level updates. 6. **Rule count check**: Suggest consolidation if any section has 8+ rules.

Required Progress Items

1. `voice-learn:snapshots-loaded` - All three stages read 2. `voice-learn:diff-analyzed` - Changes categorized 3. `voice-learn:accumulator-checked` - Prior patterns reviewed 4. `voice-learn:proposals-generated` - Updates proposed 5. `voice-learn:user-approved` - Changes accepted by user

Step 1: Load Snapshots

Load: `@modules/snapshot-management`

PROFILE_DIR="$HOME/.claude/voice-profiles/{name}"
SNAP_DIR="$PROFILE_DIR/learning/snapshots"

# Find the most recent snapshot set
# Format: {piece-name}-{timestamp}-{stage}.md

Read all three stages for the target piece.

Steps 2-4 need Opus or larger: telling a tone adjustment from a precision edit takes the nuance Sonnet flattens. When this session runs Sonnet or Haiku, run them as one dispatch:

Agent(general-purpose):
  - model: opus
  - prompt: {Steps 2-4 of this skill, the three snapshot paths,
    and learning/accumulator.json}

The subagent sees nothing from this conversation and reads the snapshots from disk. It returns the proposals and does not write the profile. Step 5 approval stays in this session. On an Opus or Fable session, run the steps inline.

Step 2: Diff Analysis

Load: `@modules/pattern-analysis`

Compare post-review vs post-edit. Categorize every change:

| Category | Example | |----------|---------| | Tone adjustment | Softened a claim, added hedge | | Voice insertion | Added parenthetical, aside, humor | | Structure change | Broke paragraph, reordered | | Precision edit | Replaced vague with specific | | Deletion | Removed fluff or decoration | | Addition | Added context, example, anchor |

Step 3: Check Accumulator

Read `learning/accumulator.json`:

{
  "patterns": [
    {
      "id": "pat-001",
      "category": "tone_adjustment",
      "description": "Softens confident claims about tool capabilities",
      "instances": [
        {"piece": "blog-post-1", "date": "2026-04-08", "diff": "..."}
      ],
      "target": "register",
      "status": "accumulating",
      "first_seen": "2026-04-08",
      "last_seen": "2026-04-08"
    }
  ],
  "staleness_threshold_days": 30
}

Match new changes against existing patterns:

• Semantic similarity (same category + similar description)
• If match found: merge instance, check if threshold reached
• If no match: create new accumulator entry

Step 4: Generate Proposals

For patterns that reach threshold (3+ instances or 1-2 matching prior accumulator entries with 2+ instances):

Apply (strong evidence)

## Proposed Update

**Pattern**: {description}
**Target**: {register file or craft-rules.md}
**Evidence**: {N instances across M pieces}

| Piece | Date | Change Made |
|-------|------|-------------|
| ... | ... | ... |

**Proposed edit**:
- File: {path}
- Section: {section name}
- Current: "{current text or 'new addition'}"
- Proposed: "{new text}"

Hold (insufficient evidence)

Add to accumulator with current instances. Report:

Holding: "{pattern description}" (N instances, need 3+)

Contradictions

If a new pattern contradicts an existing accumulator entry:

Contradiction detected:
- Existing: "{accumulator pattern}"
- New: "{contradicting pattern}"
- Resolution required: user must choose

Step 5: User Approval

Present proposals to user:

Learning found N patterns ready to apply:

[1] {pattern}: {proposed change}
    Evidence: {N instances}
    [a]pply / [s]kip / [v]iew evidence?

[2] ...

Apply approved changes to the target files.

Staleness

Patterns in the accumulator expire after `staleness_threshold_days` (default 30). If a pattern hasn't recurred within that window, it was likely a one-off preference rather than a voice trait.

On each learning pass, prune stale entries:

# Remove patterns older than threshold with < 3 instances

Snapshot Capture

The learning system captures snapshots automatically when voice-review completes. Snapshot naming:

{piece-filename}-{YYYYMMDD-HHMMSS}-pre-review.md
{piece-filename}-{YYYYMMDD-HHMMSS}-post-review.md
{piece-filename}-{YYYYMMDD-HHMMSS}-post-ed