Style Writer

Style Writer

Write content in any stored writing style. Discovers available styles from the writing-styles/ collection, loads the appropriate guide into context, applies it during writing, and optionally self-evaluates the result against the style's rubric.

When to Use

• User wants to write content in a specific stored style
• User asks to "list available styles" or "what voices do we have?"
• User wants to edit or rewrite existing text to match a style
• User asks to evaluate or score text against a style rubric

Style Discovery & Selection

List directories in writing-styles/ (exclude _template/). Each directory is an available style.

Selection logic:

Context Loading

Load compact files first — voice card + do/don't list give enough context for most tasks.

Read files using the Read tool.

Writing

Apply the loaded style to the user's content request:

1. Use the voice card as the primary compass
2. Cross-check against the do/don't list while writing
3. For long pieces: write in chunks, periodically re-read the voice card to prevent drift
4. Incorporate style-appropriate transitions, sentence patterns, and rhetorical moves

Fiction: Narrative Authenticity

When writing fiction or creative writing, also load references/narrative-authenticity.md and:

1. Pre-write: Review the pre-writing checklist. Plan structural choices (timeline, subplots, character moral complexity) before drafting.
2. During writing: Vary scene types and emotional intensity. Resist the pull toward clean single-track plots, explicit thematic statements, and embodied-metaphor-only emotion.
3. Post-write: Run the post-writing checklist. Audit for AI narrative defaults — especially Claude-specific fingerprints (flat event escalation, epilogue endings, low event diversity).

If the style guide includes narrative construction notes (from fiction extraction), those override the general guidance in narrative-authenticity.md. The style guide captures THIS author's narrative choices; the authenticity guide captures general AI defaults to avoid.

Task-Specific Approach

Self-Evaluation

Triggered when user asks "evaluate", "score", or "how well does this match?"

1. Load evals/style-rubric.md
2. Score the output on each of the 5 rubric dimensions (1–5)
3. Report total score with per-dimension breakdown
4. For fiction: Also score on 5 narrative construction dimensions (structural diversity, thematic subtlety, emotional range, character agency, event/temporal complexity). See the Narrative Construction Evaluation section in references/evaluation-guide.md.
5. Suggest 2–3 specific improvements

See references/evaluation-guide.md for detailed scoring guidance and common pitfalls.

Anti-Patterns

• Don't over-load context. Start with voice card + do/don't, escalate to full guide only if needed.
• Don't apply style mechanically. The voice card captures the spirit, not a formula.
• Don't ignore the do/don't list. It catches the most common style violations.
• Don't mix styles. Apply one style consistently per piece.
• Don't skip evaluation. If the user cares about accuracy, score the result.

References

• references/evaluation-guide.md — Detailed self-scoring guide with common pitfalls and improvement strategies
• references/narrative-authenticity.md — AI narrative debiasing guide with Claude-specific fingerprints (fiction only)