Prose Mechanics

Diagnostic audits for sentence-level prose quality on near-final drafts. Each audit is a focused pass that produces a flagged-issues report; the author reviews and applies fixes before moving to the next audit. Works for fiction or nonfiction.

Session Continuity

Audit work is iterative across multiple sessions. Claude has no memory between them, so the audit tracker is your persistent state.

• At session start: Read audit-tracker.md in the project directory (copy from assets/audit-tracker-template.md if it doesn't exist). Check which audits have run on which chapters and what's still outstanding.
• At session end: Update the tracker with what was audited, what was flagged, and what was resolved. Note the stopping point.
• Surface unresolved questions: If prior audits flagged issues the author hasn't addressed, list them at session start so the user can decide what to take on.

Core Concept

Audits are diagnostics, not edits. Each pass scans for one specific class of failure, produces a list of flagged passages, and stops. The author—not the audit—decides whether each flag is a real problem and how to fix it.

When to Use

This skill is for:

• ✅ Sentence-level audits on finished or near-finished drafts (fiction or nonfiction)
• ✅ Final polish before submission, query, or publication
• ✅ Catching the mechanical failures that survive developmental editing
• ✅ Complementing fiction-workshop or narrative-nonfiction work after structural editing is done

When NOT to Use

This skill is NOT for:

• ❌ Early drafting—audits are for revision, not generation
• ❌ Developmental editing tasks—use fiction-workshop or narrative-nonfiction instead
• ❌ Manuscripts that haven't stabilized structurally yet (you'll audit prose that's about to be cut)
• ❌ Single scenes or short passages where eyeballing is faster than running a pass
• ❌ Heavy stylistic departures (e.g., experimental prose) where the audits' assumptions don't hold

Audit Passes

Run audits one at a time, in order. Order matters because earlier audits change what later audits flag. Recommended sequence: frequency-family first (they inform the crutch-words config), mechanical passes next, semantic passes last (judgment-expensive; don't run on prose about to be reworked).

Load only the reference file matching the currently invoked audit. Do not preload all references at session start — it wastes context budget. If switching audits mid-session, load the new reference file and treat the prior one as out-of-scope.

Finding Format

Every audit emits findings conforming to ../../references/finding-schema.json: audit, technique, severity (note/suggestion/warning), location (file/line/quote), issue, exemplar (optional), confidence (deterministic/judgment). Render findings grouped by severity, resolving any exemplar reference against references/exemplars/<ref>.md.

Engine Hook

Before running any audit, check command -v scriptorium. If it succeeds:

• Deterministic audits: run scriptorium prose audit <name> <chapter>

and treat its JSON as authoritative — do not re-derive findings by hand.

• Hybrid audits: run scriptorium prose prepare <name> <chapter> to

get code-detected candidates, judge each per that audit's reference file, then scriptorium prose submit-findings <state> --findings-file <file>.

• Semantic audits: run scriptorium prose prepare <name> <chapter>

(no candidates — pure judgment over the raw chapter), judge per the embedded instructions, then submit-findings.

If scriptorium is unavailable, perform the audit conversationally using the reference file's detection rules directly — this is the audit's full specification, not an abbreviated fallback.

Workflow

For each audit:

1. Scope: Confirm which chapter(s) or passage(s) to audit. Audit one chapter at a time unless the user asks for broader scope.

1. Run the pass: Read the prose looking for the specific patterns named in the relevant reference file. Do not look for other classes of issue—keep the lens narrow.

1. Produce a flagged-issues report: For each flagged passage, output:

- File and approximate location - The flagged text (short excerpt) - Why it was flagged (which detection pattern matched) - A suggested fix or two—not a mandate, just options

1. Stop. Wait for author review. Do not apply fixes automatically. The author decides which flags are real and which are stylistic choices.

1. Apply approved fixes: Use str_replace for surgical edits on the flags the author confirms. Skip the rest without comment.

1. Update the tracker: Note which audit ran, on which chapter, what was flagged, and what was resolved.

1. Move to the next audit. Repeat.

Common Mistakes

Self-Check: Is This Working?

Use these checkpoints to verify you're running audits correctly.

Before starting an audit:

• Is the manuscript structurally stable (no pending major cuts or reorders)?
• Have prior developmental and character-work passes been completed?
• Is the audit scope clearly defined (which chapter, which passages)?

During an audit:

• Are you using only the patterns from the relevant reference file?
• Are you producing a flagged-issues report rather than applying fixes directly?
• Are you ignoring issues outside the audit's scope, even when you notice them?

After producing the report:

• Can the author act on each flag without re-reading the audit reference?
• Did you include a short excerpt and a suggested fix for each flag?
• Did you stop and wait for review rather than continuing to the next audit?

After applying fixes:

• Did you use str_replace for surgical changes only?
• Did you skip the flags the author didn't confirm, without re-raising them?
• Did you update the tracker before ending the session?

Before moving to the next audit:

• Is the current audit's flags fully resolved (applied, dismissed, or deferred with a note)?
• Have you confirmed the author wants to proceed?

If you answered "no" to any checkpoint, stop and resolve it before continuing.

Stopping Points

Each audit has a defined end. Stop at it. Do not auto-advance to the next audit, do not silently expand the audit's lens, do not apply fixes without explicit author confirmation.

If the author explicitly asks you to continue past a stopping point (e.g., "just apply the obvious fixes"), fine—but name which flags you're acting on and which you're skipping, so the scope shift is visible.

Quick Reference Commands

Files

• references/active-passive-audit.md - Voice and agency at the sentence level
• references/parallel-structure-audit.md - Grammatical parallelism in lists, comparisons, series
• references/sentence-length-variance.md - Rhythm diagnostics for paragraphs
• references/accessibility-audit.md - Readability scoring and structural accessibility
• assets/audit-tracker-template.md - Per-project tracker for audit runs and findings