Writing Analyze

Mine the session DuckDB index for assistant writing patterns, compare against the user's voice, and propose a diff to plugins/writing/wordlists/*.txt.

Arguments

Forward these from $ARGUMENTS to analyze.ts (see Run):

• --since <date>: restrict the corpus to sessions on or after the date. Default: the full index.
• --model <glob>: restrict to matching model IDs (e.g. *opus*). Default: all models.
• --judge: add the LLM-judge pass over the deliverable corpus. See Meaning-Layer Judge. Default: off.

Prerequisites

Activate the claude-code:session skill first. Run its refresh script to update the index and capture the DB path:

DB_PATH=$(<session-skill-dir>/scripts/refresh.ts --refresh)

The refresh script prints the resolved DB path to stdout.

Build the local voice baseline once, and refresh it as new writing accumulates. It is the comparison surface for deliverable-aware rule health, local-only and never committed, stored in the plugin data directory (CLAUDE_PLUGIN_DATA, else ~/.claude/plugins/data/writing-bendrucker).

# Seed from the already-present delimited corpus (no re-fetch needed)
bun ${CLAUDE_SKILL_DIR}/scripts/ingest-voice.ts --source file --file <data-dir>/voice-baseline/github-prs.txt
# Or fetch fresh merged PRs (designed to add more sources later)
bun ${CLAUDE_SKILL_DIR}/scripts/ingest-voice.ts --source github --author <user> --created 2019-01-01..2025-01-01
# Build the profile the audit reads
bun ${CLAUDE_SKILL_DIR}/scripts/voice-profile.ts

Both scripts write to the plugin data directory, outside the default sandbox allowlist, so run them with dangerouslyDisableSandbox: true (or via a terminal outside Claude Code).

If no profile exists, analyze still runs: deliverable-surface rules fall back to the chat audit and the report flags the baseline as not loaded.

Run

Pass the DB path via --session-db:

bun ${CLAUDE_SKILL_DIR}/scripts/analyze.ts --session-db "$DB_PATH"
bun ${CLAUDE_SKILL_DIR}/scripts/analyze.ts --session-db "$DB_PATH" --since 2026-04-01 --model '*opus*' --top 50
bun ${CLAUDE_SKILL_DIR}/scripts/analyze.ts --session-db "$DB_PATH" --project bendrucker.me --min-lift 7

Run with --help for all flags. --data-dir overrides where the voice baseline is read from (default: CLAUDE_PLUGIN_DATA or ~/.claude/plugins/data/writing-bendrucker). Writes a markdown report to tmp/trope-analysis-<date>.md (override with --out). The report may quote any host in the combined index, so keep it under tmp/ and never paste host-specific content into committed work.

Meaning-Layer Judge

--judge adds an LLM-judge pass over the deliverable corpus (six binary criteria, information-density first). It requires ANTHROPIC_API_KEY, prints a cost estimate before any call, and caps documents with --judge-limit (default 100, cents per run on the default Haiku-class model):

bun ${CLAUDE_SKILL_DIR}/scripts/analyze.ts --session-db "$DB_PATH" --judge

The judge prompt is a versioned artifact (resources/judge/prompt.md). The report records its hash, and numbers from different hashes are not comparable. Judge flag rates are uncalibrated until the #791 labeling passes run (a user checkpoint). The standalone runner covers ad-hoc files, the reproducibility gate, and the #769 heading baseline:

bun ${CLAUDE_SKILL_DIR}/scripts/judge-run.ts files <paths...>
bun ${CLAUDE_SKILL_DIR}/scripts/judge-run.ts gate
bun ${CLAUDE_SKILL_DIR}/scripts/judge-run.ts headings tmp/heading-labels.tsv

See the "Meaning-Layer Judge" section of references/methodology.md for the rubric, prompt versioning, gate, and calibration protocol.

Hook Health

The PreToolUse dispatcher (hooks/pretooluse.ts) appends one JSONL line per run to ~/.claude/writing-hooks/log.jsonl (controlled by WRITING_HOOKS_LOG, see the plugin README). That log is the runtime half of this skill's audit: the wordlist analysis judges rule precision from session history, and the health check judges hook behavior from what the dispatcher actually did.

Run it before or alongside the wordlist audit:

bun ${CLAUDE_SKILL_DIR}/scripts/hook-health.ts
bun ${CLAUDE_SKILL_DIR}/scripts/hook-health.ts --since 2026-07-01 --json
bun ${CLAUDE_SKILL_DIR}/scripts/hook-health.ts --log /path/to/log.jsonl

It reads the default log (plus its .1 rotation) and reports run volume, outcome and tool breakdowns, latency percentiles for the silent hot path, and per-category fire/suppress counts. The report ends with an opportunities list, each naming a concrete fix:

• a category dominating injections points at a precision audit of that rule (sample its firings from session history, then demote, narrow, or retire)
• a category suppressed as often as it fires means the reminder is not changing behavior within sessions
• any ask outcome is a regression (the numbering demotion expected zero) and names a rule to demote
• silent-run p95 over budget points at the dispatcher's checker order and prefilters
• a high skipped-scratch share warrants checking recent Write paths against isScratchPath for swallowed durable prose
• a clean report is itself the signal: after two consecutive stable audits, flip the WRITING_HOOKS_LOG default to off

Fixes land in the plugin (wordlists, detection/, hooks/), then the next audit's log confirms or refutes them.

Metrics

Lift: how distinctive a phrase is to assistant output vs. user text. lift = rate_assistant / rate_user_smoothed, where rates are per-million-token frequencies. A lift of 10.0 means the assistant uses the phrase 10x more per token than the user. The --min-lift threshold (default 5.0) gates new candidate phrases only. Rule keep/remove uses a direct rate comparison plus --min-count, not lift (see methodology).

Session count: distinct sessions containing a phrase. Candidates require session count >= 3 to filter project-specific jargon that dominates a single session.

Output

• Summary stats (corpus sizes, voice-baseline size, rule count, model breakdown)
• Voice delta (per-feature corpus rates beside the local baseline rates, each labeled skill-prescribed, skill-encouraged, or ungoverned so drift points at the right fix: tune the skill or build a detector; aggregate trends only, never per-document flags)
• Proposed removals, each tagged dead (model produced it fewer than --min-count times on the surface where the hook fires it) or not distinctive (baseline uses it at least as often as the model)
• Proposed additions (high-lift n-grams not already covered), each with its voice-baseline rate and a spot-checkable quote
• Rule health table (every entry with type, audit surface, and keep / remove (reason)), plus deliverable quotes for the deliverable-surface tells
• Structural pattern audit (the hook's regex patterns, hit counts across sessions)
• Structural signatures (part-of-speech tag sequences distinctive to the model's deliverable prose; word-independent, so they survive vocabulary drift between model releases)
• Meaning-layer audit (per-criterion judge flag rates with sampled spans, only when --judge is passed)
• Corrective feedback (short human messages naming a writing problem, with the preceding model output)
• Correction candidates (long-assistant, short-user pairs suggesting prose pushback)

Each rule is judged on the surface where the hook fires it. Chat-surface rules (openers, sycophantic patterns, conversational vocabulary) compare the model's chat against the user's chat. Deliverable-surface rules (flowery-phrases.txt, soft-phrasing.txt) compare the model's deliverable prose against the user's voice baseline, so a tell frequent in PR bodies and absent from the baseline reads as keep, not dead. A rule the model uses far more than the baseline is kept even when its lift reads low. Lift is not used for removal: the smo