Scribe

A speaker-agnostic pipeline that turns a sermon or theological lecture into a finished, typeset book-edition chapter — installable as a Claude Code plugin.

Give it a SermonAudio URL (or a local audio file) plus the speaker and series name. It then:

1. Transcribes the audio (mlx-whisper, Apple GPU, large-v3, word-level timestamps).
2. Revises the transcript faithfully in the speaker's own voice (a medium dial — smooths for a reader, keeps the rhetoric).
3. Bookifies it — strips oral deixis, converts handout references into figure pointers, derives headings from the speaker's own numbered points, and flags sensitive/uncertain material for review (see below).
4. Cites Scripture — verifies every quotation against the speaker's default Bible version, switches versions where his exposition leans on specific wording.
5. Marks spoken emphasis — detects acoustic prominence from the word timestamps and italicizes the prose peaks.
6. Finalizes to one Markdown deliverable (…-BOOK-final.md) with a Scripture-permissions footer for the versions used. (An HTML preview is available but optional.)

The engine and the skill are generic. Everything speaker- or series-specific lives in the data project beside the audio — never in this plugin.

How it's layered

Config resolves by walking up from a chapter file: DEFAULTS < speaker.config < series.config.

Layer	Lives in	Holds
Speaker	project root	speaker.config (name, Bible version + lookup URL, whisper-prompt base) and SPEAKER-PROFILE.md — the voice / version / sensitivity ledger
Series	each series folder	series.config (title, numbering, proper-noun seed) and SERIES-*.md precedent files
Sermon	each chapter folder	the audio, the Markdown drafts, the per-sermon audit logs

SPEAKER-PROFILE.md is a ledger: every misheard-name fix, version quirk, and sensitivity call gets appended, so later chapters — and parallel drafting agents — resolve them the same way. By a speaker's third sermon the profile is dense enough that new chapters draft as cheaply as an established one.

Install

claude plugin marketplace add ~/Projects/scribe     # or: github:zoopmaster/scribe
claude plugin install scribe@scribe

Restart Claude Code. The scribe skill loads automatically and triggers on a sermonaudio.com link or phrases like "sermon to book", "book edition", "next sermon".

Usage

Hand the skill the audio plus the speaker and series name — it does the setup itself:

• New speaker → scaffolds the project root (speaker.config + SPEAKER-PROFILE.md) from templates/.
• New series → creates the series folder with its series.config and seeds SERIES-DECISIONS.md.
• Existing speaker/series → reads the config + ledgers and continues.

Output & retention

The deliverable is one Markdown file, …-BOOK-final.md — emphasis as *italics*, scripture as blockquotes, [^footnotes], ![figures](), all portable to any Markdown reader or typesetter. finalize.py produces it from the curated source at sign-off and appends the Scripture-permissions footer.

Two outputs are opt-in and never assumed — the skill asks, and the answer can be saved to speaker.config so a batch isn't re-asked:

• HTML preview (preview: ask | yes | no) — a styled, book-like HTML rendering. When on, the edit hook keeps it fresh; when never generated, the hook leaves it alone.
• Retention (keep: ask | final | final+html | all) — a finished sermon dir accumulates audio, transcripts, intermediate Markdown, and logs. prune_sermon.sh keeps only what you choose (the final md + figures by default) and removes the rest, including the original audio. It always supports --dry-run, and the skill previews the deletion list before acting.

Flags & permissions

Sensitive or uncertain material runs through an agnostic flag process. The categories are fixed (SENSITIVE, ATTRIBUTION, VERSION, BYLINE, TRUNCATION, GRAPHIC, WORDING); the triggers and the reword/keep decision for a given speaker come from that speaker's profile. A non-destructive watchlist (flag-terms.tsv + scan_flags.py) surfaces candidates for the editorial pass to judge; raised flags go into the screen-only proof-checklist and a durable per-sermon FLAGS.md, and lint_series.py fails any flag marker that leaks into the body.

Bible-copyright notices are appended automatically at finalize, keyed by version (bible_permissions.py, override per speaker via bible-permissions.tsv). Verify the notice against the publisher's current wording before sending a book to print.

Layout

scribe/
  .claude-plugin/   plugin.json + marketplace.json
  skills/scribe/    SKILL.md — the method (the editorial judgment you perform)
  scripts/          the engine — speaker-agnostic, config-driven
  hooks/            on_book_edit.sh + hooks.json (refresh an existing HTML preview on edit)
  templates/        speaker.config, SPEAKER-PROFILE.md, series.config, sections.tsv,
                    FLAGS.md, flag-terms.tsv

The engine

Script	Does
series_config.py	layered config resolver (speaker.config over series.config)
transcribe.sh / transcribe_local.sh	mlx-whisper transcription (URL or local file) with repetition-collapse guard
apply_emphasis.py	acoustic-prominence emphasis detector → curated …-BOOK-emph.md
scan_flags.py	grep the speaker's flag-terms.tsv watchlist → flag candidates (non-destructive)
finalize.py	strip screen-only blocks + append permissions → …-BOOK-final.md (the deliverable)
bible_permissions.py	canonical Bible-version copyright notices
build_html.py	optional styled HTML preview of a chapter (opt-in)
prune_sermon.sh	keep only chosen artifacts (audio/intermediates removed); --dry-run first
build_index.py	series contents page (optional grouping via sections.tsv)
build_book.py / build_book_pdf.sh	combined review HTML → PDF
smartquotes.py	straight → curly typographic quotes, in place
lint_series.py	front-matter + cross-chapter consistency lint (+ flag-leak check)

A speaker project (the consumer)

<speaker-project>/            # e.g. a private "axehead-press" repo
  speaker.config              # speaker, bible_version, bible_url_base, alt_version, whisper base
  SPEAKER-PROFILE.md          # voice tics, version policy, sensitivity taxonomy, name ledger
  <series>/
    series.config             # series_title, numbered, whisper_prompt (+ series nouns)
    SERIES-DECISIONS.md …     # standing rulings / registry / house style (as they accrue)
    sections.tsv              # optional contents grouping for build_index.py
    NN-slug/                  # one sermon: …-BOOK-final.md (deliverable) + (pre-prune) audio.*,
                              # …-BOOK.md, …-BOOK-emph.md, FLAGS.md, logs

Per-series batch runners (e.g. batch_transcribe_*.sh) stay in the data project and call this plugin's scripts/ via ${CLAUDE_PLUGIN_ROOT:-$HOME/Projects/scribe}/scripts.

Editing the engine

This repo is the source of truth; the installed plugin is a cached copy. After editing, refresh the cache. Bump the version in .claude-plugin/plugin.json + marketplace.json, commit, then:

claude plugin marketplace update scribe     # re-read this local marketplace
claude plugin update scribe                 # pull the new version

If a refresh ever reports "not found" (e.g. the version was unchanged), reinstall:

claude plugin uninstall scribe@scribe
claude plugin marketplace update scribe
claude plugin install scribe@scribe

Restart Claude Code (or reload) so the updated skill and hooks load.

Requirements

macOS with an Apple-silicon GPU is assumed for the transcription engine (mlx-whisper, whisper-large-v3-mlx), plus ffmpeg and yt-dlp. CPU whisper medium.en is the fallback floor. Rendering and linting are pure Python 3 with no third-party dependencies.