Shownotes

Turn a developer livestream MP4 into a local transcript and ready-to-paste YouTube publishing copy. Use the local web interface to choose from five title candidates and two descriptions, or use the CLI for a repeatable pipeline.

Video stays on your computer. ffmpeg extracts audio and whisper.cpp transcribes it locally. Transcript text is sent to Anthropic for one structured-output request. The app does not upload videos to an AI service or publish anything to YouTube.

Requirements

• Node.js 22.12 or newer and npm.
• ffmpeg and whisper.cpp (whisper-cli), installed separately.
• A local ggml Whisper model. large-v3-turbo works; its q5_0 quantized variant uses less memory and is a practical choice on memory-constrained Macs.
• 1Password desktop app with CLI integration enabled, and the op CLI on PATH.
• An Anthropic API key stored in 1Password with API billing available.

No model or system binary is automatically downloaded by Shownotes. On Apple Silicon, use a whisper.cpp build with Metal support. Transcription can consume substantial memory; avoid running multiple jobs simultaneously.

First setup

npm ci
cp .env.example .env.local

Edit .env.local with a reference, not a plaintext secret:

# @initOp(allowAppAuth=true)
# ---

ANTHROPIC_API_KEY=op(op://YOUR_VAULT/YOUR_ITEM/YOUR_FIELD)
WHISPER_MODEL=/absolute/path/to/ggml-large-v3-turbo.bin
WHISPER_BINARY=whisper-cli

Replace the placeholder with the secret reference copied from your own 1Password field. Keep the 1Password app unlocked and enable its developer CLI integration. Varlock’s official 1Password plugin resolves the reference when you start the application; it may ask you to authorize desktop access. No op read, export, or plaintext secret file is needed.

If 1Password reports multiple accounts, choose the account that contains your vault and add its address or account ID as a static argument in the ignored .env.local header, for example:

# @initOp(allowAppAuth=true, account=YOUR_ACCOUNT)
# ---

Keep only one @initOp header. The installed plugin requires a static account argument and does not forward OP_ACCOUNT to its CLI subprocess. Account selection belongs in local configuration, not the shared schema.

The checked-in .env.schema declares the plugin and marks the API key required and sensitive. .env.local is ignored by Git. Both startup paths disable persistent Varlock caching. The web server still receives runtime variables through the Varlock wrapper. The CLI resolves only its four configuration keys through Varlock’s graph API and passes the API key directly to the Anthropic client without writing it into process.env. The API key is never sent to the browser or to ffmpeg/Whisper subprocesses. Tests use a synthetic override and do not access 1Password.

For this workstation, local binary/model paths and the supplied secret reference have already been configured. Do not overwrite an existing .env.local with the example.

Start the local interface

npm start

Open http://127.0.0.1:4317 at the default port. If you configure PORT, open http://127.0.0.1:<PORT> or the URL printed at startup instead; localhost at the same port is also accepted. This command builds the app and launches the server through Varlock.

1. Drop one MP4 onto the upload area, or use the keyboard-accessible file picker.
2. Leave Reuse a matching transcript checked to avoid repeat transcription. Uncheck it to regenerate the transcript.
3. Leave Include chapters checked for automatic 5–8 chapters with a ten-second minimum, choose an exact count (1–8) and minimum duration (1–3600 seconds), or uncheck it for short clips. Select Generate shownotes. Progress reports upload, extraction, local transcription, and candidate generation. It reports stages, not a predicted completion time.
4. Select a preferred title and description. Copy buttons remain disabled until a choice is made. Copy selected description includes chapters when enabled. Chapterless results hide the chapter section and copy descriptions without an empty timestamp block.
5. Copy tags and, if needed, the separate chapter list. Clipboard success or failure is announced. If the browser denies clipboard access, select and copy the visible text manually.

Keep the page open during processing. One recording can run at a time. Uploads stream to disk, default to a 4096 MB limit, and must have an MP4 filename, media type, and ISO media header. ffmpeg performs the actual media decoding. The server binds only to 127.0.0.1, checks the Host and Origin, and requires a per-session token for API requests. Do not expose it through a proxy or tunnel.

The web app saves output under output/<video-sha256>/ and shows the exact path after completion. Selecting a candidate affects clipboard output; the saved Markdown keeps candidate 1 as its default and includes all alternatives. Choices are not persisted across page reloads.

Stop with Ctrl-C. Normal errors and graceful shutdown remove temporary uploads/audio, while intentional transcripts and briefs remain. An OS crash or forced kill can leave .upload-* or .shownotes-* directories; remove these only while the server is stopped. Never delete a directory belonging to a running job.

CLI

Build once after checkout or source changes:

npm run build
npm run brief -- ./recording.mp4 --out ./output

The brief script runs the compiled video-brief CLI directly. Normal processing resolves configuration with Varlock and its 1Password plugin inside the CLI; help and diagnostic maintenance skip resolution. The resolver uses the installed Varlock internal graph API, covered by real-schema tests; verify those tests when upgrading Varlock. Help does not require secrets:

node dist/cli.js --help

Options:

Option	Behavior
--model FILE	Override the local model path from WHISPER_MODEL.
--out DIRECTORY	Output directory; defaults to beside the input.
--skip-transcribe	Require an existing valid cache; fail clearly if missing or stale.
--force-transcribe	Ignore a matching cache and transcribe again.
--no-chapters	Generate titles, descriptions, and tags without chapters.
--chapter-count N	Exact count 1–8, or auto (default: 5–8).
--chapter-min-seconds N	Minimum chapter duration 1–3600 seconds (default: 10).
--help, -h	Show usage.

Matching cache entries are reused automatically even without --skip-transcribe. Do not combine the skip and force options. The CLI chooses title candidate 1 and description candidate 1 as readable defaults; alternatives remain in the brief.

For recording.mp4, the output bundle contains:

• recording.transcript.txt: the reusable plain-text transcript.
• recording.transcript.json: versioned cache with segment text, millisecond timestamps, video SHA-256, binary configuration, and model path/size/modification time.
• recording.brief.md: default publishing copy, chapters, tags, all five titles, and the alternative description.

The cache validates video content, the binary setting, and the configured model identity. Changing any of these invalidates it. Manually replacing a model while preserving both its size and modification time requires --force-transcribe. Cache-only CLI runs can omit a model path; when a model is configured, it must exist and match. Same-name CLI inputs in one output directory overwrite the previous bundle after successful processing; their hashes prevent accidental transcript reuse. Use separate output directories when you want to retain both.

Transcripts are saved before requesting Claude, so API failure does not waste completed transcription. A failed run preserves an earlier brief; that earlier file has not been regenerated. Rerunning distillation makes another paid API request. Automatic SDK retries are disabled.

Transcription failures and diagnostic cleanup

Read failures distinguish missing from unreadable Whisper JSON. Parsing errors distinguish invalid JSON syntax from segment schema and timestamp validation errors; segment indexes are zero-based. Messages never include transcript text. Validation is not relaxed or malformed output repaired automatically.

On these output errors, any raw whisper.json and whisper.txt emitted by Whisper are retained alongside ownership/lifecycle metadata in a unique .whisper-diagnostics/run-.../ directory under that run’s requested output directory. The error reports its location. Web jobs therefore retain diagnostics under output/<video-sha256>/.whisper-diagnostics/. Cleanup attempts to remove temporary audio and uploaded MP4 copies. Audio cleanup failure does not prevent diagnostic finalization; the reported error preserves the original transcription reason and retained location alongside cleanup failures. Successful runs remove their raw diagnostic files after saving the intended transcript cache. Diagnostic directories are ignored by Git and contain private text; do not publish them.

After diagnosing the problem, use Retained diagnostics → Review diagnostic files in the UI. It lists file names, byte sizes, and run state without reading transcript text into the browser. Select Delete this diagnostic run and confirm the displayed scope. Saved transcript/brief files are preserved. The server rejects cleanup while a recording is processing, and active diagnostic runs are protected.

CLI maintenance requires no key, Varlock authentication, model, or ffmpeg. After building, list retained metadata first:

npm run diagnostics -- --out ./output

Delete one listed run deliberately, substituting its exact ID and scope:

npm run diagnostics -- --out ./output --delete RUN_ID --scope VIDEO_HASH

For CLI runs directly in the chosen output directory, omit --scope (default .). The list includes both direct runs and web runs in immediate video-hash subdirectories. node dist/cli.js diagnostics --help shows the same command without npm. There is no bulk delete or automatic expiry. Cleanup refuses symlinks, unknown files, invalid ownership/metadata, traversal paths and active runs. An interrupted/crashed run may remain marked active; it is deliberately excluded from automated deletion and needs manual inspection with the app stopped. Never edit ownership metadata to bypass these safeguards.

Diagnostic operations require a POSIX filesystem, directories/files owned by the current user, and no group/other write permission on the output or diagnostic directories. Canonical ancestors must be owned by the current user or root and not writable by other users; sticky ancestors such as /tmp are allowed. Metadata is opened with O_NOFOLLOW and checked through the file handle. Cleanup unlinks only recognized files and removes the empty run directory; it never recursively descends into unexpected subdirectories. Listing ignores a vanished run only when ENOENT is confirmed for that run itself; malformed metadata, permission failures and symlinks still fail.

This protection assumes trusted same-user processes and no ACL grants that let other accounts mutate the tree. It is not descriptor-relative isolation from a malicious process running as your user, root, or an ACL-authorized writer. Use a private local output directory. If unexpected files appear during deletion, cleanup may stop after removing recognized files; inspect what remains with the app stopped.

The earlier 64-minute failure discarded raw output, so its root cause is unknown. For a short speech-heavy test clip, omit chapters. If a valid transcript cache exists, retry the same MP4 with transcript reuse checked, or use --skip-transcribe --no-chapters in the CLI. If transcription itself failed, no valid cache exists: transcribe again, keep any retained diagnostics until the exact error has been investigated, then clean them up using the UI or CLI.

Configuration

Put non-secret overrides alongside the reference in .env.local:

Variable	Default
WHISPER_MODEL	Required for fresh transcription; no automatic download.
WHISPER_BINARY	whisper-cli
CLAUDE_MODEL	claude-haiku-4-5-20251001
SHOWNOTES_OUT	output (web interface only)
PORT	4317
MAX_UPLOAD_MB	4096

The app always requests five distinct titles, two distinct descriptions of two or three paragraphs (at most 1200 characters per paragraph), and five to ten tags. Chapters default to an automatic 5–8 with a ten-second minimum; you can request exactly 1–8 chapters, choose a minimum of 1–3600 seconds, or omit them entirely.

Before calling Claude, the app checks whether the requested count fits at the actual transcript segment boundaries, rounded down to seconds, including the final chapter’s duration. Impossible settings fail locally with guidance to omit chapters or reduce the count/minimum. The validated transcript remains cached, and changing chapter settings does not invalidate it. Feasible timing does not guarantee suitable topic changes or a valid model response: returned chapter IDs, count, spacing, and final duration are still checked.

For a cached short clip:

npm run brief -- ./recording.mp4 --out ./output --skip-transcribe --no-chapters

For three shorter timestamp entries:

npm run brief -- ./recording.mp4 --out ./output --chapter-count 3 --chapter-min-seconds 5

Counts below three or durations below ten seconds produce timestamp lists that do not meet YouTube’s manual chapter requirements. When enabled, the first chapter starts at 00:00; Claude selects supplied segment IDs and timestamps are derived locally. When omitted, the returned chapter array is empty and descriptions/Markdown have no appended chapter block.

The request allows up to 8192 output tokens; truncation and incomplete/malformed responses still fail without an automatic retry. This is headroom, not a guarantee for every language or response. Review generated copy before publication; no paid quality evaluation is included in the automated suite.

Development and verification

npm run quality

Runs Vitest, Oxlint, Stylelint, Oxfmt checks, TypeScript checks, and the build. Default tests do not need models, media, secrets, vault access, or paid requests. They cover subprocess arguments/errors/cancellation, Whisper JSON parsing, SDK structured-output parsing with a mocked transport, cache invalidation, cleanup, loopback upload protections, DOM selection/copy, drag/drop, submission and error flows. UI tests use jsdom, not a real browser.

Opt-in real-binary test on macOS (generates public-safe speech using say; no user recording or API call):

SHOWNOTES_INTEGRATION=1 WHISPER_MODEL=/absolute/path/to/model.bin npm test -- tests/integration.test.ts

Set WHISPER_BINARY too if it is not on PATH. The test encodes generated speech as MP4, extracts PCM WAV, runs Whisper, and asserts recognized speech and timestamps. Allow several minutes for model loading on a constrained machine. Each subprocess has a five-minute cancellation bound.

Modules are separate and take explicit inputs: audio.ts, transcription.ts, and distillation.ts. pipeline.ts manages caching/output; cli.ts and server.ts share it. Browser modules import no Node runtime dependencies.

Calavera Composer supplied TypeScript, Oxlint, Oxfmt, Stylelint Standard/Baseline, Varlock, and the checked-in skills. Its recipe, artifact lock, and managed state are retained; downloaded package caches are ignored. Project-specific package scripts and TypeScript settings were intentionally adapted for Node ESM, browser modules, tests, and a separate build. Preview future Calavera reapplication and reconcile these intentional changes before applying it. quality deliberately does not resolve live secrets.

Verified upstream contracts

• whisper.cpp CLI parser and JSON writer: text/JSON flags, extensionless output prefix, and segment millisecond offsets.
• Anthropic TypeScript Messages source and Zod helper: messages.parse, output_config.format, original-schema validation.
• Varlock 1Password plugin: plugin initialization, desktop authentication and op() references. The installed parser is exercised with synthetic overrides in tests.
• YouTube chapter requirements: first timestamp 00:00, ordered chapters and a minimum ten-second length.

MIT licensed. No automatic upload, speaker diarization, remote hosting, or automatic model download.