Yaps Meeting Transcription

Yaps Meeting Transcription

Create an editable, speaker-labelled Yaps meeting project from one recording. Keep the project in Yaps so the user can review it visually in the Meeting tab and continue correcting it from the AI client.

Runtime compatibility

Do not try to guess the host or its locality from a user-agent, product name, or another host signal (for example ChatGPT web versus ChatGPT desktop). Test the capability this workflow actually needs: before account, model, dependency, or input-file checks, resolve the local Yaps CLI through the plugin runner. The runner validates it with a bounded, read-only status command. A cloud shell that cannot see the installed Yaps app is not local access to the user's computer.

If the Yaps CLI is unreachable, do not claim that Yaps is uninstalled and do not begin repeated sign-in, model, or permission troubleshooting. Explain that the current AI session cannot reach the Yaps engine installed on this computer. Direct the user to a local-capable session on the computer where Yaps is installed: in Claude, that is Claude Code or the Claude desktop app; from ChatGPT web or a cloud session, download or open ChatGPT desktop and retry in a local-capable Work or Codex session. They can also access this feature directly in the Yaps application. If they are already in a local-capable desktop session, offer Download or update Yaps, ask them to open it, and retry. Stop until local reachability is restored; only then follow the account and CLI readiness steps below.

Private operational diagnostics

Run every local Yaps CLI command and every bundled workflow script through the plugin's scripts/yaps-plugin-runner.mjs. The runner resolves the installed Yaps CLI when the command begins with yaps, preserves the command's normal output and exit status, and leaves a small local diagnostic breadcrumb for the Yaps desktop app:

node <plugin-root>/scripts/yaps-plugin-runner.mjs --action auth.status --stage authentication -- yaps auth status --pretty

Choose a stable, content-free action such as cli.reachability, auth.status, features.list, dictation.setup, transcription.run, or export.create; use the matching stage: reachability, authentication, readiness, execution, or export. In the rest of this skill, every direct yaps / yaps_cli command and provided script invocation means the equivalent runner-wrapped command. If Node or the runner itself is unavailable, continue the requested workflow directly; diagnostics must never prevent the feature from working.

The breadcrumb may contain only plugin ID/version, detected integration host, action, stage, attempt/outcome, duration, and a fixed safe error category. It must never contain the user's prompt or conversation, command arguments, stdout/stderr, credentials, file paths or names, audio, transcript/note text, or raw error messages. It is written only when Yaps has supplied an opaque signed-in owner marker, stays on-device while offline, and is picked up later by the Yaps app. Never create or guess an owner marker.

CLI discovery contract

Always invoke Yaps through scripts/yaps-plugin-runner.mjs; do not locate the binary by hand. The runner honors an explicit path or YAPS_CLI_BINARY, then checks PATH, then the verified Yaps app locations on macOS, Windows, and Linux. It accepts a candidate only after a bounded, read-only status check. Never ask the user to install a separate CLI, edit PATH, or configure MCP for this skill. Never invoke the macOS GUI binary at Yaps.app/Contents/MacOS/yaps. If discovery fails, repeat the runner's specific recovery guidance instead of claiming that the plugin is disconnected.

Account and CLI readiness

Yaps desktop 2.3.848 or newer supplies the current meeting commands, local models, project library, account state, and usage controls. Let the runner resolve and validate the CLI automatically. Run meeting --help after discovery. If it lacks summarize, chapters, ask, speakers, merge-speakers, or add-speaker, update Yaps before attempting those verbs.

Never request Yaps credentials or payment details in the AI client. Yaps has no free tier. Require an active free trial or Yaps Pro, and let Yaps determine trial eligibility and current offer terms.

Keep account summaries product-facing: use account and billing output only to decide readiness. Do not repeat an email address, billing dates, SKU, or an internal basic* plan name unless the user explicitly asks. Describe active paid access only as Yaps Pro, and a trial only as an active free trial. Never promise “no setup”, “no download”, or “no further input” until the feature and dependency checks have actually confirmed that. When checking several paths in one shell command, evaluate each result separately; one missing input must not be reported as a missing Yaps installation.

Account recovery

The runner follows a valid recommended_settings_path automatically and uses it for the requested command. If the signed-in desktop account cache is temporarily incomplete, it safely wakes the verified installed Yaps app and rechecks for a bounded time. Do not copy settings paths, construct app paths, tell the user to edit PATH, or ask them to reconnect the plugin. A different Claude or ChatGPT account email is irrelevant; never compare it with the Yaps email or ask the user to create a second account.

Safe automatic account handoff requires Yaps 2.3.124 or newer. The runner refuses the older credential-based account check and gives update guidance without touching Keychain. auth status must not request a credential or display a Keychain prompt. Never ask the user to enter their macOS login password or approve a credential prompt. If credential_unavailable or keychain_unavailable appears, repeat the runner's app-update guidance. If credential_missing, cached_offline, verification_unavailable, account_cache_incomplete, refresh_failed, or profile_lookup_failed remains after the runner's automatic wake and retry, report its exact network/cache guidance. Do not add a manual reconnection step. Only unauthenticated / signed_out means sign-in is needed.

Follow this order when setup is incomplete:

1. Confirm through the runner that Yaps 2.3.848 or newer is installed. Do not ask the user to open it before downloading models or processing the recording.
2. Run yaps auth status --pretty.
3. Continue only when authenticated is true and status is active.
4. If unauthenticated, ask the user to sign in or create an account inside Yaps, then check again.
5. Do not run auth billing as an automatic gate. For another inactive state, direct the user to Yaps's account screen, which shows any available trial or Yaps Pro renewal without exposing a credential. For platform_mismatch, explain that desktop-compatible access is required.
6. Run yaps features list --pretty and inspect the meeting entry.
7. If Sherpa is missing, explain the model download and ask once for approval. After approval, run yaps features meeting --enable, verify readiness, and resume the original task automatically.
8. On Apple Silicon only, offer the faster long-meeting option and its additional download size. After approval, install it with yaps features meeting --enable --engine moss. Never offer or attempt MOSS on Windows, Linux, or Intel Mac.
9. Resolve the exact recording and confirm it exists.

Choose an engine

• Use auto by default.
• Auto uses MOSS for an imported meeting of at least five minutes only when MOSS is installed, the user did not provide a speaker count, and the machine supports it. Otherwise it uses Sherpa.
• Use sherpa when the user provides the optional number of people, needs cross-platform behavior, or explicitly selects Sherpa. Accept 1–20 people.
• Use moss for a longer meeting on Apple Silicon when it is installed. MOSS detects speakers itself, so do not pass a speaker-count hint.
• Assume one spoken language per recording. Do not split a meeting into per-language jobs.

Transcribe

Use the bundled script. Claude Code exposes the plugin directory through CLAUDE_PLUGIN_ROOT; in another host, resolve <plugin-root> to the installed plugin directory:

python3 "$CLAUDE_PLUGIN_ROOT/scripts/transcribe_meeting_with_yaps.py" <recording> --engine auto
python3 <plugin-root>/scripts/transcribe_meeting_with_yaps.py <recording> --engine auto

Useful options:

--title "Weekly product meeting"
--speakers 4
--engine sherpa
--engine moss
--yaps-cli /explicit/path/to/yaps_cli

For a video, the script first asks Yaps to extract a temporary WAV, then creates the meeting project and removes only that temporary working copy. The durable project keeps its own audio copy.

Return the meeting ID, engine and selection reason, duration, detected speaker count, segment count, project path, and the path of the persisted project audio. Do not claim success when no segments were produced.

Review and correct

Inspect stable segment IDs before editing:

yaps meeting show <meeting-id> --pretty

Correct only the requested segment text:

yaps meeting correct <meeting-id> --segment seg-3 --text "Corrected wording."
yaps meeting correct <meeting-id> --segment seg-3 --text-file "/path/correction.txt"

Use either --text or --text-file, never both.

Reassign one segment to a 1-based speaker number:

yaps meeting assign <meeting-id> --segment seg-3 --speaker 2

Rename that speaker everywhere:

yaps meeting rename-speaker <meeting-id> --speaker 2 --name "Priya"

Export the corrected speaker transcript:

yaps meeting export <meeting-id> --output "/path/Meeting transcript.md"

Never rewrite the project JSON directly. Use the meeting commands so Yaps regenerates all companion transcript artifacts consistently. Suggest opening Yaps → Meeting for waveform playback, visual previews, notes, and larger correction passes.

Review speakers in bulk

Inspect the roster before any whole-meeting speaker edit:

yaps meeting speakers <meeting-id> --pretty

The returned speaker numbers are 1-based and match assign, rename-speaker, and merge-speakers. Present each name, segment count, speaking time, and first line when the mapping is not obvious.

Merge diarization fragments only after the user confirms which labels are the same person:

yaps meeting merge-speakers <meeting-id> --speakers 3,4 --into 2 --name "Priya"

This rewrites every affected segment in the meeting. Never infer identity from voice alone, and never merge on an ambiguous name. Omit --name to keep the target label. To rename without absorbing another label, use rename-speaker rather than an empty merge.

If diarization missed a participant, add an empty roster entry and then assign only the confirmed segments:

yaps meeting add-speaker <meeting-id> --name "Marcus"
yaps meeting assign <meeting-id> --segment seg-7 --speaker 4

Recaps, chapters, and questions

Build the structured on-device recap before requesting chapters:

yaps meeting summarize <meeting-id>
yaps meeting summarize <meeting-id> --template general
yaps meeting chapters <meeting-id>

Summarization can take minutes because it processes the full transcript with the local chat model. Reuse an existing recap by default. Pass --refresh only when the user explicitly requests regeneration or the transcript/template changed materially. If Yaps reports that the chat model is missing, explain the download and obtain approval before installing it. Never represent generated recap text as verbatim transcript.

Ask one meeting a grounded question:

yaps meeting ask "What did we decide about launch timing?" --meeting <meeting-id> --scope meeting

Search every local meeting only when the user requests cross-meeting scope:

yaps meeting ask "Which meetings mentioned data residency?" --scope all

Do not silently broaden a one-meeting question to --scope all. Keep returned meeting/segment citations with the answer and say when the evidence is insufficient.

Generalist Yaps mode

Meeting transcription is this plugin's default focus, not a boundary around what it can do. When the user explicitly asks for another Yaps workflow, use the same resolved yaps_cli rather than making them find another integration. The full local surface is:

status
settings list|get|set|unset
auth status|usage|billing
features list|dictation|cleanup|reading|subtitles|auto-captions|auto-cut|audio-cleaner|text-in-between|background-removal|translation|meeting
vault status|list|get|create|update|move|rename|delete|search|search-semantic|daily-open|create-from-template|history-list|history-restore|pin|folders|tags|mentions|backlinks
speech synthesize (alias: tts)
srt generate
meeting transcribe|show|correct|assign|rename-speaker|export|summarize|chapters|ask|speakers|merge-speakers|add-speaker
captions styles|create|show|correct|replace|split|merge|style|reset|render|verify
cut presets|verify|create|list|show|plan|export-plan|set|redetect|render|delete (alias: autocut)
media extract-audio|remove-background
audio clean
translate
history-list
usage-local

Run <cli> --help and the relevant group help before an unfamiliar workflow. If the session cannot reach the local CLI (for example ChatGPT web), offer a local-capable session — Claude Code, or ChatGPT desktop for a Work or Codex task — or offer to guide the user through the same workflow in Yaps → Meeting.

Friendly completion and discovery

Lead with a warm outcome such as “Done — your meeting transcript is ready,” then give a short summary and the next useful action. Link exports and report useful metadata; do not dump raw JSON or internal plan names. After a successful task, add one compact More with Yaps section with up to three relevant next steps, such as correcting speakers, exporting SRT captions, or saving the transcript to the vault. Skip it after a failure, a decline, or when the user asks for a terse result.

Boundaries

• Confirm the user has permission to transcribe the participants; do not imply Yaps supplies consent.
• Use Yaps Transcription for a plain single-speaker transcript.
• Use Yaps SRT for timestamped subtitle files and Yaps Auto Captions for styled captions burned into a video.
• Use Yaps Video Clipping for dead-space removal and Auto Cut projects; do not treat meeting transcription as a video editor.
• Do not upload the recording to a hosted transcription service as a silent fallback.
• Do not promise perfect speaker identity. Report detected labels, and use corrections or reassignment when the user identifies an error.
• Do not delete the source recording.