CtrlK
BlogDocsLog inGet started
Tessl Logo
PluginPublic

jbaruch/speaker-toolkit

Twelve skills for conference speakers covering rhetoric analysis, talk authoring, deck building and export, CFP submissions, publishing, illustrations, shownotes, screencast production, and vertical shorts, informed by a 113-entry Presentation Patterns taxonomy (83 observable: 64 patterns + 19 antipatterns; 30 unobservable: 21 patterns + 9 antipatterns).

Quality

92%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Medium

Suggest reviewing before use

Overview
Quality
Evals
Security
Files

local-media-acquisition.mdskills/vault-ingress/references/

Bounded Local-Media Acquisition

Read before using local audio/video or the YouTube Whisper fallback. The public entrypoint remains fetch-transcript.py; its module docstring owns arguments, JSON output, exit codes, and transcript-replacement authorization.

Runtime Gate

Require the database-configured interpreter's lanes before using them:

"{python_path}" "{speaker_toolkit_root}/skills/vault-ingress/scripts/check-runtime.py" \
  --lanes core,source-media,whisper --require-lanes core,source-media,whisper

For the YouTube download fallback, include youtube-download in both lists. For probe-only library work, require core,source-media without Whisper. check-runtime.py's LANE_REQUIREMENTS owns the lane dependencies; pyproject.toml owns their versions. A failed optional lane does not invalidate independent captions or slides.

Acquisition Ownership

  • Pass the original local locator to fetch-transcript.py --audio; never pre-open, hash, hydrate, copy, or invoke ffprobe on the source yourself.
  • local_media_evidence.py's probe_local_media owns availability, generation binding, snapshot hashing, probe facts, and private-workspace cleanup.
  • local_media_contract.py's named MEDIA_* constants, admission helpers, and bounded_whisper_result own supported containers and resource ceilings.
  • local_media_transcription.py's transcribe_local_media accepts a current in-memory media/video probe from its owner. It reuses those facts and rechecks source generation. A serialized or hand-constructed receipt is not authority to skip acquisition.
  • transcribe_local_words uses the same acquisition owner for bounded interior samples. Read sampled-word-evidence.md before using this independent word-evidence lane; it does not write catalog transcripts.
  • local_media_download.py's download_youtube_audio yields a private local artifact plus provider duration and removes the workspace when its consumer exits. It does not register or preserve a recording in the vault.
  • Worker arguments contain no source locator or model path. Authenticated private request payloads carry them; diagnostics remain bounded and redacted.
  • Full Whisper acquisition uses greedy decoding without previous-window text prompts. Its bounded result gate rejects extreme unspaced periodic letter runs with whisper_repetitive_text; see skills/vault-ingress/scripts/local_media_contract.py — WHISPER_REPETITION_POLICY_VERSION and _reject_repetitive_whisper_text. This is a decoder-loop guard, not a complete hallucination detector. A refusal returns no truncated transcript and preserves the prior bundle. It does not migrate an existing quality receipt or authenticate a recording's catalog identity.

Declaring an Audio Recording

A transcript Whisper'd from a local audio file binds its local_media_duration receipt only after the owner declares that file as the talk's recording with the declare_local_audio mutation; see schemas-db.md. The declaration binds transcript and timing receipts, never delivery_video evidence.

A recording over the MEDIA_MAX_INPUT_BYTES ceiling in skills/vault-ingress/scripts/local_media_contract.py cannot be acquired whole. Extract its audio stream first:

"{python_path}" "{speaker_toolkit_root}/skills/vault-ingress/scripts/extract-local-audio.py" \
  --recording "{recording_path}" --output "{vault_root}/recordings/{talk_stem}.m4a"

Exit 0 prints {ok: true, output, written, media_sha256, size_bytes, duration_seconds, mode}: output is installed and owner-probed, and written: false means identical audio was already there and was left as is. Exit 2 prints {ok: false, reason_code, error} with the message on stderr: nothing was installed, an existing output was never replaced, and the staged file was removed (or the error names it). Follow the error's action; when it says the codec does not fit, re-run with a .wav output, and for output_exists choose another output name. Exit 3 is an unexpected failure: stdout is empty, stderr carries the closed failure document, and nothing was installed; report it. The script's docstring owns the output modes, bounds and reason codes. .m4a copies the stream. Transcribe the reported output with --audio, then declare it by its vault-relative path (recordings/{talk_stem}.m4a). Never rename a local transcode to {youtube_id}.mp4 to satisfy the YouTube filename rule: that fakes provider identity.

Refusal and Recovery

The fetcher reports worker refusal through its existing failure JSON and stderr. Use LocalMediaError.reason_code for library callers. Repair unavailable runtime lanes, restore a locally available supported source, or retry after source edits finish. Never bypass worker admission or increase a threshold from skill prose.

A source just written into a Google Drive vault can have its change time touched while Drive finishes the upload, often mid-transcription. A drift that moves only the change time is re-verified by digest rather than discarded: the worker decodes a private snapshot verified against the established source_sha256, so the source's later metadata changes never touch the speech, and the owner re-probes before committing and keeps the work only when the bytes still match. Any other drift, and any byte difference, still refuses with media_generation_changed.

A failed local probe, transcription, resource check, cleanup, or precommit generation check leaves prior transcript, quality, and timing artifacts intact. This includes a stale quality receipt: failure does not authorize weakening its provenance. --force authorizes transcript replacement, not bypassing validation.

Acquisition proves source bytes and duration, not that a recording belongs to a catalog talk. Keep the separate owner-binding preflight gate; do not patch a receipt, infer a missing recording, or relabel old analysis to clear it.

skills

vault-ingress

SKILL.md

README.md

tile.json