agentleFS
fiveonecodefiveonecode/agent-skills/meeting-transcription/SKILL.md

meeting-transcription

Skill20 starsChanged 2 months agoMITView on GitHub

Transcribe long local meeting audio into a validated rough transcript with bounded chunking, repetition and output-limit detection, targeted retries, optional stronger-model rescue, and auditable token usage. Use when a user provides an audio recording or asks for reliable multilingual meeting transcription.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

9 rules. About 0 are common, 7 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈7

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Run the bundled script from the installed skill directory:

    line 47
  2. Distinctive:

    no persistent Gemini file retention.

    line 62
  3. Very specific:

    Use --prepare-only to verify source metadata, hashing, and chunking without API calls.

    line 64
  4. Distinctive:

    Only suspect five-minute chunks are split into one-minute retries.

    line 76
  5. Distinctive:

    Treat output as a rough AI transcript, not certified verbatim evidence.

    line 95
  6. Distinctive:

    Use generic speaker labels unless identity is grounded independently.

    line 97
  7. Very specific:

    Prefer [unclear], [inaudible], or [overlapping speech] over invented content.

    line 98
  8. Distinctive:

    Keep raw audio, chunks, and API envelopes outside downstream repositories unless the user explicitly requests approved retention.

    line 99
  9. Distinctive:

    Create temporary evidence with owner-only permissions and repair restrictive permissions before resuming an existing work tree.

    line 100

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Meeting Transcription

Purpose

Turn long local meeting audio into a validated transcript body without treating a successful model response as proof of quality. This skill owns media preparation, transcription, outlier detection, retry/rescue, assembly, and usage evidence. Downstream business summaries, CRM changes, knowledge-base updates, and customer communication remain the consuming repository's responsibility.

Dependencies

  • ffmpeg and ffprobe
  • node, jq, and shasum
  • The separately managed gemini-files-api skill
  • Gemini authentication configured outside this skill

Install the wrapper dependency and this skill into the same reviewed manager root before running the canonical command:

npx --yes skills@1.5.14 add fiveonecode/agent-skills \
  --skill gemini-files-api --agent codex --global --yes
npx --yes skills@1.5.14 add fiveonecode/agent-skills \
  --skill meeting-transcription --agent codex --global --yes
bash ~/.agents/skills/gemini-files-api/scripts/bootstrap.sh

For Claude Code, use the same two pinned commands with --agent claude-code; the bootstrap path is ~/.claude/skills/gemini-files-api/scripts/bootstrap.sh.

The script resolves the Gemini wrapper from:

  1. GEMINI_MM, when set;
  2. ${HOME}/.agents/skills/gemini-files-api/scripts/gemini-mm.mjs;
  3. ${HOME}/.claude/skills/gemini-files-api/scripts/gemini-mm.mjs;
  4. ${HOME}/.codex/skills/gemini-files-api/scripts/gemini-mm.mjs for legacy installs.

The bootstrap script follows the same root order and can be overridden with GEMINI_MM_BOOTSTRAP.

The wrapper must support --model, --temperature, --max-output-tokens, --json, --prompt, and --file, and return JSON containing text plus usageMetadata. The orchestration script intentionally omits model-specific thinking controls because supported knobs and levels vary by Gemini model family; a caller-selected model uses its compatible default.

Canonical Command

Run the bundled script from the installed skill directory:

scripts/transcribe_meeting_audio.sh \
  --input /absolute/path/to/meeting.m4a \
  --work-dir /tmp/meeting-transcription-account-date

Defaults:

  • gemini-3.1-flash-lite primary model;
  • 300 second primary chunks;
  • 60 second retry chunks;
  • original spoken language preserved;
  • temperature 0 and minimal thinking;
  • no persistent Gemini file retention.

Use --prepare-only to verify source metadata, hashing, and chunking without API calls. Use --resume after an interrupted run; chunk sets are reused only when their completion marker matches the files on disk. Add --rescue-model <model> only when one-minute retries still fail validation.

Validation Contract

The primary response is suspect when it has:

  • empty text or invalid JSON;
  • any candidate finish reason other than STOP;
  • output tokens at the configured cap;
  • long generated runs of I, Yeah, or uh;
  • the same substantial line repeated at least five times.

Only suspect five-minute chunks are split into one-minute retries. The script exits non-zero before assembly while any retry remains unresolved. A stronger rescue model may process only those unresolved one-minute pieces.

Outputs

The work directory contains:

  • source-metadata.json and source.sha256;
  • chunks/, raw/, retry/, and optional rescue/ evidence;
  • validation.tsv, retry-validation.tsv, and unresolved-parts.txt;
  • accepted-rescue-parts.txt, which limits assembly to rescue outputs accepted by the current validation pass;
  • assembled-transcript-body.md with global chunk anchors;
  • usage.json with modality token counts;
  • run-manifest.json with source and output identity.

Pricing is intentionally not hardcoded. Compare usage.json with current first-party pricing when cost reporting is required.

Evidence Rules

  • Treat output as a rough AI transcript, not certified verbatim evidence.
  • Preserve original languages; do not silently translate.
  • Use generic speaker labels unless identity is grounded independently.
  • Prefer [unclear], [inaudible], or [overlapping speech] over invented content.
  • Keep raw audio, chunks, and API envelopes outside downstream repositories unless the user explicitly requests approved retention.
  • Create temporary evidence with owner-only permissions and repair restrictive permissions before resuming an existing work tree.
  • Reconcile names, numbers, dates, prices, and commitments against the source before downstream use.

Completion Gate

  • unresolved-parts.txt is empty.
  • validation.tsv and retry/rescue reports were reviewed.
  • assembled-transcript-body.md, usage.json, and run-manifest.json exist.
  • bash -n scripts/transcribe_meeting_audio.sh passes.
  • scripts/test_transcribe_meeting_audio.sh passes, including managed-root dependency discovery, interrupted-split recovery, and malformed-response retry coverage.
  • A short spoken fixture passes the normal path and an injected repetition/output-limit fixture reaches the retry path.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.