trx -- Agent-First Transcription CLI

npx skills add crafter-station/trx -g

Prerequisites

trx doctor --output json

trx init

bun add -g @crafter/trx

Workflow

1. Dry-run first (always)

trx transcribe <input> --dry-run --output json

2. Transcribe

trx transcribe "https://youtube.com/watch?v=..." --output json

trx transcribe ./recording.mp4 --output json

trx transcribe <input> --fields text --output json

3. Post-process (fix whisper mistakes)

.txt

1. Punctuation: Whisper drops periods at paragraph boundaries and misplaces commas. Fix sentence boundaries.
2. Accents (Spanish): Whisper often drops diacritics. Restore: como -> como/cmo, esta -> esta/est, mas -> mas/ms.
3. Technical terms: Whisper misspells domain-specific words. Ask user for a glossary or infer from context.
4. Repeated phrases: Whisper sometimes stutters on word boundaries. Remove exact consecutive duplicates.
5. Speaker attribution: If user provides speaker names, insert

   [Speaker Name]:

   markers.
6. Filler words: Remove "um", "uh", "este", "o sea" if user wants clean output.
7. Timestamp alignment: If editing

   .srt

   , preserve the timestamp structure. Only modify text between timestamps.

4. Schema introspection

trx schema transcribe
trx schema init

Commands

init

trx init --model small

transcribe

trx transcribe <url-or-file> --output json

doctor

trx doctor --output json

schema

trx schema transcribe

Shorthand

trx <input>

trx transcribe <input>

Output format

• --output json

  : Machine-readable (default when piped)

• --output table

  : Human-readable with progress (default when TTY)

• --fields text

  : Only return transcript text (saves tokens)

• --fields metadata

  : Only return metadata (language, model)

• --dry-run

  : Validate without executing

Flags reference

--language <code>

auto

--model <size>

--output-dir <dir>

.

--no-download

--no-clean

--json <payload>

Edge cases

• yt-dlp extension mismatch: yt-dlp sometimes outputs

  .mp4.webm

  instead of

  .mp4

  . The CLI handles this by scanning for the downloaded file by prefix.
• Large files (>1hr): Whisper processes in segments. Works but is slow on CPU. Consider

  --model tiny

  for speed.
• No GPU: whisper-cli uses CPU by default. Acceptable for tiny/base/small models.
• Auto-detect language: When

  --language auto

  , Whisper detects the language from the first 30 seconds. For multilingual content, specify the primary language.