Update to Tella

{
  "description": "Find and remove recording mistakes (false starts, retakes, repeated phrases, abandoned sentences, technical interruptions) from a Tella video's transcript through the Tella MCP. Use when asked to remove mistakes, clean up retakes, or tighten a recording by cutting flubbed takes.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 377
    },
    {
      "relative_path": "references/mistake-rules.md",
      "size_in_bytes": 6656
    }
  ],
  "name": "tella-remove-mistakes",
  "skill_md_contents": "---\nname: tella-remove-mistakes\ndescription: Find and remove recording mistakes (false starts, retakes, repeated phrases, abandoned sentences, technical interruptions) from a Tella video's transcript through the Tella MCP. Use when asked to remove mistakes, clean up retakes, or tighten a recording by cutting flubbed takes.\n---\n\n# Remove mistakes from a Tella video\n\nRead the transcript, find the mistakes using the rules in [references/mistake-rules.md](references/mistake-rules.md), show them to the user, and cut the approved ones with `cut_clip_by_transcript`.\n\nThis skill needs the Tella MCP. If tools such as `get_transcript` are missing, follow the connection steps in the `tella` skill first.\n\n## Scope\n\nMistakes are not filler words or silences. For standalone \"um\"/\"uh\", use `remove_fillers`; for pauses, use `get_silences` / `remove_silences`. Only do those if the user asked for them too, and on each clip send them only after its `cut_clip_by_transcript` call has returned: calls on the same clip that run together each return the clip as they saw it, so one result can be missing the other's cuts.\n\nFinding mistakes is read-only. Do not cut anything until the user has approved the list, unless they explicitly asked you to remove the mistakes without reviewing.\n\nOnce the user has answered, cut only what they approved. A mistake you spot after that (while re-reading the seams, or in a later turn about something else) is a new finding, even when it looks like the rest of an approved one: list it and ask, don't cut it.\n\n## Workflow\n\n1. **Find the clips.** `get_timeline` lists the video's clips. Work clip by clip: transcripts, word indices and cuts are all per clip. Before cutting, call `get_clip` on each one and keep its `cuts`, so you can restore it.\n2. **Read the transcript.** Read every word of the clip with `get_transcript` `view: \"text\"`: one sentence per line with its start time, each word prefixed with its stable `index` (`12:word`), on the clip's playback timeline (existing cuts already removed). Several rules compare neighbouring sentences.\n   - Read it in windows of about 5 minutes with `startTimeMs`/`endTimeMs`, one window per step, and start each window after the first about 30 seconds before the previous one ended, so retakes that span a window edge are still visible. Finish a window (step 3 below: write down its mistakes as word ranges) before you request the next one. A mistake found twice in the overlap, with similar bounds, is one mistake.\n   - Don't look for mistakes in the sentences view, which hides the words inside each sentence: retakes and repeated sentences get missed that way. Read the words view (`get_transcript` without `view`) only for a stretch where you need exact word times or want to check a doubtful finding.\n3. **Spot the mistakes.** Apply [references/mistake-rules.md](references/mistake-rules.md). Each finding is an inclusive word-index range, a short reason (for example \"False start\" or \"Repeated phrase\"), and the words it removes. When unsure, leave it in: over-trimming is worse than under-trimming.\n4. **Show the list before cutting.** For each finding give the time, the reason, the words that would be cut, and a few words of context on each side so the user can see what the sentence becomes. Let the user drop or adjust items. Report clips where nothing was found too.\n5. **Cut.** Call `cut_clip_by_transcript` once per clip with every approved `{fromWordIndex, toWordIndex}` range. Word indices do not shift when cuts change, so the ranges stay valid. Never pad ranges with extra words \"for safety\", and never split a word.\n6. **Check the seams.** Re-read the transcript around each cut. Look at the gap between the last word before the cut and the first word after it. A word that ends right at the cut, especially one ending in a consonant, can sound clipped, and mistake cuts are where that happens, so listen to the seams once: download the MP4 that `get_clip_preview` returns (it has the clip's audio), cut each seam's range to audio with ffmpeg (`ffmpeg -ss 2 -t 6 -i preview.mp4 -vn -ac 1 -ar 22050 -b:a 48k seam.mp3`), join them into one file and listen to it with your audio tool (`describe_audio` in the Tella editor). If the listen and the transcript disagree, `get_audio_levels` for those few seconds settles whether the word is whole. If a seam sounds wrong, move the cut by a whole word rather than adjusting milliseconds. Confirm the sentence now reads as intended.\n7. **Report** what was cut per clip, what you left in on purpose, and what you verified (transcript re-read, seams listened to). If you didn't listen, say so in a few words.\n\nA cut call that returned the clip succeeded; don't send it again, even when a later result seems to lack its cuts. If a cut call fails or its outcome is unclear, re-read the transcript before retrying: words that were cut are absent from it. When `cut_clip_by_transcript` reports word indices not found, those words are already cut: drop those ranges and send the rest once. Never blindly repeat a cut.\n\nTo undo, call `update_clip` with the `cuts` you saved before cutting, then re-apply the cuts you want to keep. Prefer that to guessing at a partial undo. `cuts` replaces the clip's whole cut list, and the person may be editing the video by hand while you work: keep any cut in the current `get_clip` that neither you made nor was in the saved list, and send `cuts: []` only when that is the result: the saved list was empty and the current clip has no cuts to keep, or they asked to remove every cut.\n"
}