Repurpose a long video into viral shorts

This is the autoshorts pipeline. It chains: transcribe → identify clips → cut → optional overlay → upload to short-form platforms.

1. Locate the source video

The FFmpeg tools need a public, directly downloadable HTTPS URL (or an Upload-Post media URL from a previous upload). A YouTube/Vimeo watch page is not a file URL. If the user only has a local file, ask them to share a direct download link first.

2. Transcribe

The Upload-Post MCP does not ship a transcription tool today. Two options:

• The user already has a transcript (YouTube captions, a Whisper export, an existing autoshorts run). Ask for it and skip to step 3.
• No transcript yet. Submit a submit_ffmpeg_job with input_url and a full_command that extracts the audio (e.g. ffmpeg -y -i {input} -vn -ac 1 -b:a 64k {output} with output_filename: "audio.mp3"), then run Whisper locally (or any transcription tool the user has). Do not promise transcription as part of the plugin — be honest that the user needs to bring it.

Save segment-level timestamps. You will need them to cut.

3. Identify clip candidates

Send the transcript (with timestamps) to the model and ask for viral moments. Good criteria:

• A hook in the first 3 seconds.
• Self-contained — works without the preceding context.
• 20–60 seconds long.
• Has a strong line worth a hook overlay.

Aim for 3–8 candidates. Present them to the user with timestamps and one-line summaries; let them approve before cutting.

4. Cut with FFmpeg

For each approved candidate, call submit_ffmpeg_job:

• input_url — public URL of the source video (or an Upload-Post media URL from a previous upload). Use files (array of URLs) instead only when the job needs several inputs.
• full_command — the explicit FFmpeg command. It must start with ffmpeg and use the placeholders {input} (or {input0}, {input1}… with files) and {output} instead of real filenames. Shell metacharacters (;, |, &, $, backticks) are rejected by the API. Example trim: ffmpeg -y -ss 754 -to 801 -i {input} -c:v libx264 -c:a aac {output}.
• output_filename — optional, e.g. clip-1.mp4; otherwise auto-named.

The call returns a job_id. Poll get_ffmpeg_job (jobId) until the status is completed, then download_ffmpeg_result (jobId) to get the download URL.

If the source is horizontal, do the trim and the 9:16 reframe in the same command rather than chaining jobs, e.g. add -vf "crop=ih*9/16:ih,scale=1080:1920". Every job counts against the FFmpeg quota, so fewer passes is better.

5. Hook overlay (optional)

If the candidate has a strong hook line, generate a short overlay (large bold text, first 2–3 seconds) and burn it in with a drawtext filter (e.g. drawtext=text='…':fontsize=72:x=(w-tw)/2:y=h*0.15:enable='lt(t,3)') — ideally in the same full_command as the cut. Skip this if the user is in a hurry.

6. Schedule

Hand the resulting clip URLs to the schedule-campaign skill (or call upload_video directly with platforms: ["tiktok","instagram","youtube"]). Stagger them — do not post 5 clips at once to the same account; use addToQueue: true to let Upload-Post space them.

Limits and quota

FFmpeg processing is metered in minutes per month against the plan allowance. Call get_ffmpeg_consumption upfront to know how much budget you have. If submit_ffmpeg_job returns a quota error, surface the upgrade URL rather than retrying.