Add action mode (audio-transient detection), in-frame verification, and time-warp helper

[Erik5000]

Summary

Three additions, each motivated by a real session where the existing dialogue-only flow had gaps. The original skill is great for podcasts/interviews; these changes extend it without changing the dialogue path.

1. Action mode — Step 1B + scripts/detect_transients.py

For footage driven by repeated impacts (woodchopping, sports, drumming, percussion, throws/hits) where Whisper isn't useful. The new script detects strikes via spectral flux on a tunable freq band:

• defaults (--band 1000:6000 --min-flux 30 --min-gap 0.6) work out-of-the-box for axe-on-wood
• band can be retuned for heavy thuds (200:2000 — kick drum, body shots) or high cracks (4000:10000 — snare, whistles, clinks)
• ambient/wind sits at flux 5–15; real impacts at 50–300, so the threshold is robust

In testing on ~30 minutes of woodchopping footage, the detector found 170 in-frame strikes across 5 phone clips with one CLI invocation per clip.

2. In-frame verification — Step 1.5

Applies to both modes. Audio detection finds the sound of an event; the camera might not have captured it. The new step samples one frame at each candidate via ffmpeg -ss T -frames:v 1 and the agent reads each verify image before rendering. Drops:

• candidates where subject is bent/off-frame/behind obstacle
• thumb-on-lens shots (yes, this happens — only catchable visually)
• false-positive sounds (door slam, dropped tool)

This caught a clip where the chopper was bent low over the log on every chop in that clip — would otherwise have produced an awkward "we cut on a sound I can't see" feeling in the final edit.

3. Time-warp helper — Step 4c

For when the user wants a slow scenic/pan/reveal shot tightened ("cut this in half") without losing the arc. Trimming forces dropping part of the arc; speed-ramping with setpts/atempo preserves all beats at higher tempo:

ffmpeg -ss \"\$START\" -i \"\$CLIP\" -t \"\$ORIG_DUR\" \\
  -vf \"setpts=PTS/2,scale=1080:1920:flags=lanczos,fps=30\" \\
  -af \"atempo=2.0\" \\
  ...

Includes rules of thumb (1.5x for brisk human movement, 2x for punchier reveals, chained atempo=2.0,atempo=2.0 for 4x since a single atempo filter only accepts 0.5–2.0).

Other changes

• Frontmatter description extended to include action-mode triggers ("action montage," "cut on each chop/hit/punch")
• New ## Modes section at the top routes the agent to dialogue / action / hybrid before Step 1
• Step 5 (subtitles) now starts with "Skip in action mode"
• Pitfalls updated with band-tuning guidance and the atempo chaining gotcha

Test plan

• detect_transients.py --help works; runs cleanly on 16kHz mono WAV
• Detector smoke-tested on real footage (5 phone clips, woodchopping); top-N flag returns the same hits a human would pick
• Time-warp ffmpeg pattern verified on a 12s pan → 6s output with synced audio
• Reviewer to confirm the new sections fit the existing prose style