Resolve Rough Cut

Resolve Rough Cut

Turns a folder of raw footage into an assembled, gapless timeline. The timeline is the deliverable, not a finished video.

For editorial craft (why a cut works) see docs/guides/editorial-decision-guide.md. For the analysis layer see the resolve-media-analysis skill. This skill is the assembly recipe and the traps between them.

The deliverable contract

Deliver an assembled timeline. Do not add titles, captions, text cards, effects, transitions, music, or grading unless explicitly asked. Editors who ask for a rough cut want shot selection and pacing; styling is theirs. Adding graphics is work that gets thrown away — and see the Fusion trap below for why it also silently fails.

Render an mp4 only to preview the cut, and say that is what it is.

Workflow

1. Probe before importing. ffprobe every clip for r_frame_rate, avg_frame_rate, and rotation. Two things routinely differ from what the filenames and Resolve suggest — see Verified traps.

2. For a series, add a timeline to the existing project — don't create a new one. Check project_manager list first. One project per series keeps the bins, analysed clips and settings in one place; one per episode scatters them. Only create a project for a genuinely new series.

3. On a NEW project, set BOTH frame rates before any timeline exists. timelineFrameRate is locked once a timeline is created — set it via project_manager safe_set_project_settings. timelinePlaybackFrameRate cannot be set through the API at all, so ask the user to set it in the UI now, as a setup step:

   Project Settings (gear, bottom-right) → Master Settings → Playback frame rate

   Quote that path as written and do not extend it from memory — if the user cannot find it on their build, say you cannot see their screen and point at Blackmagic's manual rather than guessing a second location. Improvised UI directions are what issue #132 cost a user; see the retime entry in resolve-edit for the rule.

   Do not defer this to the end or write it off as cosmetic; a mismatch has been reported to affect playback and output, not just the display. On an existing project, read both back and only raise it if they are wrong.

4. Create a bin named for the episode, media_pool set_current_folder to it, THEN import. That order is mandatory, not stylistic: ImportMedia has no destination parameter and always lands in the current folder, so importing first puts the clips wherever the current folder happens to be — see the MediaPool.ImportMedia (current-folder destination only) entry in the api_truth ledger. Restore the previous current folder afterwards if it matters. One bin per episode, not one shared dump — it keeps analyze_bin scoped and the Media Pool readable across a long series.

5. Analyse — media_analysis analyze_bin, sampling_mode="adaptive_capped". Complete the commit_vision loop for every clip; leaving one in pending_host_vision_analysis is a failure, not a partial success (AGENTS.md).

6. Find shots via contact sheets, not one frame at a time: python3 scripts/contact_sheet.py <clip_dir> <out_dir> — see Shot finding.

7. Assemble in one call — media_pool create_timeline_from_clips with positioned clip_infos. Name the timeline for the episode.

8. Verify — timeline detect_gaps_overlaps must return zero of both, and check the total duration against the target.

Shot finding at scale

media_analysis extracts frames at full source resolution. A 35-minute 4K clip yields 80 frames at 2160x3840. Reading those individually to satisfy commit_vision costs ~1,100 tokens each — a four-clip shoot exceeds 250 frames.

Tile them instead. scripts/contact_sheet.py burns frame index, timestamp and selection_reason onto each tile, so per-frame findings stay reportable by index. Roughly a 15x saving with no loss of coverage. It needs Pillow (pip install Pillow) — the repo treats Pillow as optional elsewhere, but this script hard-fails without it rather than silently producing nothing.

Ask for short clips per subject rather than 20-35 minute takes when the shooter can choose. adaptive_capped tops out at 80 frames per clip, so a 35-minute take gets sampled only to about its first 16 minutes — the tail is never seen.

When footage arrives pre-culled (the shooter has already thrown away the unusable takes), scale the effort down: fewer frames per clip, fewer sheets, and trust the selection rather than hunting for the good moments. The expensive part of shot finding is separating usable from unusable, and that work is already done. Don't re-litigate it.

Verified traps

Each one silently produces a wrong result rather than an error.

Every row states the build it was confirmed on. A date alone is not reproducible — the scripting API changes per patch release, so a trap confirmed on one build is only a prior on another. Re-confirm before relying on a row whose build is older than yours, and update the stamp when you do.