← Recipe index / Structure
R02_chapter_reset / 180 frames / 30 fps
Chapter reset → return to context
Separate two arguments without losing the task. Use between related demonstrations, not in the middle of an uncompleted cause-and-effect beat.
Source evidence
Workspace agents in ChatGPT: Third-party risk management agent ↗ · official OpenAI channel · published 2026-04-24
Inspected source window 68–76 s. Consecutive subset 69.5–70.75 s: 75 frames. A longer excerpt is not an every-frame edit list.
Source evidence, not a Resolve recreation. Silent excerpt; the player counter starts at zero.
Observed mechanic and limits
A white chapter card separates two views of the same agent setup. Cut onsets outside the native subset are only sample-bracketed.
Two-pane application → centered single-line chapter → two-pane application.
Letter construction on the title, then product return and later composer input.
Setup summary → test chapter → same setup context → next request.
Complete chapter is present at 70.917, 71.633 and 72.367 s; the app has returned by 73.083 s. These bound, not exactly measure, the hold.
Established two-pane task
Chapter sentence still growing
Complete chapter holds
Return to the same task context
Construct in Resolve
Documentation-checked; application-unverified. Use the environment, frame-counting, geometry and font setup first. Timeline timecode starts at 01:00:00:00. Tables use zero-based recipe-local frames and half-open intervals. Clip-local frame equals recipe frame minus that clip’s In; add any Fusion composition offset separately. Values are chosen implementation defaults.
| Track | Asset / native tool | In (inclusive) | Out (exclusive) |
|---|---|---|---|
| V1 | context.png | 0 | 45 |
| V1 | paper.png | 45 | 120 |
| V1 | context.png (duplicate) | 120 | 180 |
| V2 | Text title | 45 | 120 |
- Build three adjacent V1 clips with durations 45, 75 and 60 frames. Use the same Media Pool context asset for both outer clips; do not independently rescale the return.
- Add a basic Text title over paper.png only: “Now inspect the result”, DejaVu Sans Bold, Size 72, Alignment Center, Anchor horizontal Center / vertical Center, Title Position X 960 / Y 540, color #23312e. Settings → Transform Position 0,0 and Zoom 1. No animation or transitions.
- Compare the last pre-card frame (44) with the first return frame (120). The fixture should be identical. For real captures, preserve the visible task state, not necessarily every pixel.
Inspector and keyframes
| Control | Recipe-local frames | Value and interpolation |
|---|---|---|
| All context clips | 0–44 and 120–179 | Zoom 1; Position 0,0; no keyframes |
| Title/card | 45–119 | 75 frames; full-opacity title for entire card |
Hold, handles and retiming
Chapter title has 75 complete frames (2.5 s); returned context gets 60 frames (2 s).
Stills/generators need no temporal handles. For real captures reserve 15 source-time frames before and after each used interval at a 30 fps capture rate; keep those handles outside the edit. No dissolve or optical-flow handles are needed here.
100% speed. No Optical Flow, Speed Warp, frame interpolation or simulated product wait.
Replace assets / change duration
Lengthen the middle card by ripple-moving the return and every downstream marker. Replace both context instances with the same asset/state. Do not manufacture a completed backend action across the title.
Compare the mechanic, not the branding
Exact pixel equality on return is a fixture contract; source task identity, not identical backend state, is evidenced. Source title letter construction is not recreated.
No application render or editable project has been produced. Timing, framing, motion, alpha, continuity and readability comparisons remain incomplete. A source clip or fixture PNG is not that verification.
Construction reference: official Resolve 21 manual, p. 967, p. 1212, p. 1219, p. 1246. Controls still need confirmation in the installed edition/version.
All recipe specifications · Original preparation kit · Import/export and verification procedure