Technique / ChapterReturn

Interrupt a task and return to its state

Mark a new part of the argument without losing the task.

Run the original

6.2 s · 30 fps · silent

Original neutral assets · 6.2 seconds · phase values are chosen defaults, not source timings.

npm run render -- ChapterReturn out/ChapterReturn.mp4 --codec=h264 --muted

Download the complete pinned project · Inputs, phase meanings & full source code

Official source / inspected interval

Workspace agents in ChatGPT: Third-party risk management agent

Watch the official OpenAI upload ↗ · Published 2026-04-24T13:00:56-07:00

Inspected 68–76 s. Consecutive subset: 69.5–70.75 s, 75 frames. Dates come from the official watch-page publication field.

Established two-pane task at source 68.000000 seconds — 68.000000 s · Established two-pane task

Chapter sentence still growing at source 70.183333 seconds — 70.183333 s · Chapter sentence still growing

Complete chapter holds at source 72.366667 seconds — 72.366667 s · Complete chapter holds

Return to the same task context at source 73.083333 seconds — 73.083333 s · Return to the same task context

Source: OpenAI. Bounded excerpt and annotated stills for the analysis below; not included in the project download.

What changes across the edit

Shot boundary: A white chapter card separates two views of the same agent setup. Cut onsets outside the native subset are only sample-bracketed.
Framing: Two-pane application → centered single-line chapter → two-pane application.
Layer ownership: Chapter replaces the entire product composition. No background task is implied to complete behind it.
Motion: Letter construction on the title, then product return and later composer input.
Typography: One centered heavy sentence grows; complete text is visible by the 70.917 s sample.
Action order: Setup summary → test chapter → same setup context → next request.
Pacing & readable holds: 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.

Recreation versus observation

Matches

A title interruption and return to an identical original task snapshot.

Differs

Exact pixel identity across our return is an implementation invariant, not a measured source guarantee. Copy, timing and geometry differ.

Measured

Native inspection 69.500–70.750 s; broader interval 68–76 s.

Inferred

Task continuity comes from the same agent setup, not from a claimed unchanged backend state.

Chosen defaults

All Remotion phase frames, cubic Bézier (.16,1,.3,1), geometry, colors, font choice and reading holds are chosen defaults. None is a recovered OpenAI specification.

{
  "title": "Make room for the idea.",
  "prompt": "Arrange a small exhibition.",
  "result": "Three studies, ready to compare.",
  "label": "EXHIBITION / 03",
  "accent": "#40685b",
  "action": 30,
  "change": 66,
  "settle": 126,
  "hold": 60,
  "focusX": 0.66,
  "focusY": 0.64,
  "zoom": 1.55
}

Increase hold to extend only the final quiet state. Change the phase frames to change choreography; preserve their ordering. Content limits and the exact meaning of each phase are in the handoff.

Exact original geometry, layer ownership and asset contracts · Schedule against an actual recording