# Type a multi-paragraph prompt caption layered behind the foreground subject

Manual DaVinci Resolve / Fusion construction guide generated from the recipe. It is not a tested .drp.
The accompanying Remotion project (recipe.json) is the executable definition; the numbers below are copied from it.
Neither construction establishes source fidelity: this is an original vector approximation of an inspected editing moment.

## Source and clock
Reference: OpenAI's "OpenAI deep research in practice." (zm6F0vo2E64), published 2025-02-04.
Source interval [1710, 1898) on the 24000/1001 fps master: 188 frames, 7.841167 s. Local frame 0 = source frame 1710.
Create a 1920x1080 progressive timeline at 24000/1001 fps and a Fusion composition exactly 188 frames long. Keep integer frames; do not round the frame rate.
Background field: #2d3a3c. Export silent H.264 yuv420p to match the intended playback container.

## Observation
Seven frames of a soft laptop close-up, then a hard cut (~f7) to a wide office shot with a seated figure at centre-left. Large white caption text is already present: two finished paragraphs at the top and a third paragraph mid-type. The third paragraph continues typing character by character (f7-170), and wherever a line crosses the seated figure the letters are hidden behind her, so words appear to pass behind the subject. The complete caption holds until a cut at f188.

## Mechanism
Depth-sorted caption: a type-on text layer is composited behind a matted foreground subject, so the typing reads as living inside the shot rather than on top of it.

## Phases (local frames)
- 0: Laptop close-up
- 7: Cut to wide shot; paragraphs one and two already set
- 8: Paragraph three types on behind the subject
- 170: Caption complete, hold

## Layer inventory (absolute 1920x1080 top-left pixels, back to front)
- closeup [group] at (0, 0) size 1920x1080 visible [0, 7)
  - cu-bg [rect] at (0, 0) size 1920x1080 fill #6f7f82
  - cu-screen [rect] at (80, 60) size 1200x720 radius 10 fill #c9d2d4
  - cu-keys [rect] at (0, 760) size 1920x320 fill #3a4144
  - cu-hair [ellipse] at (1200, -100) size 900x900 fill #5a3f33
- office [group] at (0, 0) size 1920x1080 visible [7, 188)
  - sky [rect] at (0, 0) size 1920x1080 fill #c7d0d6
  - city [rect] at (0, 620) size 1920x460 fill #a9b3b8
  - floor [rect] at (0, 880) size 1920x200 fill #5a5f5e
  - mullion1 [rect] at (600, 0) size 28x900 fill #3e4446
  - mullion2 [rect] at (1200, 0) size 28x900 fill #3e4446
  - frame-l [rect] at (0, 0) size 60x1080 fill #3e4446
  - frame-r [rect] at (1860, 0) size 60x1080 fill #3e4446
- caption [group] at (40, 60) size 1840x900 visible [7, 188)
  - para-a [text] at (40, 60) text "Please provide a list of sources that you referenced / and prepare your responses in a research report format /   / Use bullet points to make the text easier to read, or tables / to present data. Include the source for each datapoint" 30px weight 400 #ffffff
  - para-b [text] at (40, 336) text "Please review recent articles from industry publications, / consulting and analyst briefs, and publicly available / data, as well as reports from semiconductor associations." 30px weight 400 #ffffff reveal chars frames 8-170
- subject [group] at (560, 300) size 420x560 visible [7, 188)
  - s-table [rect] at (710, 600) size 330x18 radius 4 fill #7a5a44
  - s-leg [rect] at (860, 618) size 16x240 fill #5a4032
  - s-chair [rect] at (560, 630) size 140x230 radius 10 fill #8d6b52
  - s-body [rect] at (600, 440) size 180x300 radius 50 fill #8e2d78
  - s-head [ellipse] at (640, 340) size 120x130 fill #7b5a48
  - s-laptop [rect] at (770, 550) size 110x60 radius 4 fill #2b2b2b

## Construction steps
1. Timeline 1920x1080 at 23.976 fps; Fusion composition 188 frames. Local 0 = master frame 1710; the cut to the wide shot is at local 7.
2. Node order (back to front): office background plate, caption Text+ layers, then the foreground subject isolated with a Magic Mask or rotoscope matte, then the untouched frame edges.
3. Caption: two Text+ blocks, 30px white, 46px leading, left edge x=40; block A (paragraphs one and two) fully visible from frame 7; block B at y=336 with Write On 0 to 1 over frames 8-170 in character mode.
4. Track the subject matte to the footage so the occlusion follows her as she shifts; the caption itself does not move. The third paragraph's lines must cross the subject's head and shoulders for the effect to read.
5. Easy mistake: keying the text over the subject. The whole point is that the letters vanish behind her, so the matted subject must be merged on top of the text.
6. Hold the finished caption to frame 188; the exit is a hard cut to the results screen.

## Content variants
Default copy:
- paraA: Please provide a list of sources that you referenced / and prepare your responses in a research report format /   / Use bullet points to make the text easier to read, or tables / to present data. Include the source for each datapoint
- paraB: Please review recent articles from industry publications, / consulting and analyst briefs, and publicly available / data, as well as reports from semiconductor associations.
Adaptation changes:
- paraA: Summarize the current state of grid-scale battery storage / and present it as a briefing for a non-technical board /   / Use short sections with headings. Put every figure in a table / and cite the primary source next to each number
- paraB: Please prioritize regulator filings, utility annual reports, / and peer-reviewed studies from the last three years, / and flag any figure that comes from a vendor press release.

## Limits
Easing names map to Fusion spline shapes: linear = linear, in/out/inOut = ease curves on the same frames, hold = step, spring/overshoot = slight overshoot past the target before settling.
Text reveal frames give the first and last unit; distribute characters/words evenly between them. Cursor click rings last 14 frames.
Vector illustration is an approximation of layout and motion, not the product UI or recovered production assets.