Recreating Deck From Recording
Given a recording, produce a slide-accurate PowerPoint:
- Scan the full recording, identify every timestamp range where slides are visible.
- Build a slide inventory (one entry per distinct slide, with timestamps).
- Pick the clearest, most complete frame for each slide.
- Deduplicate near-identical captures.
- Build the PPTX — screenshot-based or editable text, never invented content.
- Add source timestamp + confidence to speaker notes.
- Deliver the file plus a completion report.
Never fabricate text, numbers, or layout. If something can't be read, mark it [Unreadable in recording] rather than guessing.
Progress:
- Step 1: Inspect recording, identify slide-visible segments
- Step 2: Build slide inventory (timestamps, titles, quality flags)
- Step 3: Select best frame per slide
- Step 4: Remove duplicate captures
- Step 5: Reconstruct PPTX (image fill or editable text/shapes)
- Step 6: Handle unreadable/obstructed content
- Step 7: Add traceability notes (timestamp, status, method, confidence)
- Step 8: QA pass against checklist
- Step 9: Deliver file + completion report
Step 1 — Inspect the recording
Ignore: camera-only segments, intros, loading screens, app-switching, notifications, meeting controls, chat panels, overlays. Detect slide changes via visual diffs, titles, layout shifts, page numbers.
Step 2 — Slide inventory
Per slide, record: sequence number, first/last appearance timestamp, visible title, full-visibility flag, animation-reveal flag, image quality, and uncertainty notes.
Animation handling:
- Progressive reveal of the same final slide → capture only the final complete state.
- Animation stages carrying materially different information → preserve as separate, clearly labeled "animation stage" slides.
- Cursor movement, presenter camera, or control-bar changes alone are never grounds for a new slide.
Step 3 — Best frame selection
For each slide, scan multiple frames in its interval and choose the one where: content is fully visible, text is sharpest, no motion/transition is in progress, cursor/camera/controls don't obscure content. Prefer a later stable frame for gradually-revealed content. Crop precisely to slide boundaries; correct perspective only if needed; never stretch or distort.
Step 4 — Deduplication
Same title + layout + text + visuals + page number + final animation state = duplicate → keep clearest only. Do NOT merge slides that differ in figures, dates, labels, highlights, or animation content even if visually similar.
Step 5 — Reconstruction priority
- Clear source → full-slide cropped screenshot image.
- Readable text → editable PowerPoint text boxes with exact original wording.
- Simple shapes/tables/diagrams → editable objects, only if accurately reproducible.
- Never invent missing content, alter numbers/names/dates/labels/conclusions/chart values.
- Unrecoverable slide → clearly labeled placeholder, not a guess.
For screenshot slides: image fills the slide, full original preserved, no decorative overlays, no unnecessary margins.
Step 6 — Unreadable/obstructed content
Search other moments in the recording for a clearer view before giving up. Use only visible information. Mark gaps as [Unreadable in recording], note the timestamp and cause in speaker notes.
Step 7 — Traceability (speaker notes only, not on-slide)
- Source timestamp
- Capture status: Complete / Partial / Unclear
- Reconstruction method: Screenshot / Editable reconstruction
- Obstruction/unreadable details (if any)
- Confidence: High / Medium / Low
Step 8 — QA checklist
- Full recording reviewed start to finish
- All distinct slides considered
- Slide order matches recording
- Duplicates removed
- Images sharp and correctly cropped
- No mid-transition/incomplete-animation frames used
- No notifications/chat/controls/overlays unless part of source slide
- Wording, figures, dates, labels preserved exactly
- No invented content
- Consistent 16:9 (or source-native) layout
Example 1: Input: 40-minute screen-recorded webinar, slides visible 00:02:15–00:35:40, presenter camera in corner throughout. Output: 18-slide PPTX; 14 slides as screenshots (complex charts), 4 as editable text reconstructions (title/agenda/summary/closing slides); 3 duplicate frames removed (re-shown during Q&A); speaker notes on each slide with timestamp and confidence; completion report listing one slide marked Partial due to presenter camera covering a bottom-right legend from 00:18:02–00:18:40 (clearer frame found at 00:18:55 used instead).
Example 2:
Input: Recording where a budget table's last column is cut off-screen for its entire visible duration (00:12:00–00:14:30).
Output: Slide recreated with all readable content; final column cells marked [Unreadable in recording]; speaker notes cite timestamp range and state "right edge of table out of frame for full duration, no clearer alternate frame found"; flagged in completion report as requiring manual review.
- Treat the recording as sole source of truth — no external images, stock content, or inferred facts unless explicitly requested.
- Default to 16:9, original language, and original sequence unless told otherwise.
- Prefer editable text when it can be reproduced with certainty; fall back to image when fidelity is at risk.
- Always scan forward/backward in time before declaring content unreadable.
- Keep traceability data in speaker notes only — never clutter the visible slide.
- Treating cursor movement, camera overlay, or control-bar changes as new slides.
- Capturing a frame mid-transition or mid-animation instead of the stable final state.
- Deduplicating slides that differ only in data (dates, figures, highlights) — these are distinct, not duplicates.
- Guessing obscured text instead of marking it
[Unreadable in recording]. - Including chat panels, notifications, or meeting-control UI that weren't part of the actual slide.
- Stretching/distorting screenshots to fit a fixed slide size instead of preserving aspect ratio.