Method, limits & data

100 distinct source films. Specific examples, not a claim to have reverse-engineered OpenAI's production system.

This public edition contains 100 official OpenAI videos, 1,075 timecoded shot groups, 637 app/browser occurrences, and 37 close motion studies. It expands an earlier 36-film collection with 64 additional films. The study date is September 4, 2026; the library is not a complete channel archive or a list of the latest products.

What was inspected

• All 100 films have source metadata, a full-duration visual sampling pass, an authored shot-group timeline, representative stills, framing observations, adaptations, and limits.
• The first 36 retain their earlier two-second overview samples and selected closer inspections. The added 64 use four-second full-duration contact sheets. Short shots between samples can be missed.
• Close motion studies use the sampling interval printed on each card. The 16 added in this edition were inspected at half-second intervals. A statement such as “visible by 31.0 seconds” refers to a sample, not an exact cut boundary.
• The source metadata was checked against the official OpenAI channel identifier, and source file hashes were checked against the local analysis records. The public source index includes the identifiers and hashes, not downloader logs or media-server credentials.
• This is visual analysis. No listening pass, soundtrack analysis, frame-exact edit reconstruction, independent performance benchmark, or current feature-availability audit is claimed.

What the counts mean

A film is one distinct official source video, not another excerpt from the same film. A shot group is an editorial beat that may contain several cuts. Groups cover the source timeline, but their boundaries are approximate; this is not an edit decision list.

An app/browser occurrence is a group with identifiable interface content: a complete screen, a staged or cropped composer, a menu, an embedded artifact viewer, or a readable filmed-device insert. A group may also contain a presenter cutaway. Occurrences are not unique features or unique UI components.

An isolated generated image, an output reel, a title, an advertising chat-text overlay, or an unreadable background laptop is not counted as an app-screen example. Films with zero screen examples remain useful references for human storytelling, output presentation, or typography. Filter the shot list by type to see those differences.

A motion study is a short excerpt with its own narrow observation and adaptation. It is not another source film and is not added to the shot-group total.

Observation is not a production preset

The guide separates three things:

1. Observed evidence: a credited frame or excerpt and a description of what is visible at the sampled source time.
2. Editorial interpretation: why that framing or sequence may help an audience understand the task. The techniques page compares examples; its conclusions are this study's interpretations.
3. Proposed adaptation: a way to use the idea in another production while retaining the real product's behavior and that presentation's identity.

No inferred duration, crop, font size, animation curve, or UI token is presented as an official OpenAI film specification. Automatic scene-change candidates are not treated as a confirmed cut list. Visible processing labels and quick edits do not establish latency.

How to use the library

Start with the cross-film techniques, then choose a film format. Use the screen index for an app moment and the motion index to inspect a transition. The shot list gives the full narrative context around each example.

Each still can be enlarged and has a link to the corresponding official source moment. Film pages include a “Do not infer” note to distinguish an attractive example from a verified product claim. Filters and pagination keep large indexes readable; with JavaScript disabled, the underlying entries remain available.

Applying this to a paired presentation

These are proposed workflow safeguards, not observations about how the source films were produced.

• Treat the supplied talking-head recording—including its actual speech and pauses—as the timing master. Derive the transcript timing from the recording, not script length.
• Keep the talking-head video and generated presentation video as two separate assets. The presentation should not duplicate the spoken narration or add music by default.
• Start the presentation with a title or logo. A recognizable content transition can cue a manual switch to the OBS playback scene. Record both cue time and the expected human-response offset; do not assume simultaneous playback.
• Compare both assets on one shared timeline. If the talking head begins at presentation time offset, its shared ending is offset + talking_head_duration. The presentation ending must precede that ending by the agreed buffer. An opening lead-in and an ending buffer are separate requirements.
• Confirm that the actual supplied footage contains the ending wave or pause. Do not invent or alter the speaker's performance to fill a missing buffer.
• Plan one consolidated storyboard and demo review. Capture genuine product behavior using approved demo data; do not recreate source-film capabilities that the target product lacks.
• Set final aspect ratio, resolution, frame rate, codec, and playback constraints against the supplied assets and meeting setup. The study clips' 960-pixel width and 30 fps encoding are only browsing choices.

Source footage and publication boundary

This is independent research, not an official OpenAI publication or an endorsement. Source footage, brand marks, and example data belong to their respective owners. The stills and brief silent excerpts are presented with specific visual commentary and official source links. Their presence here does not grant permission to reuse them in a new presentation.

Full source movies are not hosted here. This standalone edition also excludes private component catalogs, internal product source code, private-site files, local machine details, and raw downloader metadata. For a new production, borrow the editorial idea and create original, faithful footage of the real product.

Download the research

• Complete study data — JSON: all films, methods, shot groups, observations, and adaptations.
• 100 official sources — JSON: source titles, channel identifiers, dates, links, durations, and reference hashes.
• Timecoded shot list — CSV: all 1,075 editorial groups.
• App/browser inventory — CSV: the 637 interface occurrences only.
• Close motion studies — JSON: excerpt bounds, observations, and linked shot groups.
• Publication manifest — JSON: the built public files and checksums.

The data uses seconds from the start of each source film. A source URL is provenance, not a claim that the historical interface or advertised service is available unchanged today.