I turn AI’s promise into proof.

A description of the picture and on-screen text, with approximate times. Product details and offers are shown as they appeared in the 2026 films.

1. 0:00

   An AI chat opens on “Legible Film, an Agent Skill for Claude Code and Codex” and the line “Your coding agent can make and edit videos for you.” A request is typed: “Make a launch film for Atlasfold, my offline photo-map app.”

2. 0:06

   The agent’s film plays, labelled “Made by the agent · Atlasfold is an invented app”: photos gather into a map, then title cards for a searchable map that runs offline, and “Atlasfold, out today”.

3. 0:16

   “Turn an idea into a launch film.” A second request asks for a scene to be cut into a short captioned clip.

4. 0:23

   “Change the format of your videos without losing focus.” A 21-second scene from Tears of Steel becomes a 6-second captioned 4:5 clip that keeps the speaker’s face in frame.

5. 0:30

   “Clean up your recordings without losing a word.” In the paper edit of a reading, a 1.49-second gap is cut, then put back by editing one line of text. Tightening is labelled experimental.

6. 0:37

   A third request asks for a 25-to-30-second trailer of a 7½-minute animated short. “Turn a film you love into a trailer.” The agent maps 459 words and 76 shots, gives each cut a reason, and cuts around a shot that flickers about 12 times a second.

7. 0:54

   “Time your sound to the frame.” A sound peaks 165 ms before she speaks, and her voice stays at least 12 dB above it.

8. 1:00

   “Change your cut by saying what you want.” An example review note becomes an exact edit, 2.58 seconds shorter with one line removed, and every version is kept.

9. 1:07

   “Hand off to your editor, and check what comes back.” In the editor’s returned cut, it finds a caption 1.6 seconds late and proposes a fix, marked “proposed, not applied”. The XML hand-off is labelled experimental.

10. 1:14

    “Know what’s left to check before you post.” The trailer’s delivered file: 15 checks pass, 4 to look at, and 3 need a person, who watches it muted, hears it with sound and checks names and numbers in its captions.

11. 1:21

    Short examples: words on screen as they’re spoken, captions timed to every word, shorts from one long video, titles behind a moving subject, and a note pinned to one frame of a cut, labelled as an example.

12. 1:34

    “And more.” It also scores cuts, converts files and checks your setup.

13. 1:39

    “This film was made with Claude Code and Legible Film.” The end card gives the repository, github.com/JamesDimachkie/legible-film, the install command and the credits.

A description of the picture and on-screen text, with approximate times. Product details and offers are shown as they appeared in the 2026 films.

1. 0:00

   A video player cycles through sample subjects before landing on a kung fu course. A speed menu offers ‘skip watching’ and the player advances to the end.

2. 0:06

   A Claude Code prompt runs /skipwatch with a video link. The film presents a 48-minute video becoming a skill in one command; the kung fu timeline is labelled ‘Illustrative run.’

3. 0:12

   A white dithered figure bows, takes a horse stance, punches, then blocks and counters. Some steps carry source-time tags. A quoted example and a SKILL.md panel show how the steps become a reusable skill.

4. 0:30

   A prompt invokes Skill(kung-fu-basics). The figure responds to another with the displayed answer: ‘Block. Step in. Counter.’

5. 0:36

   The SkipWatch mark appears as examples cycle through kung fu, sourdough, TypeScript, linear algebra, cinematography and quantum physics. The screen labels these illustrative examples.

6. 0:44

   The closing card shows installation and usage commands, github.com/robert-botto/skipwatch, and ‘Free and open source (MIT).’

A description of the picture and on-screen text, with approximate times. Product details and offers are shown as they appeared in the 2026 films.

1. 0:00

   An Outlook Send button gives way to headlines about real email disclosures. The film identifies the quoted voices as synthetic.

2. 0:23

   Emails rebuilt with invented names show a party invitation with 2,743 people's details attached, a clinic newsletter exposing 781 addresses, 245 applicants visible to one another, bank accounts sent outside, payroll sent to a scammer, and a password sent to an impostor.

3. 0:58

   The password email is shown again with Guardcell. On Send, a dialog identifies the password and offers Send anyway, Review or Don't send. The review pane highlights the issue; the sender removes the sentence and sends again.

4. 1:10

   The film illustrates checks of recipients, message text and attachments with examples including passwords, card numbers, lists, margins and client remarks. A laptop diagram says the checks run without a server.

5. 1:29

   A grid of rebuilt emails accompanies the film's result: Guardcell asks about 35 of 43 documented incidents that went outside. The case study explains that these replays informed development, rather than proving historical incidents would have been prevented.

6. 1:37

   An administrator uploads a manifest and assigns users. The closing card calls Guardcell free and open source under Apache 2.0 and points to its GitHub repository.

A description of the picture and on-screen text, with approximate times. Product details and offers are shown as they appeared in the 2026 films.

1. 0:00

   A green line travels across a paper-like scene as Decision Models is introduced as inference for open-weights decision models. Cards representing new models accumulate around the question: ‘Which one do you build on?’

2. 0:12

   A JevBench scoreboard credits founder Florian Standhartinger and highlights the open-weights models among the leading entries. The film presents five models through one Jev-compatible API, with separate service terms.

3. 0:24

   An illustrative support request asks to update an invoice address. A decision model selects billing with a displayed 0.87 score; the film says the probabilities are illustrative, not measured API results.

4. 0:32

   A dated, company-measured comparison shows 5.28 versus 31.73 decisions per second for Perplexity Decider v1.1, using several questions per request on the same RTX PRO 6000 GPU. The film also notes a smaller gain for single-question requests and that some answers differ.

5. 0:38

   An illustrative cost counter reaches $6.67 for one million S1-FAST decisions on the peer-to-peer tier, using 351 input tokens per decision, output free and excluding tax. The film compares that 2026 list-price example with Jev.

6. 0:44

   The film points to public REAX Engine code, then shows a globe and an invitation to become a hosting provider, with hardware and registration requirements. It closes with the Decision Models mark and the line ‘Served by the people, for the people.’

A description of the picture and on-screen text, with approximate times. Product details and offers are shown as they appeared in the 2026 films.

1. 0:00

   A compact grid of points spreads into a network as ‘One data center’ changes to ‘Served by the people.’ The heading identifies REAX Network and Bittensor subnet 92.

2. 0:05

   An illustrative decision card asks whether to escalate a ticket. A line travels through the network to a GPU run by an independent miner; the example displays 0.91 for escalate and 0.09 for keep.

3. 0:10

   A validator appears. The film says miners are checked on fresh questions and notes that network processing is opt-in on the peer-to-peer tier, company GPUs provide fallback, and answers are checked on a sampled basis.

4. 0:16

   The REAX mark closes the film with ‘AI by the people, for the people’ and decisionmodels.io.

A description of the picture and on-screen text, with approximate times. Product details and offers are shown as they appeared in the 2026 films.

1. 0:00

   REAX Competitions asks whether a 9B model can take first place. A public development risk-assessment item shows likelihood 3 and impact 4. Because a test is overdue, the effectiveness rating is void; the displayed score of 12 exceeds the limit of 8. Ranked items are described as sealed.

2. 0:08

   Illustrative routing examples send a duplicate-invoices SQL task to a mid-size language model and a higher-stakes legal question to a large one. The film names demand forecasting and risk assessment as tracks and invites openly licensed models of up to 9B parameters.

3. 0:16

   Two sealed draws represent scoring on unseen questions: the first has 300 fresh items, and the second follows 12 hours later. The 2026 closing card shows reax.co/competitions and an October 30, 12:00 UTC deadline.

A description of the picture and on-screen text, with approximate times. Product details and offers are shown as they appeared in the 2026 films.

1. 0:00

   An animated comparison introduces ‘Same GPU. 6× the decisions.’ For Perplexity Decider v1.1, the bars show 5.3 decisions per second for the BF16 reference server and 31.7 for REAX Engine using FP8 and shared state.

2. 0:10

   The film attributes this result to company-measured Decision Index requests in October 2026 on the same RTX PRO 6000 Blackwell GPU. It separately reports 3.1× on single-question requests, says the score is within noise, and notes that about 2% of answers change.

3. 0:14

   The closing card points to github.com/reaxlabs/reax-engine and describes a miner competition as coming soon at the time of the film.