Legible Film · James Dimachkie

SC 01

Overview

Legible Film is an Agent Skill: instructions and scripts that Claude Code, Codex and other coding agents load when a job calls for them. With it, an agent can draw a launch film in code, cut an interview or a tutorial from its transcript, caption and score it, follow an editor’s revisions, and check the file a viewer will receive. I released the first public version, built from my launch-film method, on 27 October 2026, and version 2.1 on 3 November.

SC 02

The problem

An agent can assemble a video and read its frames and measurements, but it can’t watch the cut at normal speed or hear it. A person can watch and listen, but can’t easily see what the agent did. Mistakes slip through between the two: a cut in the middle of a word, a caption left on screen after its line, music that buries the voice, a revision that quietly changes a scene someone approved, and no record of which version is which.

SC 03

What I did

I designed it so that both of them can read the edit. The edit is a script, film.md: who the film is for, what it should leave them with, then each line in the order it plays, with what is on screen beside it. Strike a word and the agent cuts it. Every render is a new version that never changes, and a diff lists what changed. Picture and sound share one cue list, so retiming the picture makes the old sound stale and the tools refuse to use it. Checks run on the delivered file and report in plain states: pass, look, fail, needs you and not assessed. Nothing counts as watched or heard until a person has watched or heard it.

The agent picks one of six paths by who owns the timeline and loads only what that path needs. A Screening Room page on the person’s own machine holds the cut, its script, the checks, notes pinned to moments and every version. Teams of AI agents wrote the scripts and the documentation under my direction, and a model from a different family reviewed each part.

Legible Film's Screening Room reviewing version 14 of a Tears of Steel teaser. A readiness line reads: v14 checked, 14 checks pass, 3 to look at, 3 need you, 1 not assessed. Buttons for Map, Plan, Cut, Check, Screen, Revise and Deliver sit above the film. Beside it, the script states the story, who it is for and what they should leave with, above a selected line and its note on the join. — The Screening Room on a one-minute teaser that agents cut with the skill from Tears of Steel, with two small fixes by hand (© Blender Foundation, CC BY 3.0).

SC 04

How it is tested

The repository marks each part as tested or experimental and says what each rests on. On its test material, all 240 cuts inside spoken lines landed in silence measured in the audio. Cutting at the transcript’s word edges instead put 52 cuts in speech and 78 in breaths, and clipped 6 words.The README’s Evidence section at release 2.1.0. Test material, not a rate on real projects: cutting at the transcript’s word edges instead put 52 cuts in speech and 78 in breaths, and clipped 6 words.Source When the skill read one professional editor’s two real Premiere revisions, every clip and every placed word matched, the words to within a millisecond.The README’s Evidence section at release 2.1.0. One editor’s real revisions, read by the skill; handing a cut to Premiere is still experimental.Source A self-test of about 2,150 cases with known answers runs automatically on Linux, macOS and Windows on every change, and records any case a machine can’t run.The README at release 2.1.0 and its CI workflow. Each machine records the cases it cannot run. On the 2.0.0 release candidate, before 2.1 added cases, 1,633 to 1,666 passed per system and none failed. A pass shows the scripts measure synthetic known answers; it says nothing about the quality of a particular film.Source In a published test of version 1.0, five requests with one run each on one model, agents using the skill met 35 of 35 of the method’s expectations, against 10 of 35 without.Five requests, one run each on one model (GPT-6.1 Terra through the Codex CLI), graded by a model from another family. The results say what the test could not control: blinding was partial, and the author's own instruction file loaded into both arms. The expectations measure adherence to the method, not the quality of a film; no film was rendered, watched or heard.Source

The parts new in 2.1, including the six paths and the cue list shared by picture and sound, are still experimental. So are the hand-off to Premiere or Resolve, tracking and mattes, tightening a recording and the scorecard. Film-level evaluations for each kind of video haven’t been run yet.

SC 05

The launch film

On 3 November I posted a 1:47 launch film for version 2.1. Claude Code made it with the skill, and I directed it. It shows a request for a launch film for an invented app and the film an agent made with the skill. Then it cuts a 21-second scene into a captioned 6-second clip that keeps the speaker’s face in frame, puts back a pause by editing one line of text, cuts a trailer from an animated short with a written reason for each cut, and shows the checks that still need a person before it closes on the install command. Before it went out, six fresh AI viewers, three Claude and three GPT, each saw one frame a second of an earlier cut, with no context. Every one named the product and listed what it does. Each also flagged something confusing, and those notes shaped the final cut.