cinetic: an agent skill that makes a coding agent direct a launch film

Idea before code, and three concepts tested before one is chosen

The first thing cinetic does is not code. The README describes an idea first approach: before any code, the agent writes three concepts through different lenses, tests them, and picks one story device the film owns. That ordering is the difference between this and an animation component library with a prompt wrapper. The claim about the resulting work is that a film owns a single story device rather than a pile of effects, and the Kiln example illustrates it. A 25-second launch teaser was made from a two-sentence brief, and the described scene is one changed line turning a wall of 494 build targets red as everything rebuilds, with 491 coming from cache so the timer stops at 0:09, and a red keystone completing the arch of the kiln logo. The README is explicit that the agent branded, directed, scored and rendered it. The teaser exists as a 1920x1080, 60 fps MP4 with sound at 7 MB, with a GIF as the excerpt.

A technique library that draws the plan for you

The second mechanism is a library of 273 techniques across 14 categories, each with a build recipe and timing at 60 fps. A script called pick.py draws the film's plan, weighted for the format and the energy of the piece, and the reason given is that two films should not share the same fade-up, push-in and end card. This is a meaningful design decision. Most generative video work is reproducible in the bad sense: the same default transition appears in every output. Here the plan is sampled rather than fixed, and the weighting is by format and energy, so a vertical Reels cut and a horizontal landing page loop do not draw the same vocabulary. The techniques are described as measured moves rather than presets, and the timing data at 60 fps is what lets the picture, the sound and the QA scripts share a common clock later.

The brand comes from three words, not a template

The look is derived rather than selected. The README says the brand comes from three personality words, so a calm notes app and a fast CI tool do not end up looking alike, and the Kiln example is credited to an agent that invented the brand. This is a defensible answer to a problem most AI video tools avoid. If the system has one house style, every product's film looks like the same product's film. Three words as the seed of a visual identity gives the agent a handle to differentiate on, and the Halden example shows the payoff in a small piece. A 6 second logo sting for a calm note-taking app, mixed at minus 16 LUFS, with four ProRes 4444 alpha versions for held and cleared light and dark footage. In it, the app's text cursor settles into a dot and the h draws its arch over it like a fermata, the musical sign for a hold.

Copy arrives in fragments and the music is measured

The typographic and audio rules are the most specific part of the README. Copy arrives a few words at a time, split around what the product does, with a quoted example of here is the newest, then the product, then its name. Holds are readable and drift rather than freeze, and there are stretches of pure picture between. The fragment grammar is said to come from a frame-by-frame study of eight professional films, and a crisper profile exists when a brief asks for energy. On the sound side, a file called timeline.ts drives the picture, the soundtrack and the QA scripts, so every hit lands on its frame, and the effects come from a separate sound kit called earshot, where sounds are designed in code and rendered with measured sync points, one soft sound per text fragment and one clean element on the lockup. Typing, clicks and ticks follow measured norms of top-tier launch films: soft, short and mostly left to the picture.

A beat grid, then a measured finish with a QA pipeline

Everything is quantised to one beat grid, and the output checks are described in production terms. The finished film has true motion blur, no banding, BT.709 colour and a sync-checked mux, and then it goes through pixel forensics, an audio-sync audit, contact sheets and critic passes against a ship gate. That last phrase matters. A ship gate implies a pass or fail decision on a set of criteria, which is what separates this from a prompt that asks for a video and returns something. The Quire Streaks example shows the same machinery applied to a vertical cut: a 12 second video for Reels and TikTok in the brand the brief supplied, where a blue thread stitches each day's pages into the Streaks screen with one rising note per stitch, and on Sunday at 11:41 pm the thread hangs loose until one tap on Log pulls it tight and the week gathers into a quire, the bookbinding word behind the name.

A linter and critics enforce what a generated video must not look like

The most interesting claim in the README is the negative one. There is a no generated look policy, implemented as hard bans enforced by a linter and by the critics, which keep out the looks that mark a video as generated, and a brand you supply always wins. This is the mechanism that determines whether the tool is usable for a real launch, because the standard failure mode of AI video is not bad motion, it is the specific sheen of synthetic imagery. Encoding that as a lint rule rather than a hope in the prompt means the failure is caught before export. The README also documents the evaluation side: four films were made by an agent with development snapshots of cinetic 1.0.0, in evaluation rounds 6 and 7, from one-paragraph briefs that are the same briefs used in the evaluation section, and the brands were invented. The repository carries an evals directory alongside the skills directory, a .claude-plugin directory, docs and a CHANGELOG.

Installing it as a skill, and what the machine needs

cinetic is distributed as an agent skill through the skills CLI, and the install is one command.

npx skills add Leonxlnx/cinetic

After that you ask your agent for a film in plain language, and the README gives the example of asking for a 20-second launch teaser with sound for a database branching tool, inventing the brand. It states the skill works with Claude Code, Codex, Cursor and the other agents the skills CLI supports, which is the point of using the skills format rather than a Claude-only plugin. The machine requirements are Node 22 or newer, ffmpeg 6 or newer, Python 3.11 or newer and a headless Chromium, and the README asks you to do a first test render before your first real film, with the steps in the setup documentation. The rendering target is Remotion with React by default, or HyperFrames with HTML and GSAP. The project is MIT licensed, and the release history shows v1.0.0-beta on 2026-09-28, v1.0.0 on 2026-10-04 and v1.1.0 on 2026-10-18, with the last push on 2026-10-18.

Editorial conclusion

cinetic fits a team shipping a product and needing a short launch film or logo sting without hiring a motion designer, provided somebody on the team is willing to art-direct the result and to accept a rendering pipeline that needs Node 22, ffmpeg, Python and a headless Chromium. It does not fit a team expecting broadcast-grade finish, and it does not fit a workflow where the brand cannot be invented by the agent, because the hard bans are designed to remove anything that looks machine-generated. Verify first that a first test render completes on your machine, review the hard bans and the critic passes before you believe a film is shippable, and read the LICENSE file rather than relying on the badge.

Frequently asked questions

Official sources

1. Issues
2. Leonxlnx/cinetic on GitHub
3. License: MIT
4. README
5. Releases