A new open-source project called beast-skill is packaging the production principles behind one of YouTube's most-watched channels into three workflows for Claude, an AI assistant. The project translates a public onboarding document written for new production hires into skills that can plan a long-form video, render an edit from a typed plan and clean chatbot-style language out of the copy surrounding a release.
The document became public in 2024 and is a 36-page guide to how the most-subscribed channel on YouTube approaches planning, structure and editing. Its central claim is blunt: in one video, 21 million of the 60 million people who clicked were gone within the first minute. The playbook calls that above average. The toolkit tries to turn that observation into repeatable checks and production steps.
From Internal Playbook to AI Workflow
The playbook is built around an average video length of 13 minutes and 37 seconds. In that frame, the first minute is a hook that must match the thumbnail. The next two minutes are meant to show crazy progression without telling viewers what is happening. Re-engagement one arrives around minute three and is described as a simple, exciting, fast scene change. Re-engagement two lands around minute six, where longer explanations can live because the viewer has already committed. The back half is designed never to signal the end, and the ending cuts as soon as the payoff lands.
For videos that are not 13 minutes long, the structure scales. The toolkit places re-engagements at roughly 23 percent and 46 percent of the total runtime. An eight-minute video would put them near 1:50 and 3:41. A ten-minute video moves them to about 2:18 and 4:36. At 15 minutes, the targets are close to 3:27 and 6:54. At 20 minutes, they land near 4:36 and 9:12. The math is simple, but it gives editors and producers a shared map before a single frame is shot.
The planning skill is not a single template. It has six modes that are selected from what the user brings to the table: packaging-first planning, a minute-by-minute map, edit notes, retention diagnosis, sponsor integration and production operations. Every note is supposed to carry a timecode and the metric it protects, whether that is click-through rate, average view duration or average view percentage. The skill also marks which recommendations come directly from the playbook and which ones are scaled interpretations for a different length or format.
Three Skills, One Production Pipeline
The project is split into three parts. The first, beast-skill, plans and diagnoses long-form videos. It can start with titles and thumbnails, build a minute-by-minute structure, write an editor brief, diagnose a retention problem, integrate a sponsor or organize shoot logistics. A sample request asks for a 15-minute video about 48 hours on a deserted island with $10, then asks for titles, a thumbnail concept, structure and an editor brief. The output is meant to be specific enough to hand to a production team.
The second, beast-remotion, is a working Remotion and TypeScript project that renders the plan as a MrBeast-style edit. Remotion is a framework for making videos with code, and this template uses one typed edit plan to drive the whole video. The plan includes a title, a preset, frame rate, dimensions, a list of thumbnail promises, clips, word-level captions, overlays, effects, sound effects and markers. A clip can point to real footage or to a placeholder that names the shot it stands for. That means a creator can plan the entire video before shooting and see a moving storyboard instead of a static outline.
The template includes word-pop captions, camera moves, graphics and synthesized sound. Captions can run in two-word chunks in the classic style or as keywords only in the modern style. Money turns green, numbers and keywords turn yellow, and danger words turn red. Captions step aside for big statements. Camera work includes punch zooms, decaying shake, slow push, tighter second-angle crops and freeze frames. Graphics include big text that slams in, money counters, challenge timers that go red in the last ten percent, day stamps, stair-stepping price tags, circle and arrow callouts, elimination stamps, name tags and white flashes.
The sound design is built from eight effects: whoosh, pop, ding, boom, cash, riser, tick and buzzer. They are synthesized from math by a project script, so there is nothing to license. Music ducks under speech automatically. The checker is the most important part of the rendering skill. It fails a plan when thumbnail elements do not appear before the hook ends, or when the plan has no re-engagements or payoff. It warns on late re-engagements, long shots, pacing problems, end signals, dead air and subscribe asks after the payoff. On the demo plan, the checker passes because a thumbnail element appears before 14 seconds, re-engagement one lands at 22.6 seconds against a 22-second target, re-engagement two lands at 44.8 seconds against a 44-second target, pacing hits 25 cuts per minute before the first re-engagement and the video ends 2.1 seconds after the payoff.
Rendering the Edit as Code
The rendering template ships with two presets. The classic preset reproduces an older, faster retention style. The modern preset follows newer pacing and is the default for new uploads. Public reporting has counted 38 cuts in one minute of a 2023 video and 23 in a 2024 video, after the creator said he had slowed down and focused more on story. The two presets let a production team choose the older density or the newer, more story-led rhythm without rewriting the underlying plan.
The checker is designed to catch structural mistakes before an editor spends hours on a timeline. It does not judge taste. It enforces the parts of the playbook that can be measured: whether the thumbnail promise appears early, whether re-engagements arrive on time, whether the first section moves quickly enough, whether the ending signals itself too early and whether the video keeps talking after the payoff. For a creator who is learning retention, those checks can be more useful than another round of subjective notes.
The project also accounts for licensing. Remotion is free for individuals, non-profits and companies with up to three employees. Larger companies need a Remotion company license. That detail matters because the template is meant to be copied and used, not just read. A team can install the three skills together as a plugin or copy one skill at a time into a personal skills folder. The planning skill can be used on its own, and the rendering template can be used as a standalone project.
Cleaning Up the Language Around the Video
The third skill, deslop, targets the words around the video. A video can be cut at a rapid pace while its description, pinned comment or documentation still reads like a chatbot wrote it. The skill finds and rewrites common tells in English and Korean. It knows more than 40 patterns, including em dashes, the it is not X, it is Y reveal, bold-label bullets, delve and tapestry vocabulary, chatbot openers and sign-offs, vague attributions, Korean translationese and stock closers. It skips code blocks and can run as a zero-dependency TypeScript checker on Node 22.18 or newer.
The checker can be pointed at a draft and given a maximum slop density. If the draft reads too much like a bot, the process exits with a failure code. That makes it usable in continuous integration, where documentation can be checked before it is published. The project says its own documentation passes the check. A before-and-after example shows the intended effect. The before version opens with a fast-paced world, calls video editing a superpower, asks the reader to dive in, adds a moreover, says mastering retention is crucial and highlights the pivotal role of the first minute before ending with a hope that this helps. The after version says most viewers decide in the first minute whether to keep watching, so that is where editing effort pays off first.
The language skill is a natural companion to the planning and rendering skills. A production pipeline that uses timecodes, retention markers and a strict ending can still lose credibility if the public-facing copy is full of generic AI phrasing. Deslop treats the copy as part of the edit. It removes the phrases that make a viewer or reader feel like they are being addressed by a template rather than a person.
Who It Is For and What It Signals
The toolkit is aimed at creators, editors, producers and small teams who want a structured starting point for long-form video. It is not a replacement for taste or footage. It is a way to make the planning conversation concrete. Instead of asking whether a video is engaging, a team can ask whether the thumbnail promise appears before the hook ends, whether the first re-engagement lands at 23 percent of runtime and whether the payoff is followed by a clean cut. Instead of arguing about pacing, a team can compare a cut count against a chosen preset. Instead of rewriting a description by feel, a team can run a checker and see which phrases read like a chatbot.
The project is unofficial and is not affiliated with MrBeast or Beast Industries. It does not contain the original production document. That boundary is important. The value is in the translation: a public playbook becomes a set of prompts, a typed edit plan, a checker and a language filter. For creators who already study retention, the toolkit offers a way to encode what they know. For teams that are new to the format, it offers a map with timecodes, metrics and failure conditions. The final piece of the project is not a promise that every video will go viral. It is a system for making the first minute, the re-engagements and the ending deliberate rather than accidental.
Source: GitHub News