Put differently, plays went up 5x: how an open-source skill for '$0 What If animations' got its stars

Baptiste O'Donovan, an AI consultant, broke down the 'What If' stick-figure animations that October blew up on YouTube into a formula and made it an open-source skill that runs on both Claude Code and OpenCode. In the same repo he posted two videos: the first on 'how to do it with Claude', 35k views; the second on 'do it without paying for Claude, $0', 195k views in 5 days, and the repo's stars rose from 37 to 245 along with it. For people who want their open-source project to get stars, the takeaway from this case isn't the code. It's the path of 'riding a content format that's currently hot + making your own video demo', and where that path's ceiling is.

A 7/15 storyboard

In Oct 31, O'Donovan posted 'I found a way to make these AI animations for $0' on YouTube. In the video he runs a comparison: the same skill, the same computer, the same prompt 'What if you could see one minute ahead?', with one side handed to paid Claude Opus and the other to the time-limited free GLM-5.3-Flash Free in OpenCode, each run to completion without interruption. The free model took 28 minutes and rendered the video in one pass. Along the way it first drew a still frame for each of 15 shots and checked them itself, judging 7 of them 'hard to understand', for example a ball covering the character's face, and only started animating after fixing them all.

His conclusion in the companion article is that the two finished videos are 'hard to tell apart'. The real differences are just two: Opus searched and found that five channels had already done the 'one minute ahead' topic and switched its idea, which the free model didn't search for; and the free model tried several times to adjust sound effects and the mix. 'Run skills with free models, build skills with Claude.'

By Nov 5 this video had 194,698 views, while his channel had only 4,060 subscribers.

What it is

A folder you drop into ~/.claude/skills/. You type one line, 'What if you could rewind 10 seconds?', and it produces a 15–25 second stick-figure animation in dusk tones, through an eight-step process:

1. Write the script using the 'seven-beat What If formula' (the author extracted it from five scripts across four channels: a few quick demos up front, a title card, the fun part, and a final twist based on real, checkable facts);
2. Voice it with ElevenLabs, get timing for each word, and tighten the pauses to this kind of video's pace;
3. Storyboard word by word, with the opening at about $27 per minute;
4. Generate a still for each shot first and assemble them into a storyboard for you to review;
5. Self-check the stills against a 11-item checklist and fix anything that's hard to read;
6. Draw with SVG and render with Remotion; the character is the same function in code, 'so the character never drifts';
7. Add a sound effect to each action, aligned to the onset, plus ambience and background music;
8. Master to -14 LUFS and report how much the episode cost.

The costs are spelled out clearly: the README says the first episode used 152 ElevenLabs credits, and each episode after that costs 170–242; the free allowance is 10k credits a month, enough for about 40 episodes, but the free plan is limited to non-commercial use and requires attribution, so 'read the terms before you think about monetizing.' There's no image model and no video model. That addresses the biggest pain point for people who want to replicate this kind of channel: using video models like Higgsfield costs about $0.4 per 5 seconds, and the character looks different in every shot.

Where the stars came from

(Daily GitHub star records; view counts are from YouTube pages, as of around 11-05.)

Three things stand out.

First, the hook matters more than the product. The first video and the author's X post, said to have 1,689 bookmarks, together brought only about 30 stars. Real traction came from '$0, no Claude subscription needed': functionally it just added another way to run it, yet plays were 5.6 times the first video's. People who watch these videos are mostly those who want to make a 'faceless YouTube channel', not developers, and for them 'free' is decisive.

Second, the stars came from YouTube viewers, not GitHub. It never made GitHub Trending. 72 forks, close to 30% of stars, is much higher than typical projects. The README installs by cloning into the skills directory, so non-developers following the video step by step often clicked fork along the way. Issues are 0, so it's impossible to tell how many people actually produced a video.

Third, it rose fast and fell fast. After the peak, daily stars halved repeatedly and reached only 19 by 11-06. Without a third video from the author, the curve will likely return to flat (inference).

How crowded this niche is

'Having an agent write code to animate' became a small hit in October, but the What If slot was already taken:

(GitHub, as of 11-06; 'about' indicates rounding.)

Placed in this table, 245 stars is mid-to-low. Of the 10 tools in the same X long post, the author's own three are a 3D cross-section backyard (the original was a new TikTok account, 48 videos since Oct 11 with 58.6M views), a stick-figure survival story, and a lecture-style explainer, each riding a hot format. The format itself isn't scarce. Almost every viral short-video format gets a skill on GitHub within a week or two of appearing.

Why it got reposted

Looking closer, two things are its own earned edges:

• It does one format, and does it from one sentence to a finished video. Others are mostly style libraries or toolboxes. This is a complete pipeline of 'script → voiceover → storyboard → self-check → animation → sound effects → mastering', plus a selling point that can be said out loud: the character is code, so it never drifts.
• It states its costs and limits upfront. How many credits per episode, how many episodes the free allowance covers, that the free plan can't be used commercially, that the hand-drawn style isn't supported, and that 'it's not a single prompt; the storyboard step is where the video gets better.' For non-developers, these lines get people to act more than a feature list does.

But the reach came from the author's own distribution: he talks about 'Claude tools' on YouTube and Instagram, and his Instagram grew to 10k followers in 61 days. Each video comes with a text version on freshstack.ai, and beneath the article sit his consultancy's 'AI bottleneck test' and booking link. For him the repo is a lead-generation sign, and he used the same approach for tool-reel-system (a skill for editing Instagram reels, 41 stars).

If it were you

One judgment: the path of 'turning a hot short-video format into a one-line-to-video agent skill, then making your own video demo' is replicable, and a few hundred stars isn't hard. But this slot is already crowded. To beat it, you need to be faster, narrower, and better at telling the story, not better code.

Who it suits: developers who can write Remotion or front-end animation, who already scroll TikTok or YouTube Shorts, and who can spot a format within one or two weeks of it going hot. Ideally someone willing to appear on camera or record their screen, even if the channel only has a few thousand subscribers. This case shows that 4,000 subscribers can still produce a 200k-view video.

Who it's NOT for: anyone who won't make a video and is just hoping people find the repo on GitHub (it never hit Trending, its stars came entirely from outside videos); anyone trying to build a general-purpose "agent video tool" (lemo-opuscar, Remotion's official one, and Chase AI already own that); anyone chasing the What If stick-figure lane itself (kaomei, mromidev and this one are already in it, you're late).

How to do it:

• Pick the format, not the tech. Choose a short-video format that only started charting in the last two or three weeks and is pulling views well beyond its channel size. First check whether GitHub already has a skill with the same name. If it does, pick a different format, or build the part it does badly.
• Reverse-engineer its structure: pull the script beats, cuts per minute and visual rules from several episodes of the top channels and write them into the skill. Draw characters and scenes in code, and sell it on "never off-model, nearly zero marginal cost."
• Fix its weak spots: the voiceover and sound effects depend on ElevenLabs, and the free tier can't be used commercially. A local, free voice version that's cleared for commercial use is itself the selling point (in the same X long post, @finnardley's space narration is what you get when there's no key and it falls back to local voice). Hand-drawn and ink-wash styles are something it openly says it can't do, so that slot is empty (inference). A one-click Windows install with no command line is a hard requirement for non-developer viewers.
• Put the cost per episode, how many episodes the free quota covers, and honest limitations in the README. Include a storyboard image and one finished video.

How to get users: The main battlefield is YouTube and X, not GitHub. Record a demo titled something like "I built X with X and it got a million views," and pair it with a text tutorial; repost the same repo with a sharper hook, and the $0 version gets 5.6 times the views of the first one. Some people on X run "free Claude video tools" roundups, like @theoxmotion. Pitch them the repo directly; getting into a roundup can bring a real wave of traffic.

Before the window closes: each format has about two to three weeks before it's crowded with a dozen copycats. The free model it relies on is "free for a limited time," and the $0 hook can stop working at any moment. From 11 through December, agent-made video is still getting heavily reposted, so move before that window shuts.

Signs you've misread it: (1) format video skills launched over the next month or two mostly stall at a few dozen stars, meaning viewers are already tired of this kind of roundup; (2) Remotion's official project or a large toolkit ships "popular format templates" as a built-in one-click option, and the case for a single-format skill disappears; (3) your demo video passes 100K views but the repo has fewer than 100 stars, meaning this audience watches but doesn't install. Then switch to a different audience.

Data and sources

GitHub repo page and README (11-06); daily GitHub star records (from 10-29); freshstack.ai articles "How to Make AI Animations for $0 Without Paying for Claude" (10-31) and "how to make What If cartoons with Claude Code" (10-24); YouTube videos avIK9YTKpjU, XeZTgiEFxY0 and channel @baptisteodonovan page (around 11-05); X: @theoxmotion 11-01 long post; competitor star counts are figures around GitHub 11-06. The 1,689 bookmarks on the author's original X post are quoted from the @theoxmotion long post; the original post was not read directly.