<- all tokdocs
2026.09.16@itstundealao

The /brag Claude Code skill turned a storage-cleaner app into a launch video for about $6 of Opus 5 tokens

AI
Watch on TikTok

View on TikTok ->

The skill in this video is real, the $6 price tag is roughly right, and the cost breakdown he reads out loud leaves out the line item that accounts for a third of the bill. Tunde installs /brag from the latent-spaces/brag repo, runs it against a storage-cleaning app he says he shipped the week before, waits about ten minutes, and gets a product launch video with music, motion graphics, and marketing copy. The session panel he shows on screen has enough numbers in it to check his claims against Anthropic's published pricing. Most of them hold up.

What /brag actually does under the hood

/brag is a Claude Code plugin and Agent Skill published by Latent Spaces under an MIT license. The repo went up on 2026-06-16 and has 5,548 stars and 335 forks as of today, so Tunde's "I've never heard of it before" reflects his feed rather than the project's reach. Installation is two commands: /plugin marketplace add latent-spaces/brag then /plugin install brag@brag. Non-Claude agents get it through npx skills add https://github.com/latent-spaces/brag --skill brag.

The SKILL.md describes a four-step pipeline. Claude reads the project source and answers a nine-point planning rubric, writes creative direction and a beat-by-beat storyboard, hands a brief to Hyperframes, then validates with npx hyperframes check and renders the MP4. The split matters: /brag owns the story and the tone, and Hyperframes owns composition, animation timing, and render orchestration. Output lands in a brag-output/ directory alongside the plan, the composition brief, and the share copy. The skill treats 15 to 25 seconds as a hard rule, which is where the video's "20 second launch video" line comes from.

Hyperframes is the part doing the actual video work, and it is not a generative video model. It is HeyGen's open-source HTML-to-MP4 renderer, Apache-2.0 licensed, 51.5k stars, published to npm as hyperframes (latest 0.8.50). You write an HTML composition, headless Chrome captures frames deterministically, and FFmpeg encodes them. That is why the output in the video looks like a polished product page in motion rather than an AI-generated clip. It is a web page being filmed frame by frame. Voiceover is off by default and opt-in through --voice, which routes to Kokoro, hexgrad's 82M-parameter open-weight TTS model.

Where the six dollars went

Tunde reads two numbers off the screen: 61,000 output tokens and 600 input tokens. The panel behind him shows four. Input was 676, output was 61.4k, cache read was 4.5M, and cache write was 237.7k. Model was Opus 5 at 100 percent, cache hit rate 95 percent.

Run that against Anthropic's published Opus 5 rates of $5 per million input, $25 per million output, $0.50 per million cache reads, and $6.25 or $10 per million cache writes depending on the cache TTL:

Line item Tokens Rate Cost
Input 676 $5/MTok $0.00
Output 61,400 $25/MTok $1.54
Cache read 4,500,000 $0.50/MTok $2.25
Cache write (1h) 237,700 $10/MTok $2.38
Total $6.17

The panel shows roughly $6.1, so the arithmetic lands within a few cents if the cache writes used the one-hour TTL. Swap in five-minute writes at $6.25 and the total drops to $5.28, which is further off.

The interesting part is what the voiceover skips. The two numbers he reads account for $1.54 of a $6.17 run. Cache reads alone cost more than every output token the model generated. That is normal for a long agentic session where the same repo context gets replayed on every turn, and it is why the panel's own tip says longer sessions get more expensive even when cached. Anyone budgeting for this skill should price the cache, not the completion.

The ten-minute claim, and one number the video never says out loud

"That took around 10 minutes" undersells it slightly. The session panel reads API 12m 28s and Active 14m 6s. Close enough for a TikTok, and the gap between API time and active time is the local rendering work Hyperframes does outside the model loop.

The panel also breaks down what consumed the context window: /hyperframes-cli at 36 percent, /brag:brag at 11 percent, the brag plugin itself at 11 percent, /hyperframes at 4 percent, and /hyperframes-core at 2 percent, with 47 percent of the session running above 150k context. The Hyperframes documentation is the single largest context consumer in the run, well ahead of the brag skill that gets top billing. The session also wrote +946 lines and deleted 20.

What the video does not verify

The app in the output is branded EasyGB, with slides reading "Finds every duplicate," "Keep the best one. Bin the rest," and "Compress large videos, frees about 24.5 GB." I could not find an EasyGB storage cleaner on the App Store or anywhere else, so there is no way to confirm the app exists publicly, confirm the 24.5 GB figure, or confirm that the copy on those slides reflects anything the app actually does. /brag writes marketing claims by reading source code, which means a number like 24.5 GB could be read from a real calculation in the codebase or invented to fill a slide. Nothing in the video distinguishes the two.

One run is also not a benchmark. The repo's own launch page shows three curated examples (Horse Tinder, Fish Flight School, Taxi for Taxis), and a demo video by definition shows the take worth posting. The skill's requirements are also not trivial for a casual user: Node.js 22 or later, FFmpeg on PATH, the Hyperframes CLI, and an agent that supports Agent Skills. "Install it and it's ready" holds only if that stack is already there.

Key Takeaways

  • /brag is real, MIT-licensed, and installable in two commands from latent-spaces/brag, which had 5,548 stars at the time of writing.
  • The video generation is HTML rendered to MP4 by Hyperframes, HeyGen's open-source renderer, rather than a generative video model.
  • The ~$6.1 shown on screen reconciles with Anthropic's Opus 5 pricing to within a few cents, assuming one-hour cache writes.
  • The cost narration in the video covers $1.54 of a $6.17 bill. The 4.5M cache-read tokens he does not mention cost $2.25 on their own.
  • The run took 12m 28s of API time and 14m 6s active, slightly longer than the "around 10 minutes" in the voiceover.
  • Unverified: the EasyGB app itself, the "24.5 GB freed" claim in the generated slides, and whether any of the marketing copy in the output is accurate to the product. I found no public listing for the app.
  • Running this needs Node 22+, FFmpeg, and the Hyperframes CLI already installed, which the video does not mention.

Resources

Published September 16, 2026. Writeup generated from a favorited TikTok.