Brainrot

Brainrot

Sludge content, but the top half is something you need to learn. One HTML file shaped like a phone (9:16, TikTok-style): on top, the lesson in short spoken beats with the current word highlighted and blown up as a caption; underneath, Minecraft-style 3D parkour: a blocky guy running over a void, jumping gaps and fences and sidestepping walls, forever (or real gameplay footage if the user supplies a YouTube link). The parkour is synced to the lesson: every beat recolors the world into a new biome, the beat's key phrase launches him off a slime block through an arc of coins, and finishing turns everything gold. Narration is pre-rendered with a neural TTS voice at build time (OpenAI or ElevenLabs, whichever key is in the environment), TikTok narrator energy, with word timings driving the captions. The eyes have somewhere to go, so the ears can do the work.

You write the beats. A bundled script turns them into the page and opens it.

Two modes

teach — the user names a topic and you write the lesson. If the topic is a real API, library, spec, or CLI, fetch the docs first (WebFetch, or the defuddle skill for articles) so method names and flags are real and current. Remembered API surfaces are where narrated lessons go wrong, and a confident voice reading a wrong flag is worse than a hedge.

focus — the user hands over text, a file, or a URL they need to get through. Read the whole thing before writing a single beat. Beats carry the source's actual content, chunked and lightly rephrased for the ear. Put the analogy in the eli5 field so text can stay faithful. Never add claims the source doesn't make; if a passage is ambiguous, say so in the beat.

Pick focus whenever there is source material. Pick teach when there isn't.

Workflow

1. Get the material (see modes above).
2. Write the beats JSON (schema below) to a temp file, for example ${TMPDIR:-/tmp}/brainrot-<slug>.json.
3. Build and open:

   node <skill-dir>/scripts/build.mjs <beats.json> [--video <youtube url|id>] [--out <file>] [--no-open]
   

   <skill-dir> is the directory containing this SKILL.md. With OPENAI_API_KEY (or ELEVENLABS_API_KEY) in the environment the script narrates every beat with a neural voice first (about 10 to 30 seconds for a dozen beats, roughly a cent, cached by text so rebuilds are free), then prints the output path (default ~/.cache/brainrot/<slug>-<stamp>.html) and opens it. Read its stderr: it warns about beats that are too long or whose key isn't in the text, and tells you which voice narrated. Fix warnings and rebuild before handing over.
4. Reply in one or two lines: the path, and "click to start, space pauses, ⚙ for speed, C for smooth brain mode." If the build fell back to the browser voice, say so and name the missing key. The page is the deliverable. Don't paste the lesson into chat.

Beat schema

{
  "title": "PKCE, but make it sludge",
  "subtitle": "OAuth for apps that can't keep a secret",
  "mode": "teach",
  "source": "optional label: file name, URL host, or 'from memory'",
  "video": { "youtube": null },
  "beats": [
    {
      "emoji": "🔑",
      "title": "The problem",
      "text": "Your phone app wants to log in with OAuth. But it can't keep a secret, because anyone can unzip the app and read the code.",
      "key": "can't keep a secret",
      "eli5": "It's a diary with the key taped to the cover.",
      "code": "optional, eight lines max",
      "lang": "ts"
    }
  ]
}

Required: top-level title and each beat's text. Everything else is optional. mode defaults to teach. Always write key anyway; see below.

Writing beats

• 8 to 20 beats. Fewer and the runner barely gets going; more and attention leaves before the end.
• text is 1 to 3 sentences, 40 words or fewer. It is read aloud. Browser speech gets flaky past about fifteen seconds of continuous speech, and nobody holds a sixty-word sentence by ear. The build script warns at 60.
• One idea per beat. If a beat needs "and also", split it.
• Write for the ear. Short sentences. Contractions. Concrete nouns. No parentheses, no semicolons, no lists inside text. Skip things speech engines mangle: raw URLs, snake_case identifiers, hex, version strings. Put those in code instead, where they're shown, not spoken.
• key is the punchline: 1 to 5 words copied verbatim from text. When narration reaches it the runner does his stunt and the phrase slams onto the video in big letters, so pick the words that carry the idea ("no token", "never leaves your app"), never a connective. Case-insensitive match; if it isn't found the stunt fires about 70% through the beat.
• title is spoken first as its own sentence. Five words or fewer.
• eli5 is shown, never spoken. One sentence. It's where the analogy lives so text can stay precise.
• code is eight lines or fewer and looks runnable. It's the one place precision beats voice.
• Order: the hook (why care) → the one idea → mechanism in steps → the gotcha → what to do now → a closer a sleepy person could repeat back.

Narrator voice

Unhinged is the delivery. The facts stay boring and correct. The runner and the captions are the entertainment budget; the content doesn't need to spend any.

• Talk like a friend explaining something at 1am, not a textbook. Second person, present tense.
• One concrete metaphor per beat, from daily life: bouncers, bank tellers, group chats, lost luggage. Not three metaphors.
• Hype only where it lands: "this is the whole trick" on the one idea, not on every beat.
• Fine: "okay so", "here's the thing", "no really", "this part is dumb but it matters". Not fine: stacked slang, fake enthusiasm on every line, anything you wouldn't say out loud. One "cooked" per lesson, max.
• Mild swearing is fine when it fits. Don't reach for it.
• Never invent. No made-up stats, versions, dates, or flag names. In teach mode, if you can't verify a name, fetch the docs or write "check the docs for the exact name" into the beat. In focus mode, keep the source's meaning exactly; rephrase only for the ear.

Textbook: "PKCE mitigates authorization code interception by binding the token request to the client that initiated the authorization request via a cryptographically random code verifier and its transformed code challenge."

Brainrot: "Here's the whole trick. Before you ask to log in, you make up a big random secret and send only its hash. Later, when you trade the code for a token, you show the original. Doesn't hash to the same thing? No token." (eli5: "You mail a photo of your key first, then show up with the key.")

Focus mode, source says: "Retries use exponential backoff with full jitter, capped at 30 seconds." Beat: "When a request fails, it waits a random amount before trying again. The wait roughly doubles each time, but never goes past thirty seconds." (eli5: "Knock, wait a bit, knock, wait longer. Stop waiting longer at thirty seconds.")

Options

• --video <url|id> starts on real gameplay footage instead of the built-in runner. Use it when the user names a video or asks for "actual subway surfers". Don't guess a video id from memory: ids can't be verified without fetching, and the owner may have disabled embedding. The page has a paste box, so the user can supply one later either way.
• --tts auto|openai|elevenlabs|none picks the narrator. auto (default) uses ElevenLabs if its key is set, else OpenAI, else the browser voice.
• --voice <name|id> overrides the voice: an OpenAI voice name (default ash; onyx, echo, verse are the other male options) or an ElevenLabs voice id (default: the first premade male voice on the account).
• --style "<instructions>" replaces the OpenAI delivery instructions. The default asks for a fast, upbeat TikTok narrator. Use it when the user wants calmer or weirder.
• --no-align skips whisper word alignment on OpenAI (captions then use proportional timing; the key-phrase stunt may land a few hundred ms off).
• --no-cache re-renders clips even if cached.
• --out <path> overrides the default output location.
• --no-open skips launching the browser.

Exporting a video

The page has a 🎬 button (or press E) that records the whole lesson as a clean 1080×1920 file with narration and sound effects, ready to post. It uses the browser's tab capture: the share dialog opens, the user picks this tab (keep "share tab audio" on), the bar and cursor disappear, the lesson restarts from beat one, and about four seconds after GG the file lands in Downloads as <slug>-9x16.mp4 (WebM on browsers that can't mux MP4). Clicking the red REC pill stops early. Works in Chromium browsers (Chrome, Helium, Brave, Edge, Arc); Safari and Firefox can capture but can't crop to the phone, so the export would include the page margins. Tell the user to leave the tab visible while it records.

Gotchas

• Don't ship this through an Artifact tool. Speech synthesis and the YouTube iframe are blocked in artifact sandboxes. The page needs a real tab. The build script writes the file and opens it; that's the delivery.
• The 3D runner loads three.js from cdnjs. Offline, the video panel shows a one-line notice and the lesson still narrates; 📺 with a YouTube link also works once online. Nothing to fix in the beats.
• Browsers won't speak until the user clicks. The page opens on a "tap to start" gate on purpose.
• Narration quality depends on a key being present. With OpenAI or ElevenLabs the voice is neural and TikTok-paced. Without a key the page falls back to browser speech synthesis, ranked by name toward known male voices, and on browsers without Google or Edge voices (Safari, Helium, Brave, ungoogled Chromium) that means a compact macOS voice like Daniel, which sounds robotic. Tell the user which happened. There is no voice or pitch control in the page by design; use --voice and --style.
• Audio is embedded as data URIs, so the page stays one file. A dozen beats add roughly half a megabyte. The build prints the final size.
• Export needs a click in the share dialog. That's the browser's rule for capturing a tab; it can't be automated. If the dialog shows other windows, the user should pick the current tab, not the whole screen.
• Chrome's cloud voices don't report word boundaries. The page estimates word timing for those; local voices (Samantha and friends on macOS) highlight exactly. Nothing to fix.
• Changing speed restarts the current beat from the current word. Browser speech can't change rate mid-utterance.
• "Too cringe" means the beats, not the format. Calm the voice in the JSON and leave smooth brain mode off. The runner alone is quiet.

Example requests

• "brainrot me on how DNS resolution works" → teach, about 12 beats, from general knowledge (no API surface to verify).
• "/brainrot ./docs/ARCHITECTURE.md" → focus, read the file, chunk faithfully.
• "I can't focus on this RFC, https://… just read it to me with the subway surfers thing" → focus, fetch with defuddle, chunk; mention the paste box if they want real footage.
• "teach me tanstack router loaders, make it dumb" → teach, fetch the docs first, then write; code fields carry the real function names.