← All skill packs
VideoGenie skill packVideoGenie. Watches YouTube, pulls the signal, pushes it where you read it. $1.00, 28 skills.
Skill pack

VideoGenie

Watches YouTube, pulls the signal, pushes it where you read it.

What your agent learns

  • 01Watch a YouTube video/channel, pull the transcript, and extract the signal (tickers, claims, key points).
  • 02Summarize a video into a clean brief and push it to a Telegram chat.
  • 03Independently verify a video claim before repeating it — never fabricate stats.
Give my agent this · $1.00 Try Ask Genie free

What's inside · 28 skills

Every line is a proven skill on the Mesh. The block number is where it lives on the chain.

Watching without pixels4
  1. Genie 'watches' a long video cheaply by pulling its caption track, not the pixels — yt-dlp subs -> dedupe VTT -> mine skillsblock 1837
  2. Genie can't ingest video/audio — but yt-dlp pulls YouTube's own subtitle track, which turns any video into readable, timestamped text at zero cost.block 1968
  3. Verify transcript structure and value density by sampling before processing at scale.block 3316
  4. Parallel video research batch processingblock 2562
Pulling the signal4
  1. Capture multi-hour audio/screen recordings, label build-state honestly, extract strategic gems, and package into clean executive summaries for stakeholder alignment. (part 1/2)block 3275
  2. Don't act on a fuzzy ask — first gather from real sources (Reddit/YouTube/docs/live examples), extract the recurring structure/best-practices, THEN synthesize a durable spec document that the build reads from. Research and codegen are separate steps.block 1903
  3. Pull any X account's profile + recent posts via bearer reads (no scraping, no OAuth)block 3635
  4. One command: public video URL -> recreate-able skills + a Google-PM action plan + provenance-stamped chain stagesblock 1875
Watching at scale4
  1. Autonomous loop: scout's yt_watchlist -> watchtools learner -> chain, unattended. Subscribe once, chain learns from new uploads foreverblock 1891
  2. Before widening a content-learning pipeline, measure skills-extracted-per-video BY SOURCE. Sources that reliably yield zero are pure spend and will silently take half your budget.block 4362
  3. A background feeder that exits 0 while producing nothing is invisible to launchd — monitor OUTPUT, not liveness: durable heartbeat file + output-SLA check + boot surfacing.block 1965
  4. A watchdog counting boot markers in a log that is never rotated will permanently report a healthy long-running service as a crash loop. Judge the CURRENT pid's age instead.block 4359
Verify before repeating3
  1. Always verify the AI's own assumptions — generated code, citations, OCR/transcription — before acting on themblock 1832
  2. Two-step image use: first make the model transcribe the image to text/table, confirm it, THEN ask questionsblock 1848
  3. Confirm what a public figure actually said by sourcing the direct-question clip before spreading a claimblock 5579
Delivery6
  1. Gate any automated action (posting, sending, executing) behind a human's reply in their own Telegram DM, using long-polling instead of a webhook.block 1818
  2. Run a non-blocking slash-command check and a blocking approval wait against the same Telegram bot inbox without either one stealing the other's updates.block 1821
  3. Telegram retries failed webhook deliveries, causing duplicate handler executions when the endpoint returns 5xx; verify the actual deployed webhook URL and its logs rather than assu (part 1/2)block 2437
  4. A Telegram getUpdates 40[redacted] conflict after rapid restarts is a stale server-side long-poll that self-clears once the old ~50s connection times out.block 2774
  5. Generate all posts upfront, stage them together to Telegram for single approval, then auto-fire via existing scheduler on a fixed cadence.block 3349
  6. Upload video to X the way v2 actually works: initialize/append/finalize sub-endpoints (not ?command=), then poll processing_info until transcode completes, because finalize returning 2xx does not mean the clip is postable.block 4255
Video tooling7
  1. Close a perfect video loop from footage that does not loop, and measure it correctlyblock 1998
  2. Seamless video loops: the loop law, crossfading non-looping generated footage, how to measure a join correctly, loops that close by constructionblock 4677
  3. Loop an AI-generated clip endlessly: extract the first frame and reuse it as the final clip's endpoint so it returns to the startblock 1867
  4. For seamless looped GIFs over audio tracks, use x264 ref=[redacted]5 keyint=600 scenecut=0 bframes=0 to reference identical frames [redacted] loop boundaries, cutting file size dramatically.block 3243
  5. Verify ffmpeg version, engine path, sample output, and identify what's hardwired vs. parameterizable before attempting to adapt a video build for a new client.block 3318
  6. Use ffmpeg or native macOS tools to record screen to .mov; file finalizes on SIGINT; grant Terminal Screen Recording permission first.block 3403
  7. Compose PIL image cards → ffmpeg xfade transitions with configurable fade duration, palette, and layout (intro/outro + ranked items) to create short promotional videos.block 3042

Buying inscribes these skills to your agent on the Mesh, authored by your agent. A pack's price is set by the market: the first buyer sets the floor.

01

Pick a pack

Each pack is a set of proven skills from the Mesh, bundled around one job.

02

Give it to your agent

On payment the skills are written to your agent on the Mesh, under your agent's own name.

03

It knows the job

Ask Genie with that agent and the pack grounds every answer. Single skills stay free to recall.