Skip to content
Haavn' Fun · Past
091 April 2026

Agent research

Extending your AI into the corners of the internet it can't see. Twitter, YouTube, research papers, and the walled gardens in between — with the CLI tools to reach them.

Your AI is blind to the best signal

Your AI has broad knowledge and can search the web — but its reach ends at the walled gardens. No Twitter, no TikTok, no Reddit, no research papers, no paywalled news. The best signal lives in the places AI can't see. This session mapped those blind spots and covered the CLI tools that extend what your AI can actually reach: Bird for Twitter, yt-dlp for YouTube, Scrape Creators for TikTok, Last 30 Days for cross-platform, and agent browser use as the universal fallback.

Key takeaways

Web Fetch returns summaries, not source text. The search engines powering LLM web access summarise articles before passing them to the model. Your answer comes from that summary, not the original. Quotes pulled this way have a good chance of being wrong. If a source matters, feed the actual PDF or Markdown file into the context window.

CLI tools unlock walled gardens without paid APIs. Bird piggybacks off logged-in Safari cookies to give your agent full Twitter access — no API key required. Sebastian used it to surface chatter about the March 21st Claude Code source-map leak that no LLM could find. yt-dlp handed a full Diary of a CEO transcript to the context window in seconds. Abdel had already used yt-dlp to extract transcripts from 70 training videos in one run.

Deep research shines on evergreen topics, not breaking news. ChatGPT's deep research runs GPT Pro under the hood, spends 20–30 minutes, and processes 50–200 sources. Sebastian uses it before every new client meeting for competitive research and industry primers. It has no access to Twitter, YouTube, or TikTok — anything on a short timescale needs the CLI tools instead.

Tools covered

  • Bird — Twitter/X access via logged-in cookies. By Peter Steinberger. Search, pull, and do anything you can do on Twitter, from code. To install, tell Claude: "Read bird.fast and set up bird CLI for me."
  • yt-dlp — YouTube video and transcript downloads. Long-standing, reliable, hands transcripts to your context window. To install, tell Claude: "Read github.com/yt-dlp/yt-dlp and set up yt-dlp for me."
  • Scrape Creators — cross-platform social media search (TikTok, X, Instagram). Built for creator analytics, usable for research.
  • Last 30 Days — comprehensive recency search across Reddit, X, YouTube, HN, Polymarket, and the web. Wraps Bird and Scrape Creators.
  • Agent Browser + Dogfood — when no API or CLI exists, drive the browser directly. Dogfood skill records screens and reports on UI flows.
  • Deep research — built into ChatGPT, Claude, Gemini, Perplexity. ChatGPT's is best; uses GPT Pro under the hood. Best for evergreen topics, competitive research, industry primers.
  • Built-in connectors — Claude connects to Google Drive, Gmail, Notion, HubSpot, Clay, Apollo. Your own documents are the richest source AI can access.
  • NotebookLM — upload books as text files, turn them into queryable sources. Best for knowledge that isn't on the public internet.
← All workshops