Update the mental model of difficulty

  • Theo uses an XKCD example contrasting an easy GPS lookup for national parks with the once-difficult task of recognizing a bird in a photo. His point is that engineers' old assumptions about implementation difficulty can hide tasks that AI now makes accessible. 3:52 4:15 4:35
  • Some feature requests are really requests for information. If an agent can access and analyze that information, the user may be able to obtain the result without a dedicated product interface. 6:59 7:26

Tool-call audit

  1. Theo had audited agent histories and found Claude issuing ripgrep commands with Bash-compatible syntax and newline formatting that failed under Zsh. He associates this with higher tool-call failure rates on his Zsh machines; no percentage is given. 5:02 5:29
  2. A friend wanted a product feature ranking tool calls by duration. Theo instead opened a projectless T3 Code thread and asked for a breakdown of tool calls on the current machine to identify slow operations. 5:54 6:19 6:39
  3. The agent returned a report as a page without an explicit request for that format. It identified a home-directory search lasting 41.7 minutes and another lasting almost 10 minutes. The actionable lesson was to search a repository or known directory. 8:57

A dashboard can be a conversation

  • Theo had been designing a task dashboard that could inspect threads, suggest merges or closures, rename work, and recommend follow-up prompts. The example of 40 threads, eight to merge and 20 to close, illustrates the proposed workflow rather than a reported audit result. 7:51 8:16
  • As buttons accumulated, he reconsidered the interface as a meta thread connected to T3 Code instances across environments. Users could request the particular dashboard or operation they needed. 8:36

Context audit

  • His next request asks for a breakdown of the current thread's context window and a visualizer. A similar earlier request revealed that his Railway proxy setup was adding about 70,000 tokens to every request. 9:22 9:44
  • He says the extra context had gone unnoticed for two months and uses that experience to argue against obsessively managing context. This is his inference from one setup, not a general performance comparison. 10:06
  • The visualizer finishes later in the video. It can then be revised through follow-up questions or saved as a skill. 13:21

When this approach is useful

  • Let slower requests run in the background when the result is useful but not immediately needed. The approximately three-minute estimate for the first report is an expectation, not a measured benchmark. 6:59 10:06
  • Ask for small, personally useful outputs that would not justify a standalone product. Theo cites a visualizer for one bug and combining music listening history with old Last.fm scrobbles. 10:34 11:02