Video summary

Your prompts kinda suck

Main summary

Key takeaways

Technology

Summary

Theo argues that engineers often make AI prompting harder than it needs to be. Instead of endlessly refining prompts or building a dedicated feature for every task, engineers can describe the outcome they want and let an agent use available tools and data to produce it. Many useful “features,” he says, can simply be requested when needed.

Examples and demonstrations

  • Ask for information instead of waiting for a feature: Theo asks T3 Code to analyze tool-call history, identify slow commands, and create a visual report. He also asks it to visualize how a thread’s context window is being used. These examples show how agents can create useful, one-off views without a permanent dashboard feature.
  • Delegate through prompts: He asks an agent to launch two specialist agents to investigate performance, summarize their findings, and start another thread to implement the results. His point is that users can specify the desired workflow in a prompt instead of needing a new model-selection control.
  • Turn repeated prompts into reusable tools: If a prompt or workflow proves useful, it can later become a skill or an agent instruction file.
  • Use a coding agent for support and troubleshooting: T3 Code’s npx t3 triage command opens a coding agent with machine information and a detailed Markdown playbook. The agent asks what went wrong, gathers reproduction steps, investigates the source and existing issues, and helps resolve the problem or file an issue. Theo says this replaced a much more elaborate feature concept with a small command and instructions.
  • Ask agents about their capabilities: Theo asks a desktop agent to describe its computer-use tools, such as reading interfaces, using screenshots, clicking, typing, and handling browser tabs. He then uses that description to guide development of similar capabilities elsewhere.
  • Connect agents to more of your work: Theo describes linking multiple email inboxes to ChatGPT and using agents to monitor messages, investigate a potentially high AWS bill, and initiate work in other coding tools. He frames this as a way to handle tasks he might otherwise not have built software for or pursued manually.

Engineering implications

Theo distinguishes between prompting for a one-off result and engineering the systems that make powerful prompts possible. He recommends building the underlying primitives—such as integrations, MCP access, data pipelines, and useful context—so agents can act across applications. In his view, AI shifts engineering effort away from implementing every user-facing control and toward enabling agents to access systems and carry out more complex work.

He also acknowledges that AI can make engineers feel as if their hard-won skills matter less. His counterargument is that, once engineers move past that discouraging transition, they can tackle more ambitious systems and problems because agents make more work feasible. He presents this as an argument for rethinking engineering, not abandoning it.

Sponsor segment: Macroscope

Theo showcases Macroscope, a team-activity and code-review tool. Its dashboard summarizes recent work, shows contributor activity, and reports what has been merged. He also demonstrates configuring custom agents to run on pull requests, for example, to check UI consistency and avoid unnecessary new components.

Guide or tutorial content

This is primarily a conceptual guide to prompting and agent workflows, with demonstrations rather than a step-by-step tutorial. The practical advice is to ask for the result you want, explore what an agent can already do, and turn successful recurring workflows into reusable skills or instructions.

Main speakers and sources

  • Theo — presenter and source of the opinions, demonstrations, and T3 Code examples.
  • Macroscope — sponsor featured in the product demonstration.
  • AI coding and assistant tools — including T3 Code, Claude Code, Codex, and ChatGPT, shown in Theo’s examples.

Rate this summary

Your feedback will help improve summaries.

Improve this summary

Reprocess with a stronger model when the summary feels incomplete or inaccurate.

Pro

Translate summary in another language

Pro

Ask questions to this video

Chat for follow-up questions, clarifications, and source-backed answers.

Coming soon

Share this summary

Original video