Video summary

You're Using GPT 6 Astra Wrong (Here's How to Fix It)

Main summary

Key takeaways

Technology

Summary

The presenter argues that GPT-6 Astra can perform well with less effort and fewer resources than users may expect, particularly when used through Codex. The video covers five common mistakes and how to address them.

1. Setting effort too high by default

Higher effort does not guarantee better results. On the Deep Sweep benchmark, maximum effort scored 73% at about $12 per task, while low effort scored 67% at about $2.19. Other benchmark examples also show modest score gains for substantially higher costs—and sometimes a lower effort setting performs as well as or better than maximum effort.

A website-design demo produced broadly similar results at easy and maximum effort, although maximum effort used more tokens. The presenter recommends starting at easy or medium effort and increasing it only when the results require it.

2. Underusing browser and computer control

Codex can use browser and computer tools to gather information from services that lack an accessible API, MCP connection, or CLI.

In the demo, Astra searched Dribbble for hotel-site inspiration, opened examples, took screenshots, and used the references to create a revised website. It also produced a reference table listing the screenshots and links it used. The presenter notes that this workflow could apply to other websites and applications.

3. Keeping outdated or unnecessary skills

Older skills may be irrelevant to newer models, and a large collection can clutter the context. The presenter demonstrates a skills-audit tool based on Anthropic’s skill-creation approach. It reviews skills and usage logs, checks descriptions, and sorts skills into categories such as “fix now,” “review for decommissioning,” and “save.”

The goal is to identify unused skills and test or improve those that may still be useful.

4. Ignoring voice mode

The presenter describes Codex’s voice mode as fast and responsive, with Astra available at different effort levels. Voice can be used within an existing chat to give instructions directly, or as an orchestrator to manage multiple agent chats.

In the video, the presenter assigns a research task to a new chat through voice and has it produce a short HTML brief. The presenter sees orchestration as especially useful when several agents are working at once.

5. Not specifying how Astra should handle ambiguity

The presenter says Astra may ask for clarification at decision points where older models might make assumptions. For longer agent tasks, prompts should state whether the model should pause and ask questions or proceed toward a specified end state.

Another approach is to have Astra seek approval only after preparing a concrete, reviewable output—and to present proposed solutions alongside problems. Without guidance, it may produce less directed results, so prompts should provide a clear “north star” and relevant constraints.

Guides and demonstrations

The video covers how to:

  • Choose an effort level based on task complexity and cost.
  • Use browser and computer automation to collect visual references and inform a website redesign.
  • Install and run a skill-audit tool.
  • Use voice mode inside a chat or as a multi-chat orchestrator.
  • Write prompts that clarify how the model should handle questions, decisions, and approvals.

Main speaker and sources

  • Main speaker: Chase AI presenter, demonstrating Codex and Astra.
  • Sources and references mentioned: OpenAI guidance on Astra’s clarification behavior; Anthropic and Claude Code, including creator Boris Chorny; Deep Sweep and Artificial Analysis benchmark examples; and Dribbble as a source for design references.

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