I need you to hear me out (it’s REALLY good)
Summary
AI summaries can be incomplete or wrong. Verify anything important against the original video.
Theo reviews the system prompt for Codex, an AI agent based on GPT-5, finding it to be capable but outdated and overly prescriptive, particularly in its frontend guidance. He suggests improvements like simplifying the prompt and focusing on key aspects for better performance.
The video reviews the system prompt for Codex, an AI agent based on GPT-5, concluding that while it has a strong engineering core, it's overly prescriptive, especially in its frontend guidance, and feels dated. The reviewer, Theo, rates the prompt a 7/10 for Codex runtime and a 4/10 for prompt implementation. He identifies several key issues: the prompt is heavily coupled to one specific runtime, the frontend section is too prescriptive, and some rules are presented as universal truths when they are merely taste preferences. He also notes that modern models understand tool schemas well without the main behavioral prompt repeatedly describing implementation details, and that these instructions should be generated by the harness, not embedded in the reusable agent prompt. The prompt also includes unnecessary rules and confusing persistence with permission, creating risks. Ultimately, Theo suggests that while a capable model would obey these rules, it might ignore existing design systems, user requests, or domain conventions to satisfy arbitrary global constraints. He highlights that the prompt's over-reliance on specific instructions and lack of flexibility makes it rigid for modern models. The reviewer offers a better approach, suggesting explicit implementation requests, autonomous proceeding through implementation and verification, and handling analysis, review, explanation, or ambiguous requests without modifying files without clear authorization. He concludes by saying that the prompt is useful but could be improved with simpler, more flexible guidelines.
Verdict
Claims & arguments
-
Codex system prompt
The Codex system prompt is capable but outdated and overly prescriptive, especially in its frontend guidance.
- 0:18 The prompt has a strong engineering core but treats a modern agent like an older model.
- 0:39 It needs exhaustive behavioral patching, aesthetic doctrine, repeated prohibitions, and behavioral patches that make it rigid for modern models.
- 1:50 It spends too many tokens prescribing aesthetics, runtime mechanics, response formatting, and edge-case behavior.
- 5:25 Modern models understand tool schemas well without repeated descriptions of implementation details.
- 7:48 The frontend section is far too prescriptive.
-
Codex system prompt
While the prompt has some good elements, it suffers from significant flaws in its design and implementation.
-
Codex system prompt
The system prompt needs to be rewritten to address its flaws, focusing on a clearer structure and better workflows.
Key Points
- 0:17 The prompt has a strong engineering core but is outdated and overly prescriptive.
- 1:04 Frontend guidance is particularly problematic, being too prescriptive.
- 1:50 The prompt spends too many tokens on aesthetics, runtime mechanics, response formatting, and edge-case behavior.
- 5:25 Modern models understand tool schemas well without repeated descriptions of implementation details.
- 8:38 The prompt confuses persistence with permission.
- 10:07 The prompt mandates specific instructions for actions, which can create unnecessary risk.
- 11:43 Several rules are taste preferences presented as universal truths (e.g., color palettes, letter spacing).
- 13:38 UX guidance is questionable, like preferring symbols over text controls and not compensating for icon-only controls.
- 14:14 The ban on visible instructional text is concerning.
- 15:11 A better rule is to allow models to proceed autonomously with explicit implementation requests.
- 16:52 The reviewer explains how they would rewrite the system prompt based on their findings.
- Review of Codex system prompt, an AI agent based on GPT-5.
Worth watching if: If you're interested in prompt engineering for AI agents, especially Codex, or want to understand the design considerations behind system prompts, this review offers valuable insights and specific examples of good and bad practices.
Get every Theo - t3․gg video extracted like this
One daily email with structured extracts of every channel you follow. Free tier covers 15 videos a month.
Sign in with GoogleNo credit card. Free tier forever.