GPT-6 Astra: Streamlining Skills and AGENTS.md in Codex

X postJoe DevonResource

Joe Devon suggests running a Codex audit on existing skills and AGENTS.md files for GPT-6 Astra, citing an article by Eric Provencher on eliminating bloated instructions carried over from earlier models.

Key points

  1. Legacy instructions tailored to models like GPT-5.6 Sol or Luna overconstrain GPT-6 Astra and consume unnecessary context.
  2. Having too many or overly verbose skill descriptions causes Codex to truncate descriptions, leading to poor skill selection.
  3. Adopting progressive disclosure is recommended: skill files should serve as minimal routers pointing to supporting documents only when needed.
  4. GPT-6 Astra runs tests and checks autonomously; obsolete AGENTS.md rules often trigger redundant test runs or needless whole-repo reads.
  5. Astra takes decision boundaries strictly and can be tentative about when to stop, requiring explicit completion definitions and permissions for safe workflows.
  6. Devon enforces Single Source of Truth (SSOT) doc conventions, mandates numbered diffs before edits, and highlights configuration drift in multi-model workflows.

Why it matters

As coding models evolve, legacy workarounds and scaffolded prompts in repository configuration files actively degrade agent performance and waste context.

Analysis

The release of GPT-6 Astra highlights the problem of instruction drift: detailed recipes designed for older models like GPT-5.6 Sol now restrict newer agents with better baseline judgment. Provencher stresses minimal descriptions, progressive disclosure, and explicit completion criteria over exhaustive rulebooks. Devon's observations underline the operational friction of this transition, especially in multi-agent environments where different models require conflicting steering in a single AGENTS.md.

Original post

Joe Devon

@joedevon · 5 September 2026

Codex users, do this now for Astra: --- Start Prompt --- "Codex, read @pvncher's article t.co/V1BzZOjcwE then audit all of my skills and AGENTS.md files inside ~/Projects" --- End Prompt --- You're welcome.

2.8k likes26 replies8.6k bookmarks680.8k views

View on X
More posts in the thread (10)
  1. @nandanpri @pvncher Haha yeah common. I have strong conventions about which docs are SSOT (single source of truth) for what and to NOT write docs anywhere else to keep files from drifting. I also ask it to tell me if it finds stale instructions . Self improving LlMs are the best LLMs.
  2. @liewcf @pvncher It sure can if you steer it right, @liewcf. My favorite way is you download surf cli by @nicopreme and Chrome Dev or Beta, which will exclusively be your agents' browser. You install the surf chrome extension on Dev/Beta. Then log in to X or whatever you want your browser to
  3. @PamphileRoy @pvncher I'm not sure what you're proposing. Just my perspective, but to me LLMs are magic. It's amazing it works at all. For awhile it will need steering and it's tricky because Astra needs different steering than other models. What could the codex team have done different?
  4. @Secondmindsys @pvncher Not bad. RE: read only part, I've got instructions in my AGENTS.md that requires before editing, it must provide a numbered diff so I can accept or reject the change by number, with a cap for size. For every change it it must delete as much as it added. This forces it to speak
  5. @evoclock @pvncher It is a huge problem that models need different AGENTS instructions. I've got codex and claude $200 plans, and tend to only use the best models, but when they delegate to subagents, I worry about the drift. And it may mess up Fable 5.1 because I wired it up to import AGENTS.md.
  6. @renatiltskin @liewcf @pvncher Here's what I do: t.co/niIYZyOlnC
  7. @evb84 @pvncher Yes. Check out @doodlestein's insane repository of tools, it's very easy to set it up. Offhand I think it was CAUT that tracks usage.
  8. @renatiltskin @ikofranzin @pvncher Yup. There's a lot of tools for that. Here's what I do: t.co/niIYZyOlnC
  9. @Bakaburg1 @pvncher Huge problem. Let me know if you solve it lol. I'm still getting used to astra. It's not working well for me yet. Some combination of hooks and aliases to override default behaviour may be needed. Or more likely you need a default that works well for all agents.
  10. @pvncher This tweet blew up in multiple languages. Kind of a weird experience. Also, the reply counts are not accurate. A ton of replies you can only see by clicking through specific tweets. Many seem to be marked spam when they aren't.
Selected replies (5)
  • @liewcf @joedevon @pvncher You know codex can’t read X article, right?
  • @Secondmindsys @joedevon @pvncher Nice. Add this to the prompt: “Keep the audit read-only. For each proposed change, identify its original purpose, where it applies, and what evidence suggests it’s obsolete. Separate model workarounds from project requirements and permission boundaries. For every workaround
  • @nandanpri @joedevon @pvncher Ran a version of this after the Astra unlock. Pointed Codex at AGENTS.md + skills and it caught two stale path rules I’d been pasting by hand. One prompt beats half an hour of “why is it ignoring my conventions.”
  • @AndrewLachlann @joedevon @pvncher The article literally says Astra does this on its own. Why do you need to prompt it to do this? Way too much of this BS 'prompt this or that' stuff online. It's an LLM, just talk to it.
  • @joko76ers @joedevon @pvncher t.co/yB4MzXIpSw t.co/BF5OJUCGKC

Mentioned links

Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • Video

    Video:CURT

    Decluttering Your AI Setup: Optimising Claude Code and Codex

    Constantly switching between AI environments like Claude Code and OpenAI Codex often leads to accumulated dead skills, outdated memory files, and conflicting instructions that consume context windows. This video presents a structured method to audit past usage, consolidate instructions around AGENTS.md, isolate project contexts, and convert complex tools into lean Markdown playbooks.

    AI & LLMs· Tutorial

  • X post:PostHog

    The Evolving Role of AGENTS.md and CLAUDE.md Context Files

    This discussion explores the effectiveness of using context files like AGENTS.md or CLAUDE.md to guide AI coding assistants. While some developers find them essential for project-specific instructions, others argue they often become technical debt that grows stale and hinders model performance.

    1.1kbookmarks48kviews

    AI & LLMs

  • X post:darkzodchi

    Running GPT-6 Astra in Production: Cost Cliffs and Optimisation

    A guide by darkzodchi analyses technical limits and pricing mechanics in OpenAI's GPT-6 Astra. Despite a 1.05-million-token context window, hidden pricing cliffs and TPM constraints can quickly trigger unexpected costs.

    2.4kbookmarks224.2kviews

    AI & LLMs· Tutorial

  • Video

    Video:AI News & Strategy Daily | Nate B Jones

    Agent Skills: Why Blindly Installing Them Fails and How to Build Them Properly

    Nate B Jones argues against indiscriminately downloading agent skills from repositories, explaining how skill bloat, vague descriptions, and instruction conflicts degrade AI agent performance. He details how loading mechanisms work and promotes building inspectable, focused skills.

    AI & LLMs· Opinion

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.