Careful — Don't Over-Write Your Prompts in the GPT-6 Astra Era!

Hi, it's me, Shiichan! Today I found a small but genuinely useful post on OpenAI's developer blog, so let's dig in.

OpenAI Developers developers.openai.com

What was announced?

OpenAI's developer blog published a post about rethinking how you write skills and prompts for GPT-6 Astra. The gist: as models get smarter, the old habit of spelling out every instruction in detail can actually work against you.

The story so far

Up until now, the standard playbook was to write long skill descriptions, load AGENTS.md with lots of documents to read, and carefully forbid anything that felt risky. That worked fine for older models. But with a model as capable as GPT-6 Astra, the same style of writing can waste context or make the model overly cautious and hesitant to act.

What changes

The post lays out several concrete adjustments for writing prompts and skills around GPT-6 Astra.

• Keep skill descriptions short and specific about when to use them (e.g. "use when adding a migration" instead of "use for anything database-related")
• For skills with multiple workflows, keep the first thing the model reads minimal, and disclose steps progressively as needed
• Trim AGENTS.md so it doesn't force multiple document reads on every edit — scope it to the actual context
• Drop instructions that nudge the model to run tests — it tends to do that on its own already
• For safe workflows like local testing, explicitly allow the model to keep going without asking for approval each time

Dive Deep

One interesting point is about judgment boundaries. Restraint instructions written for older models — the "don't do this" kind — can backfire on GPT-6 Astra. Since Astra already has a high bar for judgment, piling on strict constraints can leave it stuck rather than careful.

Another point is about defining "done." GPT-6 Astra tends to be cautious and often stops after a first pass to check whether it got things right. The recommendation is to explicitly define what counts as complete in the prompt, and instruct the model to carry the work through to that point instead of stopping early.

The post also touches on how skills work under the hood: only a skill's name and description are always loaded into the model's context. If the description is too long, it risks getting trimmed when the context gets compressed — so keeping it short and to the point pays off in practice.

Wrap-up

• OpenAI's developer blog shared pointers for rewriting skills and prompts specifically for GPT-6 Astra.
• Keep skill descriptions short and scoped to "when to use it"; trim AGENTS.md to the relevant context.
• Reconsider over-restrictive guardrails and test-running nudges written for older models.
• Explicitly defining "done" helps Astra carry tasks through to completion.
• Worth revisiting now if you're running agents with Codex or AGENTS.md-based setups.