OpenAI: trim your prompts and skills for GPT-6 Astra
OpenAI's Eric Provencher says bloated skill descriptions and blanket AGENTS.md rules slow Astra down in Codex. Here is what he suggests instead.
In short
OpenAI developer Eric Provencher recommends cutting skills, AGENTS.md files, and task prompts back to task-specific essentials, because GPT-6 Astra decides for itself which material it needs to read.
At a glance
- Eric Provencher of OpenAI lays out leaner instruction practices for GPT-6 Astra in Codex in a developer blog post.
- Skill scopes should be narrow: a migration skill fires only when a migration is created or changed, or its rollout checked.
- Too many broadly scoped skills force Codex to truncate their descriptions.
- Provencher says Astra may stop earlier than GPT-5.6 Sol, and literally read old rules make that worse.
- Neither source reports measurements for context use or error rates.
If you run Codex on GPT-6 Astra, the instruction files you accumulated for earlier models are now part of the problem. OpenAI developer Eric Provencher argues that stronger models need less hand-holding, and he names three places to cut: skill descriptions, AGENTS.md, and the task prompt itself. Astra works out for itself which material it has to read.
Narrow the scope of every skill
A skill is a Markdown file that bundles resources and scripts. Provencher wants its scope described in a short, precise line. His example is a Postgres migration skill that fires only when someone creates or modifies a migration, or checks its rollout, rather than on general database work.
The main document should point at supporting files briefly instead of pulling them in wholesale, so the model reads only what it needs. Pile up enough broadly scoped skills and Codex starts truncating their descriptions. Contradictory wording compounds this by loading guidance nobody asked for.
Stop making the agent read everything first
AGENTS.md carries repository-wide rules, and Provencher suggests revisiting it on a regular basis. Forcing the model through several mandatory documents before every change spends context without earning anything back. Selective pointers do the same job: architecture notes for service boundaries, database docs for schema changes, deployment guidance when shipping.
Safe workflows deserve standing permission rather than repeated approval requests. In his framing, the agent should be able to run tests, fix errors caused by the change it was asked to make, and re-run the affected tests without asking again.
Say what counts as done
Task prompts work better with clear completion criteria than with a long list of granular steps. State whether the agent should implement, verify the result, and repair what broke. This matters because Astra may stop earlier than GPT-5.6 Sol even with no restrictions in play, and it can read old limits so literally that it halts while you still want it working.
What the guidance does not tell you
Every recommendation here is qualitative. The sources give no figures for how much overlong instructions degrade output, and none for context consumption, premature-stop rates, or error rates. Whether trimming pays off in a particular repository is therefore something teams have to measure themselves.
FAQ
Do I need to rewrite my AGENTS.md for GPT-6 Astra?
Provencher recommends reviewing it. Requiring several mandatory documents before every change burns context, while targeted pointers to architecture, database, or deployment docs do the same work for less.
How narrow should a skill description be?
Narrow enough to name the trigger. His example is a Postgres migration skill that activates only when a migration is created or modified, or when its rollout is checked.
Why would Astra stop before finishing a task?
It may stop earlier than GPT-5.6 Sol even without restrictions, and it can interpret old limits literally. Stating up front what counts as finished work is the fix Provencher suggests.
