Rethinking skills and prompts for GPT-6 Astra

An X Article by an author whose profile bio reads “Codex DX” at OpenAI, written for the arrival of GPT-6 Astra, on what to remove from the instruction surfaces around a coding agent: skill files, AGENTS.md, and task prompts. Category: practitioner guidance from inside the vendor, ~900 words, two illustrated before/after examples, no measurements. Its argument is that instructions accumulate as compensation for a previous model’s failures, and a new model release is the moment to find which compensations have become dead weight.

The argument

Skills are a retrieval problem. Every installed skill contributes a name and description to the always-loaded catalog, and Codex shortens descriptions when the catalog outgrows its budget, so adding skills degrades the model’s view of each one. The fix is short descriptions that state only when the skill applies: the illustrated example keeps the same “create and validate Postgres schema migrations” intent but narrows the trigger from “databases, queries, models, or persistence” to “adding or changing a migration, or reviewing its rollout” (fig-3). Two further rules came with an update to Codex’s skill-creator: multi-workflow skills should be a minimal router over supporting documents rather than a single body, and skills written as “elaborate itineraries or recipes” now hinder a model that handles nuance unaided. A multi-model caveat closes the section: repository skills are read by other contributors’ agents, and guidance tuned for one model may overconstrain another.

AGENTS.md carries scar tissue. Instructions to read a stack of documents before every edit, or to run tests and check work, were prompts for a model that would not otherwise do so; the new model does both unprompted, so the instructions now burn context or trigger unnecessary test runs. The illustrated replacement is contextual pointers (“database.md for schema changes”) over a blanket read-first rule (fig-4).

Boundaries and persistence are the new tuning axes. Strong “ask first” language added against an earlier model’s overreach now makes the new one stop where the user would rather it continue, because it “takes your boundaries seriously.” The post concedes a behavior change relative to its predecessor, GPT-5.6 Sol: Astra is “more tentative about when to stop” and may return after a first implementation. The remedy is to define completion before starting and to grant standing permission for workflows known to be safe, with a sample AGENTS.md clause authorizing a local test loop without per-step approval. The closing advice is to ask the model itself to audit the repository against the article.

Assessment

The post is written from inside a model vendor’s product team, and it reads as a field instance of weng2026a‘s prediction that harness functions internalize into models while interfaces persist: read-first and test-after scaffolding is exactly the kind of function that just internalized, and the description catalog is exactly the interface that did not. Against vincent2025, whose 2,249-memories null result argued that most accumulated lessons are already absorbed, this post gives the absorption a clock: a model release.

• Durable: descriptions as index keys competing in one fixed budget, so a description’s job is to say when, not what; the router-over-recipe shape for multi-workflow skills; and defining completion up front rather than stopping for review by default. The multi-model caveat also travels — a skill shared across contributors is written for a reader roster, not a model.
• Era-bound: the model roster and its behaviors (Astra stops earlier than Sol; Astra tests unprompted) are unverified vendor assertions about a days-old release, with no measurements. The description-shortening mechanism is Codex’s; other harnesses cap the catalog differently.
• Weaknesses: the audit question is framed as “is the model smarter now?”, which conflates two kinds of instruction. Procedure (read these first, run the tests) decays with model capability; environment facts (which hooks merge commits bypass, what a subagent inherits, which commands need a tty) do not, because no amount of judgment recovers what the model cannot observe. Every instruction the post proposes retiring is of the first kind, and its closing advice to have the model audit and prune invites a wholesale rewrite of the very surface zhang2025 shows collapses under wholesale rewrites — the consensus loop on this shelf (Self-evolving agents) wants evidence-first, itemized subtractions, not a release-day sweep. The “decision boundaries” section also reads as tuning for autonomy without distinguishing gates that existed for a weak model from gates that exist because a wrong call is irreversible.
• Author’s stake: the author works on Codex’s developer experience; the post is de facto release guidance and a reframing of a behavior regression (earlier stopping) as user-side configuration.
• In this library: the description rule was already written in this repository’s create-skill skill and had drifted anyway — six of eight skill descriptions restated their bodies’ sections when audited on 2026-09-07 — so the durable lesson here was less the rule than that a rule with a checkable shape needs its check (hooks/lint.sh now caps description length). The ingest-post skill took the router shape the same day.