GPT-6 Astra Best Practices: Sharing AGENTS.md and Skills with Fable 5.1

Introduction

OpenAI recommends auditing the skills and AGENTS.md files that GPT-6 Astra reads. Its official guide describes a model that follows instructions closely but may pause for clarification when the user expects it to continue. It includes prompts for adjusting autonomy and making instruction priorities explicit.

Reading that guidance reminded me of Anthropic's recommendations for the Claude 5 generation, which I covered previously. Anthropic encouraged developers to remove excessive constraints and let the model exercise judgment.

The advice can initially seem contradictory, but OpenAI also recommends keeping AGENTS.md short. This article introduces Astra’s best practices and considers how to share instructions with Fable 5.1. The practical question is what to keep in persistent guidance and what to load only when needed.

AGENTS.md and CLAUDE.md are instruction files used to communicate project practices to coding agents. Skills package instructions and resources for particular tasks, such as reviewing code or writing an article. This discussion assumes that the same instruction text reaches both models; how each agent discovers and loads those files is a separate concern.

How do Astra’s best practices fit with Fable 5.1?

Both vendors favor concise instructions with less ambiguity and fewer unnecessary constraints. It helps to distinguish Astra-specific behavior guidance from advice about organizing instruction files.

OpenAI’s guidance for Astra

OpenAI's Prompting best practices for GPT-6 Astra recommends auditing instructions, defining autonomy, and resolving conflicts between explicit user requests and skill guidelines. The documentation also suggests asking the model to identify the specific skill instruction responsible when it pauses. This account reflects the documentation as checked on September 16, 2026.

A useful detail is that approval should concern a concrete result: the agent should first finish the preparation it is already authorized to do. The guide also addresses writing style, delegation to subagents, and testing proportional to the change.

OpenAI’s general Codex guidance recommends keeping AGENTS.md short and accurate, with task-specific details in separate documents. Its Harness engineering article describes using a roughly 100-line AGENTS.md as a map to deeper documentation. That is a case study, not a universal line limit.

Where Fable 5.1’s guidance aligns

Anthropic's earlier July 24 article, The new rules of context engineering for Claude 5 generation models, reports removing over 80% of Claude Code's system prompt for models including Opus 5 and Fable 5, without measurable loss on its coding evaluations. That is a result for Anthropic's system prompt, not a target reduction for every user's instructions.

The Fable 5.1 prompting guide also recommends completing already-requested work and making the scope of the deliverable explicit. For shared instructions, I would express when to proceed and when to ask, rather than issue a blanket prohibition on questions.

The Claude Code documentation also recommends targeting under 200 lines per CLAUDE.md file. This is general guidance rather than a model-specific rule or a truncation threshold. Keep material needed in every session there, and move task-specific procedures into skills or appropriately scoped rules.

How should we write instructions both models can share?

I would keep shared guidance concise, make outcomes and decision criteria explicit, and state when to consult detailed material. Within those constraints, the agent can choose its execution methods.

When a correction becomes too broad

The following is my interpretation of the two vendors' guidance, not a reported experiment comparing the proposed instructions.

Suppose frequent clarification requests lead someone to add this sentence to a shared skill:

Do not ask the user questions. Work autonomously until the task is finished.

The sentence makes no distinction between a routine implementation choice and an unresolved product requirement. Either model could read it as a reason to fill in missing requirements and continue. Even if it reduces clarification requests in one test, that does not establish that it is a suitable shared rule.

Removing approval conditions to give the model more freedom creates a different problem. Whether the agent may change customer-facing behavior or send material to an external service depends on what the user authorized. Greater model capability cannot supply that authorization.

My earlier article, in Japanese, described an incident in which Opus 5 treated provisional issue notes as settled requirements and prioritized them over my correction. That was an observation from my own setup, not evidence of a universal model trait. It did expose a problem in the material I supplied: the status and priority of the written requirements were not sufficiently clear.

The following examples apply this approach to clarification, specification changes, and verification. They are proposed wording, not official vendor templates.

When to proceed and when to ask

Instead of prohibiting questions, identify the choices the agent may make. For an input-validation bug, it can choose the order of investigation. Changing which inputs the product accepts may require a product decision beyond the repair request.

Proceed within the requested scope when existing requirements provide enough
information to decide. Choose the investigation order and implementation
approaches that can readily be revised.
 
If the work requires an unapproved change to accepted inputs or user-visible
behavior, explain why and present the options. While waiting for an answer,
continue investigation and verification that do not depend on that decision.

This wording gives both models the same scope of work. It makes the need for clarification depend on the request and the effect of the change, rather than on how strongly we urge the model to be autonomous.

When the conversation conflicts with a specification

Prioritizing the latest message without qualification is also insufficient. A user may ask about an alternative without deciding to adopt it. The agent needs to distinguish a change instruction from an exploratory question.

Distinguish confirmed requirements, provisional proposals, and reference
material in specifications and issues.
 
An explicit change instruction from the current user takes precedence over
an earlier written requirement. A question or proposal alone does not change
the requirement.
 
After applying an explicit change, report documents that remain inconsistent.
If it is unclear whether a change was requested and the answer would alter
the deliverable, ask for clarification.

This defines how to reconcile a request with project documentation. It does not override permissions enforced by the environment or mandatory organizational controls. The interpretation of a document and the authority to perform an operation remain separate.

When to stop verifying

Verification benefits from both entry and exit conditions. A blanket exemption for small changes could suppress a necessary check: a one-line authentication change and a spelling correction have very different consequences.

Choose verification based on the behavior changed and the consequences of
failure. Complete the checks required by the project.
 
Once the completion criteria are met, report the result unless a new failure
or an unresolved concern justifies additional verification.
Identify anything that could not be checked as unverified.

The same distinction applies to procedures inside skills. For an article-writing skill, source verification and a required article format can be mandatory, while the order of research remains discretionary. A model should not have to guess which steps are obligations and which can be replaced by an equivalent method.

AGENTS.md can hold shared decision criteria and directions such as “Read the article-writing skill when preparing an article.” Each skill can define its deliverable, completion criteria, and mandatory steps, without making unrelated work carry the full procedure.

CLAUDE.md’s @ imports load their targets at startup as well. Splitting a file just to get below 200 lines will not reduce context use if all the pieces are still loaded every time.

After revising the instructions, give both models the same files and requests. Include routine work, provisional requirements, and a task whose final publication is not yet approved.

Check whether the agents sought necessary clarification, stayed within their authority, and stopped at the appropriate point—not just whether they produced an artifact. If only one model still has a problem, identify the specific situation before adding a correction for it.

FAQ

Q. Will the same instructions make both models behave identically?

A. No; shared wording does not guarantee identical behavior. The proposed instructions still need to be checked with the same tasks on both models.

Q. Should we remove constraints and prohibitions?

A. Keep constraints the work requires, and state when and why they apply. Distinguish those obligations from commands added solely to counter a particular model’s tendency.

Q. Should a shared file stay under 100 lines?

A. OpenAI’s roughly 100-line example is a case study, not a universal limit. Focus on separating material needed in every session from details needed only for specific tasks.

Summary

Keep persistent instructions concise and the remaining rules explicit; say when to read detailed material. Share outcomes, constraints, and decision criteria, while leaving discretion over execution methods.

Before adding or deleting instructions, I would review each rule for the decision it is meant to govern. That is the approach I would use at ZenChAIne when giving Astra and Fable 5.1 the same AGENTS.md and skills.

References

• Prompting Claude Fable 5.1 — Anthropic
• Codex Best practices — OpenAI
• Harness engineering — OpenAI
• How Claude remembers your project — Anthropic
• Using GPT-6 Astra: Prompting best practices — OpenAI
• The new rules of context engineering for Claude 5 generation models — Anthropic, July 24, 2026
• Earlier ZenChAIne article on Opus 5 and instruction files, in Japanese