In recent years, many development teams have built up their own prompt and skill collections: text modules that explain to an AI assistant how work is done in the project, which files it should read in advance and which steps it must never carry out without checking first. These libraries have grown along with the projects, mostly without anyone ever shrinking them again. That is now becoming a problem.

What OpenAI is observing

Eric Provencher, a developer at OpenAI, points out that extensive skill descriptions, blanket reading requirements and rigid approval rules can slow down the GPT-6 Astra model in Codex. The background: more capable models need less granular guidance than their predecessors. What used to be necessary guidance is now becoming dead weight.

The recommendation is accordingly sober. Instructions should be shortened on a task-specific basis instead of applying generally to all cases. And instead of prescribing every work step, teams should define clear completion criteria – that is, specify when a task counts as done.

Why this is plausible

Prompt libraries arise reactively. An assistant forgets a convention, so a rule is added. It changes a file it should not touch, so a list of prohibitions is added. It does not read the context, so reading is made mandatory. Each individual addition made sense at the time.

The result is a rulebook tailored to the weaknesses of a particular model version. When the model changes, the guidance stays. It then describes problems that no longer exist and at the same time narrows the scope for solution paths that the new model would find on its own.

Instructions that become outdated particularly quickly

  • Blanket reading requirements: Instructions to read certain files or documentation in full before every task, regardless of whether they fit the task.
  • Step-by-step choreography: Detailed process specifications that prescribe the order to the model instead of describing the goal.
  • Rigid approval rules: Requirements to check back that apply to every single action and fragment the workflow.
  • Redundant explanations: Detailed descriptions of concepts that the model mastered long ago.

Not every one of these specifications is superfluous. Approval requirements for database migrations, deployments or changes to production configurations have their reason and should remain. The question is whether a rule safeguards a real risk or merely compensates for an old model weakness.

Completion criteria instead of work instructions

The more interesting part of the guidance is the shift to completion criteria. Instead of describing how work should be done, it describes what the result must look like: Which tests must be green? Which checks must pass? Which documentation must be updated? Which interfaces must not have changed?

This approach is more robust against model changes. A criterion such as "the existing test suite runs without errors" remains valid no matter which model does the work. An instruction such as "first read these four files" is tied to a specific model version.

Incidentally, the shift forces teams to clarify something useful: anyone who wants to formulate completion criteria must know what quality means in their own project. That question is worth asking even when no AI assistant is involved.

A maintenance topic, not a one-off project

Prompt and skill libraries are part of the project infrastructure, comparable to build configurations or linter rules. They need maintenance, and not only in the form of additions. A pragmatic rhythm would be to review the collection whenever a new model enters the workflow.

A simple approach helps here: each rule is asked which concrete problem it once solved. Does that problem still exist? If nobody on the team knows the answer, that is an indication that the rule can be removed as a test. Versioning makes such experiments low-risk, because every change can be reverted.

It is also important that the effect remains measurable. Anyone who does not track how much rework an assistant causes or how often suggestions fail in review cannot judge whether a cut has helped or harmed.

What this means for web and software projects

Anyone who has firmly integrated AI assistants into their development processes should treat the accompanying instructions as a part of the project that requires maintenance – with regular review and the willingness to delete rules again. The effort involved is small, the effect on speed and result quality can be significant. And the shift from work instructions to clear completion criteria pays off twice: it makes collaboration with assistants more stable and at the same time sharpens the team's own understanding of quality.

Sources