[DigitalToday reporter Chi-gyu Hwang] Eric Provencher (에릭 프로벤처), who oversees developer experience (DX) for OpenAI Codex, has drawn attention by stressing that the arrival of GPT-6 Astra underscores the need to rethink writing prompts for coding agents.

He said, "In the past, it took a lot of explanation and scaffolding to use a model in the direction you wanted, but that has changed now." He said that for higher-performing models such as GPT-6 Astra, guidelines that used to work can instead interfere with the task.

He cited skills saved as markdown files that spell out methods needed when using specific tasks or plugins. His message is that creating many skills does not make an agent smarter.

Provencher said, "Each skill’s name and description go into the model context, and if the description is too long or there are too many skills, (OpenAI’s coding agent) Codex forcibly shortens the descriptions. That makes it harder for the model to judge which skill to choose."

He also pointed to conflicts between skills as a problem. He said that as more descriptions try to steer the model to always select a particular skill, they can pull in instructions unrelated to the actual task. De

In his post, Provencher presented 3 principles worth referring to when using skills. The first is to keep descriptions short and clear, stating only when they should be used. The second is not to put everything in the body of a skill, but to provide only guidance to consult sub-documents when needed. That is because if the model has to read all the skill content, it reduces the spare space the model can use.

The third is not to write it like an overly detailed manual. He said models now understand ambiguous situations well, so instructions that are too specific can instead get in the way.

The same applies to an AGENTS.md file that contains instructions telling an AI coding agent how to work in a repository containing code. The file keeps running automatically each time work is done in the repository, and Provencher’s view is that it is necessary to ask again whether each instruction in AGENTS.md is really needed. He said it is overkill to tell the agent to read the entire repository just to fix a single typo.

Provencher said, "GPT-6 Astra figures out on its own what it needs to look at even without such instructions," adding, "Instructions to run tests and check the results used to be necessary, but now the model does it on its own, so the same instruction can instead lead to unnecessary tests."

He said the way permission boundaries are set also needs to be reviewed. He said that because models previously caused trouble by proceeding without permission, many cases included strong language telling them to "ask first," but GPT-6 Astra’s judgment has improved so much that it can be trusted with the work.