The community debate: the stronger the model, the less you need skills?

The trigger was OpenAI's official developer blog post Rethinking skills and prompts for GPT-6 Astra (Eric Provencher, 2026-09-11), aimed at Codex users. Core argument: as models get more capable, the "scaffolding-style" instructions accumulated over the past year in AGENTS.md, skills, and prompts are not just redundant — they slow the new model down, since every file consumes context and the instructions begin to contradict one another. The advice: keep skill descriptions short and precise, use progressive disclosure to push detail into sub-files, and replace rigid step-by-step recipes with room for the model to judge.

On the "drop skills" sentiment. It has a rational kernel: "generic" skills that paper over model gaps (e.g. Superpower-style packs) do get absorbed into the model as it improves and stop being useful. But OpenAI's official post does not say "delete all skills" — it says "slim down and rebuild". What falls by the wayside is the batch of patch-the-old-model skills; high-value workflows and domain knowledge still earn their keep.

Around the same time, Anthropic also announced trimming Claude Code's system prompt sharply (community roundups of the announcement put the cut above 80%). Two frontier labs gave the same signal in the same month: instructions are shrinking, and skills are converging toward "short, precise, modular" — which is not the same as "skills are dead."