What's this about?
Responding publicly to feedback from Shopify's Tobi Lütke, an Anthropic engineer said Claude Code will become more hackable — including support for AGENTS.md and other system-prompt modifications. The stated reason for not allowing it before is that model families are not interchangeable and the system prompt materially affects performance.
What exactly is a "system prompt"?
It is the standing instruction block the model receives before you type anything — the tool's own description of how it should behave. Vendors tune it carefully against their specific model. Letting users replace it is a genuine trade-off: it gives you control, and it also means the vendor can no longer guarantee the behaviour they tested.
What's the community saying?
1,218 and 1,004 likes on the two tweets
- The main thread is portability: developers want one agent instruction file that works across tools rather than a different format per vendor, which is what AGENTS.md is trying to be.
- The counter-argument, made by Anthropic itself, is that a shared file format implies models are interchangeable when they measurably are not.
- It connects to Gergely Orosz's point below — users increasingly want to bring their own agent and have vendors expose interfaces rather than agents.