Shopify CEO Tobi Lütke is considering disabling Claude Code on Shopify until it supports reading.AGENTS.md、.agents/skillsIt's a file.
He believes that Claude Code insists on reading.CLAUDE.mdWhen team members use different AI programming tools, it can lead to different project rules and operating instructions for developers in the same code repository.
As AI coding agents enter the software development process, more and more code repositories are beginning to add specifically AI-oriented documentation.
These documents typically write down project structures, coding specifications, test requirements, development processes, and recoverable skills to give AI agents the context of a relatively stable project before processing the code.
AGENTS.mdIt's becoming a more common arrangement. Tools such as OpenAI Codex, Cursor, and AMP already support related formats, and some projects will adopt them..agents/skillsMore detailed skills and operational specifications for directory organization.
Claude Code uses another set of systems, including:CLAUDE.mdIt's the same thing..claude/skillsIt's like a file and a directory.
When team members use multiple AI programming tools at the same time, two sets of configuration inputs can cause synchronous problems. In the same code repository, developers using Codex can read.AGENTS.mdIt's a great way to help developers who use Claude Code.CLAUDE.mdContent. If two files differ in a directory, the project rules obtained by AI will differ as well.
This has caused a lot of controversy.
The developer community was quick to deliver a set of ready-made projects.
One of the most common methods is to create soft links in projects:ln -s AGENTS.md CLAUDE.md
Another way is to use Claude Code's own document reference capabilities.CLAUDE.mdIt's written:@AGENTS.mdClaude Code uploads the general rules that the team has maintained, while maintaining the code.CLAUDE.mdIncludes exclusive configuration space for Claude.
Lütke responded by posting again in response to the constantly emerging solutions to the comment area.
"A lot of people tell me how to deal with this problem and I know all about it". He further explains that the agent and Claude profiles are recurring applications along the directory tree. In a large monorepo with thousands of developers, it is difficult to guarantee that each directory will always maintain two sets of complete files at the same time.
"We can fix it with automation", Lütke writes, "but it's a stupid complexity tax that shouldn't be borne by developers".
These methods are usually easier to implement for individual projects or smaller code warehouses.
Lütke's concerns focus on the maintenance link after the scale enlargement. A monorepo such as Shopify contains a large number of directories and multiple parallel development scenarios, where profiles need to keep up with continuous changes in directory structure, project boundaries, and team specification. Soft links, reference relationships, and synchronized scripts will also become long-term maintenance objects in engineering systems.
Anthropic: Making Claude Code more flexibleAfterwards, Claude Code team member Thariq also responded to Lütke. He said that he would continue to share the progress after the release of the relevant features.
At the same time, he explained Anthropic's previous product thinking.
The Claude Code team believes that different model families have their own behavioral characteristics, and that system prompt will significantly affect model performance.CLAUDE.mdClaude Code also configures different system prompts for different models.
This is in line with recent discussions at Anthropic about "context engineering". In July this year, Tariq shared his experience in context engineering on the Claude 5 generation model, saying that the team removed about 80% of the Claude Code system prompt in the latest model, while summarizing how to write system prompt, skills and.CLAUDE.md。
Faced with Lütke's compatibility demands, Tariq also acknowledged that multiple configurations would generate additional maintenance work, and said the team would continue to follow suit.
Current short-term plans still include:CLAUDE.mdIt also cited other Markdown documents, such as:@AGENTS.mdThe content has already been uploaded.
When asked if he intended to implement it through open source harness, Tariq also tweaked "open source" before Claude Code.
Have you ever encountered a similar problem? Welcome to the comment section.
The link is as follows:
