Anthropic gave up one of its own naming conventions this week, and that is the best news in the whole changelog.
Claude Code reads AGENTS.md when a project has no CLAUDE.md
Since version 2.1.277, released on Friday 18 September, Claude Code reads the AGENTS.md file as the project instruction – in a folder that holds no CLAUDE.md. AGENTS.md is the name AI agents from other vendors settled on – Codex from OpenAI reads it, for one. Until now Claude Code saw such a file only when someone added the import @AGENTS.md to CLAUDE.md. As an editorial team we are happier about this change than about many a louder one: Anthropic followed a standard somebody else came up with, instead of stubbornly holding on to a file name of its own. For you it means you write one set of AI instructions per folder, rather than a separate set for every tool you work with in that folder.
Which files get loaded is your choice in the /config panel, under Project instructions – and it pays to be careful there, because it is one file or the other, never both. A single CLAUDE.md one floor up the folder tree is enough to keep the AGENTS.md in your project from being loaded.
Importing files with the @path syntax works in AGENTS.md exactly as it does in CLAUDE.md, so an instruction spread over several files does not have to be rewritten into one. The import itself, though, is an addition of Claude Code’s: the shared description of the AGENTS.md format provides for no import at all, and the Codex documentation describes only the stacking of successive AGENTS.md files from the directory tree. So if your AGENTS.md is meant to work in other tools as well, keep the instructions in one file.
If your working folder holds neither of these files yet, start with the lesson Your first CLAUDE.md file – we show there what belongs in such an instruction and how to check that Claude Code really reads it.
Skills and plugins from your claude.ai account turn up in terminal sessions
Since version 2.1.273, released on Tuesday 15 September, skills and plugins behave much like MCP connectors: you switch them on once in your claude.ai account, and Claude Code sees them in the terminal without you setting up anything at your end. The same sync takes those skills to the other places your account works in – to Cowork sessions in the Claude app on your computer and to sessions in the cloud. A terminal session logged into that account downloads them at startup, in the background, into a folder of their own next to your own skills, and checks roughly every ten minutes whether anything changed in the account.
The route runs one way only – from claude.ai to the terminal. A skill you edit locally, in the downloaded folder, will not travel back to the account; Claude Code tells you outright when you save that the change has not been saved to your account. Two settings switch the whole sync off, syncClaudeAiSkills and syncClaudeAiPlugins – you only have to ask Claude Code for it.
Do not move everything you have in the terminal over to the account. A skill that runs commands with ! or pulls files in with @ stops doing so once it moves – in the terminal too, where it works today, because Claude Code reads those passages in an account skill as plain text. What suits the account are skills that are instructions, not the ones that reach for your files and your commands. Ask Claude Code which of your skills can make the move.
Keep in mind that every skill takes up part of the context in every turn of the conversation, so a dozen-odd skills from the account now take up room in the terminal as well. Exactly how much is what the /skill-doctor command shows – we wrote about it in the News of 7 September.
In the /config panel a click changes a setting’s value
Clicking in the terminal was possible before, just not everywhere: in selection lists – in a permission prompt or in /model – since version 2.1.187, and in questions with options to pick from since version 2.1.208. Version 2.1.271, released on Monday 14 September, adds one more step: in the /config panel, clicking a setting’s value changes it on the spot. Before that the cursor picked the row in that panel and the value was switched from the keyboard.
For someone who opened the terminal a week ago that is a sizeable difference: clicking is less daunting than arrowing down a list on which everything looks the same.
Worth remembering that cursor support works only in /tui fullscreen mode, and that it will not get along with every terminal.
The /config panel is worth a visit for more than the clicking. We go through the settings that really are worth changing in it in the tip Claude Code settings – what to change, what to leave.
To finish: a scientific paper you can talk to over MCP
On Wednesday 16 September Nature published a paper describing Paper2Agent – a tool that takes a scientific publication together with its code and its data, and then serves it as an MCP server. Instead of reading what the authors did and reproducing it at your end from the repository, you ask such a server for the result, and the server computes it by the methods the publication describes.
The authors showed this on papers from computational biology, and the Paper2Agent code sits publicly on GitHub.
If the idea catches on, the way we reach for other people’s research will change. Instead of reproducing the described method at your end, you ask the very code the publication was built on for the result – and suddenly the data and the methods are within reach of someone who would not know how to run that code. It looks to us like the start of a genuinely different way of getting at research results, though for now this is a single tool and papers from one field.
What an MCP server even is, and how it differs from a ready-made connector out of the catalogue, we explain in the tip Connector or MCP: what actually makes them different.