Claude Code reads AGENTS.md now — and the harness is becoming moddable
Two things landed together in 2.1.277. AGENTS.md — the cross-tool instruction file Codex and others already read — works in any project without a CLAUDE.md, so one file can now steer several harnesses. The bigger part is how they shipped it: as a "mod," the first public piece of an upcoming way to customize the Claude Code harness yourself, with the source published. Project instructions stop being a fixed feature and become something you can replace. If you maintain instruction files across repos, this is the moment to decide which name you standardize on.
▸Do it — 4 steps in your repos10 min
- Update to ≥2.1.277. In a repo with both files, CLAUDE.md still wins — AGENTS.md is only read when CLAUDE.md is absent
- Decide per repo: keep CLAUDE.md, or move to AGENTS.md so Codex and other tools read the same file. Don't maintain both
- Skim the mods directory — project instructions are the first thing you can swap, and the pattern shows what else will be swappable
- Not yet on Bedrock, Vertex or Foundry; check before relying on it in a work repo
Ptacek: use an LLM as copyeditor, never as ghostwriter
Thomas Ptacek's two rules for keeping your own voice: never use a single specific turn of phrase a model suggests, and don't let it encourage you. His reasoning for the first is the useful part — frontier models are stuck writing magazine headlines, so every suggestion is locally pleasing and cumulatively makes the piece read as output rather than writing. The method is: write it yourself, then hand it over only to find flaws. You write at blog.tiffanyphd.dev in a voice you've deliberately built, which makes the rule worth adopting as a hard line rather than a preference.
↗Read — sockpuppet.org9 min
- Rule 1: no specific phrase a model suggests may survive into your draft — even if you like it, even if it's better
- Why: models write every sentence like a headline; dozens of headlines in a row is the tell readers detect
- Rule 2: turn off encouragement — praise is an influence campaign on what you keep
- The method is write → feed in → ask only for flaws, never for replacements
- He shows his own copyediting tool; Willison's proofreading prompt is the same shape
Why read anyway — The argument for rule one is what makes it stick; stated flat it just sounds puritanical. original ↗
Attackers are targeting maintainers through fake job calls
The crates.io security team confirmed an ongoing campaign against rust-lang members and popular crate owners: a plausible company with a real-looking LinkedIn sets up a video call about a job or contract, then gets the target to install a "missing audio codec" or paste a command from the clipboard — and uses their publishing rights to ship malware. The arrayref crate fell to this last month. The shift worth absorbing: the soft target in a supply chain isn't the code, it's the humans with publish access, and every dependency you ship carries that network of people. The mitigation people keep landing on is dependency cooldowns — wait a few days before upgrading, so someone else finds it first.