Field notes, issue 1

This week we tested one small skill, explain-error, across three hosts. It loaded and ran correctly on Codex; Claude Code accepted the files but we couldn’t get a live prompt to run against it; Cursor stayed untested because we don’t have it installed. This is issue one: we build with AI agents in our day jobs and write up what actually works, here, in public.

Tested this week

We built one deliberately small skill, explain-error (read a pasted traceback, name the file and line, name the error and cause, give exactly one fix), and hand-copied the identical folder into Codex CLI (both its current .agents/skills/ path and the older .codex/skills/ path) and Claude Code (.claude/skills/). We separately confirmed the third-party npx skills installer for Codex and Claude Code, which writes to .agents/skills/ and .claude/skills/ respectively; it doesn’t target .codex/skills/.

Codex is the only host where we could check live loading this week; helps stays not measured anywhere, Codex included. On codex-cli 0.157.1, both project paths discovered the skill, and a matching prompt made Codex read SKILL.md and references/checklist.md (visible in the tool trace, not just claimed in the answer) before opening with the expected marker line, naming the file and line, and giving one fix. An unrelated prompt triggered none of that. Claude Code accepted the same files at the documented path (file placement checked, Claude Code 2.1.283), but we could not run a live prompt against it; see below. We don’t have Cursor installed, so every Cursor claim about this skill stays “documented, not tested” for now.

A change that matters

Claude Code v2.1.277, released September 18, 2026, added direct AGENTS.md fallback: in a project with no CLAUDE.md, Claude Code now reads AGENTS.md instead, and the choice is configurable under “Project instructions” in /config. We re-checked the release note directly before writing this.

This matters for anyone running the same repository through more than one agent. Before this release, Claude Code didn’t read AGENTS.md directly by default: a project without its own CLAUDE.md could still have other project-instruction routes, such as rules and imports, but it wasn’t picking up an AGENTS.md sitting there for Codex or Cursor to read. Mixed-agent repositories that assumed Claude Code was reading nothing from AGENTS.md should recheck that assumption on 2.1.277 and later, on the version actually installed. The feature is dated and versioned, not retroactive.

An honest failure

Two things didn’t go as planned.

First, a real bug we found on purpose. We renamed explain-error’s references/checklist.md to break the link SKILL.md points at, then re-ran the same prompt against Codex. Codex still opened with the marker line and gave a correct file, cause, and fix (because those instructions also live in SKILL.md itself), but its attempt to read the now-missing checklist file failed silently. The tool trace showed the failed read; the final answer gave no hint anything was missing. Someone watching only the answer, not the trace, would never know the reference broke. Our format checker, skills-ref validate, didn’t catch this either: it only checks frontmatter shape, not whether the files a skill points to actually exist.

Second, we couldn’t run Claude Code non-interactively in our test setup, so that check waits for next week. claude -p returned “Not logged in” on every attempt, even from inside an already-authenticated interactive session, and we traced it to a credential-isolation problem in our test setup rather than a Claude Code defect. We’re not going to script around another session’s keychain to force a result; we’ll run this properly, interactively, and report it honestly next issue.

Read the full test notes in install skills in Codex and where agent skills live.