菜单
Customize AGENTS.md file(s) for either the user or a project, including guidance on project, subdirectory, and user-level instruction scope.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Planning-centric workflow for substantial work across software, analyses, artifacts, documentation, and configuration. Invoke manually with $work-cycle when alignment matters; not for straightforward one- or two-edit requests.
Create or update developer-facing and agent-facing project documentation, or run module-scoped codebase audits. Invoke manually with $documentation for codebase maps, module documentation, codebase review/audit workflows, and README.md or AGENTS.md work.
Documentation subskill for project hygiene maintenance. Use through $documentation when the user asks to maintain, clean up, health-check, refresh AGENTS.md/notebook/TODOs, inspect git branches/PRs/stashes, or review overall project state. Do not use for codebase audit, codebase review, release-check, or module-scoped bug-finding requests; those belong to module-codebase-review.
Manual-only workflow for codebase audit, codebase review, release-check, paper-freeze-check, and module-scoped multi-agent read-only bug/code-smell audits that produce a prioritized notebook report and TODO-ready findings. Use this instead of maintain-project for requests that ask to audit or review code behavior.
Ergonomic notebook-backed TODO capture for deferred work. Invoke manually with /defer when the user asks to defer, save, park, or add a TODO for later work while preserving the current conversation context.
Initialize a project workspace for Claude Code with project instructions, repository checks, notebook scaffolding, optional remotes, and optional O2/SLURM setup. Invoke manually with /init-project for first-time project setup, notebook bootstrap, project remotes, or O2/SLURM setup.
| name | agents-md |
| description | Customize AGENTS.md file(s) for either the user or a project, including guidance on project, subdirectory, and user-level instruction scope. |
Use this subskill when changing durable agent instructions in AGENTS.md files.
Codex can use one AGENTS.md per directory, including the project root. Agents working inside subdirectories automatically load and concatenate the relevant AGENTS.md files along that path.
Use project AGENTS.md files for stable context:
Subdirectory AGENTS.md files should contain context specifically relevant to understanding or modifying that part of the codebase. For typical scientific programming projects, extensive directory-specific prompt structure is usually unnecessary.
Do not use AGENTS.md as technical documentation; either the documentation becomes stale when the code changes, or every code change incurs the overhead cost of updating the documentation file.
Global Codex instructions come from two source files:
~/.codex/user/AGENTS.md: user-editable personal instructions.<config-repo>/codex/global/AGENTS.md: shared instructions from this config repo.These are rendered into ~/.codex/AGENTS.override.md, which is the active generated file and should not be edited directly:
${CODEX_HOME:-$HOME/.codex}/bin/config-agent-tool render-override --agent codex
During first-run setup, ${CODEX_HOME:-$HOME/.codex}/bin/config-agent-tool install --agent codex invokes the renderer with --first-run, which refuses to overwrite an existing unmanaged AGENTS.override.md.
Changes to the user's ~/.codex/user/AGENTS.md are picked up by Codex only after rerunning the generation command and starting a new session.
The user-level AGENTS.md file should contain user preferences or context about the user that ought to be applied across all of their projects. Keep it short. It is prepended to every Codex session across all projects, so unnecessary content has a cumulative cost.