
Neat Freak
Inventory and sync agent memory, instructions, and skill paths across Claude Code, Codex, OpenClaw, and OpenCode from one reference.
Overview
neat-freak is an agent skill most often used in Build (also Operate, Validate) that documents where memory, instructions, and skills live on Claude Code, Codex, OpenClaw, and OpenCode for a reliable first-path inventory.
Install
npx skills add https://github.com/kkkkhazix/khazix-skills --skill neat-freakWhat is this skill?
- Path tables for Claude Code (~/.claude/projects memory, CLAUDE.md, ~/.claude/skills), Codex (AGENTS.md, AGENTS.override.
- Explains Claude memory YAML frontmatter fields: name, description, type (user / feedback / project / reference)
- Notes Codex and OpenClaw lack separate memory-index files—project facts belong in AGENTS.md or equivalent root markdown
- OpenClaw gating via metadata.openclaw (OS, env, binaries) called out as optional, not required for inventory
- Designed for step-one盘點 (inventory) before deeper config sync or cleanup
- Covers 4 agent platforms: Claude Code, OpenAI Codex, OpenClaw, OpenCode
- OpenClaw documents 7-tier skill loading priority (workspace through extra dirs)
- Claude memory index uses YAML frontmatter: name, description, type
Adoption & trust: 1.4k installs on skills.sh; 14.1k GitHub stars; 3/3 security scanners passed (skills.sh audits); trending (+100% hot-view momentum).
What problem does it solve?
You run multiple coding agents but do not know which directory owns memory, project rules, or skills—and risk writing context to the wrong file.
Who is it for?
Solo builders switching or combining Claude Code, Codex, OpenClaw, or OpenCode on one repo who need a single盘點 reference before editing agent files.
Skip if: Teams with a fixed single-agent standard and no cross-platform paths to reconcile, or anyone expecting automated file moves, linting, or memory migration scripts.
When should I use this skill?
When performing the first-step inventory of agent memory, instructions, and skill paths for the platform you are using.
What do I get? / Deliverables
You get a platform-accurate map of memory and config paths so you can inventory, sync project facts, and install skills in the correct layer before deeper agent work.
- Mental or written inventory of correct memory, instruction, and skill paths per platform
- Aligned choice of which file holds cross-session project facts (MEMORY.md vs AGENTS.md)
Recommended Skills
Journey fit
Spans multiple journey phases - primary shelf plus alternate fits below.
Solo builders first need this when wiring agent tooling—knowing where skills and memory live before coding with an agent. Maps platform-specific skill dirs, MEMORY.md/AGENTS.md, and frontmatter conventions—the core agent-tooling setup facet on Prism.
Where it fits
Before installing repo skills, confirm whether they belong in ~/.claude/skills, .codex/skills, or .openclaw/skills so load order matches intent.
Spinning up a quick Codex prototype, lookup AGENTS.md and TEAM_GUIDE.md fallbacks so project facts land in the file Codex actually reads.
After agent CLI updates, re-run the path tables to see if memory encoding or OpenClaw priority rules shifted before merging configs.
Pre-release check that production repo roots expose the correct CLAUDE.md or AGENTS.md layers reviewers and CI agents will follow.
How it compares
Reference cheat sheet for paths—not an MCP integration and not a generic productivity planner.
Common Questions / FAQ
Who is neat-freak for?
Indie and solo developers who use one or more agent CLIs and need to find memory, AGENTS.md/CLAUDE.md, and skill directories without reading four separate doc sets.
When should I use neat-freak?
Use it during Build agent-tooling setup, before Validate prototypes with a new agent, when Operate/iterate on drifted configs, or any time you start step-one inventory across Claude Code, Codex, OpenClaw, or OpenCode.
Is neat-freak safe to install?
It is read-only reference material; review the Security Audits panel on this Prism page and skim the skill files in your repo before granting broader agent permissions.
SKILL.md
READMESKILL.md - Neat Freak
# Agent 记忆与配置路径速查 不同 agent 平台的记忆系统和项目配置文件位置不一样。执行第一步盘点时按你正在使用的平台查这张表。 ## Claude Code | 用途 | 路径 | |---|---| | 跨会话记忆(全局) | `~/.claude/projects/<encoded-project-path>/memory/` | | 记忆索引文件 | `~/.claude/projects/<...>/memory/MEMORY.md` | | 全局指令 | `~/.claude/CLAUDE.md` | | 项目级指令 | 项目根 `CLAUDE.md`(可层级嵌套) | | Skills 目录 | `~/.claude/skills/<name>/SKILL.md` | 记忆文件用 YAML frontmatter:`name`、`description`、`type`(user / feedback / project / reference)。 ## OpenAI Codex | 用途 | 路径 | |---|---| | 跨会话指令(全局) | `~/.codex/AGENTS.md` 或 `$CODEX_HOME/AGENTS.md` | | 项目级指令 | 项目根 `AGENTS.md`(可层级嵌套) | | 项目级 override | `AGENTS.override.md`(若存在,覆盖同目录 AGENTS.md) | | Skills 目录 | `~/.codex/skills/<name>/SKILL.md` 或项目内 `.codex/skills/<name>/` | Codex 没有独立的"记忆文件 + 索引"机制,所有跨会话信息都直接写在 `AGENTS.md` 里。同步时把"项目事实"那部分内容统一放 AGENTS.md。 发现项目里有 `TEAM_GUIDE.md` 或 `.agents.md` 也要看——这是 Codex 的 fallback 文件名。 ## OpenClaw | 用途 | 路径 | |---|---| | 用户级 skills | `~/.openclaw/skills/<name>/SKILL.md`(首次运行自动创建) | | 项目级 skills | `.openclaw/skills/<name>/SKILL.md`(仓库根目录下) | | Workspace skills | 当前 workspace 的 `skills/` 目录 | **加载优先级**:workspace > project-agent > personal-agent > managed/local > bundled > extra dirs。同名 skill 高优先级覆盖低优先级。 OpenClaw 没有独立的"记忆文件 + 索引"机制,跨会话信息可放在项目根的 markdown(CLAUDE.md / AGENTS.md / 等价文件)里,参照 Codex 的做法。frontmatter 支持 `metadata.openclaw` 字段做加载时的 gating(按 OS、环境变量、二进制依赖筛选),但不是 neat-freak 必需的。 ## OpenCode | 用途 | 路径 | |---|---| | 全局配置 | `~/.config/opencode/` | | 项目配置 | `.opencode/` | | Skills 目录(项目) | `.opencode/skills/`、`.claude/skills/`、`.codex/skills/` 都会被扫描 | | Skills 目录(全局) | `~/.config/opencode/skills/`、`~/.claude/skills/`、`~/.codex/skills/` | OpenCode 同时读取 Claude Code 和 Codex 的目录,所以同一个 skill 装在 `~/.claude/skills/` 下的话三家都能识别。OpenClaw 走自己的 `~/.openclaw/skills/`,需要单独装一份(或用符号链接)。 ## 如果当前 agent 没有独立记忆系统 跳过"记忆"那一层,把功夫全花在: - 项目根 markdown(CLAUDE.md / AGENTS.md / 本平台等价文件) - README.md - docs/ 仍然是有效的同步——记忆是锦上添花,docs 才是项目知识的最低保障。 ## 跨平台共存策略 如果一个项目同时被 Claude Code 用户和 Codex 用户使用,推荐: - **项目根同时放 `CLAUDE.md` 和 `AGENTS.md`**,内容可以互相 symlink 或在两边维护 - 或者一份内容主文件 + 另一份用一行 `See CLAUDE.md` 跳转 - docs/ 和 README 是平台中立的,不需要分两份 # 变更影响矩阵 遇到不确定"这次改动要同步哪些文件"时查这张表。**两个方向都要查**:补漏(加到哪些文件)+ 防膨胀(应该从哪些文件删)。 ## 反向:哪些信息该从 CLAUDE.md / 记忆里删除 CLAUDE.md / AGENTS.md 不是变更日志。下面这些反模式发现了就删 / 迁: | 反模式 | 处理 | |---|---| | "X 时刻起 Y 功能上线,详见 docs/Z.md" 形式的 blockquote | 删除——指针角色已经被「深入文档」指针表占掉,叙事归 git log / `/changelog` / `docs/CHANGES.md` | | 在 CLAUDE.md 里抄 docs/ 已有的详细机制 / 数据流 / 评分公式 | 删除——AI 改到这块自然会读 docs,CLAUDE.md 只留"边界规则" | | 已经稳定 ≥ 7 天的"新功能上线"叙事 | 该融入项目概览的融入;纯历史的删 | | 一次性事故的复盘细节("X 时 Y 服务挂了 30min 因为 Z") | 留 1 行红线规则("不要再裸跑 systemctl stop X"),事故详情归 docs/PLAYBOOK.md 或删 | | 已被新版本取代的"中间态"叙事("5/6 改了 X,5/8 又改成 Y") | 只留最终态规则;中间历史删 | | 单条 memory > 100 行 + 全是事故复盘 | 提炼成一条 ≤ 30 行的"规则 + Why + How to apply";多余的删 | | 记忆条目里"已被 X 取代" / "已废弃" / "保留作历史" 字样 | 99% 真的可以删,docs 已经是权威 | 判断标准:**这条信息在下次 AI 写代码时如果没看到,会犯错吗?** 不会就删 / 迁。 ## 代码层变更 → 文档层变更 | 本次对话发生的事 | 要改的文件(按受众) | |---|---| | 新增 API / 路由 | 项目根 markdown 路由清单 · `docs/integration-guide.md` API 速查表 · `docs/architecture.md` Routes 小节 | | 新增 / 改名 环境变量 | 项目根 markdown 环境变量表 · `docs/operator-runbook.md` 环境变量章节 · `docs/integration-guide.md`(如果下游要配) | | 新增数据库表 / 列 | 项目根 markdown 数据库表 · `docs/architecture.md` Data Model | | 新增 / 改动 用户流程 | 项目根 markdown 用户流程 · README 相关命令行示例 · `docs/handoff.md` What Exists Today | | 新增大特性(能跨多文件) | 以上全部 + `docs/architecture.md` 新增章节 + `docs/handoff.md` 已完成清单 | | 新增术语 / 改命名 | `docs/integration-guide.md` 术语表(如果有)+ 全局搜索旧术语替换 | | 部署参数 / 基础设施变化 | `docs/operator-runbook.md` · 项目根 markdown 部署章节 | | 下游项目接入方式变化 | 下游项目的 `docs/<integration>.md` · 上游项目的 `integration-guide.md` | ## 记忆层变更 | 情况 | 处理方式 | |---|---| | 过期事实 | 改记忆文件,同时更新索引(如 MEMORY.md)的 description | | 相对时间("今天"、"最近") | 全部转成绝对日期(`2026-04-29` 而非"今天") | | 重复记录(多条说同一件事) | 合并为一条,改索引 | | 已完成的待办 | 删除——知识库不是历史档案 | | 推翻的决策 | 删除旧条目,留新决策 | | 跨会话只用一次的临时上下文 | 删除 | ## 跨项目影响检查 最容易漏改的场景: - **上游 API 变了 → 下游 SDK 文档**:协议变化必须两边对齐 - **共享子域 / 路由 / 环境变量改了 → 所有 consumer 项目的 setup 文档** - **认证中台变更 → 所有接入应