3-minute guide to centrally managing skills for Codex and Claude Code (with prompts included) When using multiple AI tools to write code, the same skill has to be maintained in several copies. If you modify one and forget to sync the others, you'll spend ages debugging only to find out the versions are inconsistent. Now the most troublesome thing is not that you don't know how to write skills, but that skills get scattered the more you use them. The same skill exists in one copy in Codex and another in Claude Code. At first, it's just a few more folders, seemingly trivial. The real headache comes later: you modify a skill once but forget to sync it to another tool. As a result, the same task behaves one way in Codex and another in Claude Code, and debugging takes ages because you can't figure out why. Initially, I created separate skill sets for each tool, thinking the content was roughly the same. Then one time I changed a prompt's logic but only updated the Claude Code copy — Codex was still reading the old version. It took me half an hour to realize the versions weren't synced. That's when I decided to maintain just one set. If each tool reads from its own source, skills become more and more scattered. This article solves exactly that: how to maintain only one set of skills and have both Codex and Claude Code read from that same set. 01｜Maintain only one set of skills This "AI job description" shouldn't be scattered across each tool's own folder. The recommended approach is to put skills in a unified folder, even a dedicated Git repository (think of it as a folder specialized for saving and tracking file versions), and then have the Codex and Claude Code entry points read from that one set. We use something called a symbolic link (symlink). Think of it as an entry point pointing to the source folder: when you click into it, you see the content, but the actual files are still in the source folder. It's not a copy, so if you modify the source file, all entry points will automatically see the updates. Here's a sanitized directory structure example: ~/ai-workspace/ shared-skills/ article-writing/ SKILL.md note-capture/ SKILL.md my-project/ .agents/ skills -> ../../shared-skills .claude/ skills -> ../../shared-skills shared-skills is the only place where the source files reside. .agents/skills and .claude/skills are just entry points pointing to it. When you modify something in shared-skills, all tools reading through their respective entry points get the latest version. 02｜Let the AI organize it for you You don't have to figure out which directories should be merged or which paths should be symlinked. Just give the following prompt to Codex or Claude Code, and let the AI first inventory your local structure, then present a migration plan: --- Please help me organize the AI tool skills management on my computer. I'm not a technical person, so please explain in simple terms, assuming I don't understand file systems, symlinks, or the command line. Goals: 1. Find all directories like .agents/skills, .claude/skills, or other possible locations for AI tool rules and skills in the current project. 2. Determine which of these directories contain the real source files that need maintenance and which are just copied duplicates. 3. Design a unified shared-skills directory for centralized skill maintenance. 4. Provide a symlink or unified referencing solution so that Codex and Claude Code both read the same instructions as much as possible. 5. Before outputting, list the commands you plan to execute and explain what each command does. 6. Do not delete, overwrite, or move any files directly; wait for my confirmation. 7. For all paths involving real usernames, private directories, or repository names, use sanitized paths in the summary. Please output in the following format: - Which skill or rule directories did you find - Which location is more suitable as the single source directory - Which locations are suitable to be changed to symlinks or unified references - What impact each step has - Commands that require my confirmation before execution - How to revert if something goes wrong --- The AI will first list what it found, what it suggests, and the impact of each step. It will wait for your confirmation before doing anything. If skills are originally spread across multiple Git repositories, the unified directory can also just serve as an aggregation entry point, symlinking to the source files in each repository. The unified directory doesn't have to store all skill source code. 03｜Also manage project rule files together Beyond skills, project rule files have the same problem. Some tools read AGENTS.md, some read CLAUDE.md. Essentially, they all answer the same question: How should the AI work in this project? If these files have basically the same content, there's no need to maintain multiple copies. Keep one main file and symlink the others to it: AGENTS.md CLAUDE.md -> AGENTS.md Again, you don't have to decide which files to merge and which to keep separate. Give this prompt to the AI: --- Please check the AI tool rule files in my current project. I'm not a technical person, so first explain in simple terms what each file might do, then give suggestions. Goals: 1. Find if there are AGENTS.md, CLAUDE.md, or other AI tool rule files in the project root. 2. Determine if these files have duplicate content, overlapping responsibilities, or only minor differences. 3. If they can share one set of content, suggest which one should be kept as the main file and which should be turned into symlinks or reference entry points. 4. If a specific tool truly needs its own rules, explain where the differences should go; do not force merge. 5. Output suggested commands and explain what each command does. 6. Do not overwrite, delete, or rename existing files directly; you must get my confirmation first. 7. Sanitize all local paths in the summary. Please output in the following format: - Rule files found - What each file roughly does - Whether they are suitable for merging - Suggested main file to keep - Files suggested for symlinks or reference entry points - Files that need backup before execution - Suggested commands and their functions --- Write project rules once, and all different AI tools can read the same standard. This reduces problems like "I changed the rules, but the other tool didn't apply them." 04｜When is this approach suitable? Suitable when: - You use multiple AI tools like Codex and Claude Code simultaneously - Multiple tools need to reuse the same set of skills or project rules - You frequently modify skills and don't want to manually copy them - You want to put skills in a dedicated Git repository - Multiple projects need to reuse the same workflow Not suitable when: - A particular project requires entirely independent skills - Team members are not familiar with symlinks and may accidentally delete the source directory - Windows environment doesn't have symlink permissions configured - The official documentation of a tool you use explicitly says it doesn't support symlinks (if unsure, ask the AI to check for you) - If you're not sure whether your situation is suitable, don't delete files yourself. Give the prompt above to the AI, let it propose a plan, and then confirm. 05｜A few reminders - Symlinks point to paths, not copy files. If you move the source directory, the symlink will break. - In team collaboration, prefer relative paths (e.g., ../../shared-skills) over absolute local paths (e.g., /Users/xxx/...). - If you manage the skills as a separate repository, remember to commit the actual source repository to Git, not just the symlink inside the project. #AI #AIAgent @nikitabier @XCreators @grok

~/ai-workspace/
  shared-skills/
    article-writing/
      SKILL.md
    note-capture/
      SKILL.md
  my-project/
    .agents/
      skills -> ../../shared-skills
    .claude/
      skills -> ../../shared-skills

AGENTS.md
CLAUDE.md -> AGENTS.md