Skill Guide: Turn Repeated Work Into Reusable Skills

The weekly review is itself a repeated task

In the workflow-library chapter, you set up a manual weekly review: scan recent conversations, collect repeated corrections, update your Skills. After a few rounds, you will notice that this review follows the same pattern every time. It has a defined input (your recent AI sessions), a defined output (a list of proposed Skill changes), and a procedure you repeat without much variation.

Your skills folder is plain text you already know how to back up

Your changelog can live inside each Skill file. Earlier in this guide, you learned to note what changed and why whenever you update a Skill. Backup is separate: because your skills folder is plain text, any backup method works. Cloud sync, your file manager's version history, or a git repository all give you a way to recover from a bad edit. The important thing is that one copy stays current, and you can undo any overnight proposal that makes output worse.

Schedule an overnight review that proposes Skill improvements

If your AI app supports scheduled tasks, use whatever the app calls that feature. The names and controls change quickly across tools, so look for the official scheduled-task, routine, or automation feature in the app you use, paste the review prompt, choose a cadence, and keep the first run supervised.

Setting up the scheduled review in Claude

In Claude, use the scheduled-task or routine feature if it is available on your plan. Create a local review task, name it skill-review, paste the prompt below into the instruction field, choose a daily or weekly cadence, and save. If the app labels or buttons differ, trust the current Claude help text over the wording in this guide.

The first time the task runs, supervise it. If it asks for permission to read files or run commands, allow only the folders and actions required for the review. Do not grant broad permanent access until the dry run produces a useful review without touching files.

Setting up the scheduled review in Codex

In Codex, use the Automations or scheduled-task feature if it is available in your app. Create a recurring task, paste the review prompt, choose a daily or weekly schedule, and keep the first run in a watched thread. If Codex exposes a conversational setup command, use it only after checking the generated schedule and workspace.

The review prompt

This prompt is a refined version of the weekly review template from the workflow-library chapter. It tells the AI what to look for and how to format the output. Paste it into the Instructions field of your scheduled task.

What the overnight review catches

Organize your library by scope so reviews can catch contamination

The overnight review works best when your folder structure matches the three tiers of scope introduced earlier: global Skills in the global folder, project Skills in their project folders, tool-specific Skills noted and placed accordingly. When the folders match the intended scope, the review can easily catch Skills that are leaking rules from one project into unrelated work.

If a Skill works identically in both Claude and Codex, keep it in one canonical location and copy it to the other tool's folder. The version-controlled repository is the source of truth. When the Skill updates, the update happens in one place.

The Skill file format works across tools

The SKILL.md format you have been using throughout this guide is an open standard. Anthropic originally developed it for Claude, then published it at agentskills.io as an open specification. Adoption is changing quickly, so treat the public spec as the stable promise: a Skill is a readable folder with a required SKILL.md file, name and description frontmatter, and optional scripts, references, and assets.

The portable part is the Skill file itself: the name and description at the top, the procedure body with gotchas, and the optional reference files. Installation paths differ by tool and may change, but a well-written Skill works across any tool that supports the format because the instruction package is just readable Markdown.

Portability has limits. Scripts, tool connections, permissions, product-specific steps, and install paths may need to be adapted for each tool. Keep the task procedure in SKILL.md, and keep tool-specific setup notes in a README or install note alongside the Skill.

Mini-project: schedule your first overnight review

• I created a scheduled review task in my AI tool's app.
• I ran the task manually once and reviewed the output.
• I applied at least one proposal by asking AI to update the relevant Skill.

The system that maintains itself

At the start of this guide, the problem was clear: modern AI tools remember preferences and past conversations, but memory alone does not handle repeatable procedures, review rules, and task-specific judgment.

You now have a library of Skills and foundation files, a tested trigger for each Skill, gotchas collected from real use, and a scheduled agent that scans your recent work and proposes improvements overnight. The corrections you make today become the instructions the AI follows tomorrow.

The tools will keep changing. Claude will update, Codex will evolve, new options will appear. Your skills are portable, self-improving, and independent of any single tool. That system outlasts any single vendor.