# Set Up potato-skill in Codex Source: https://www.potatoannotator.com/docs/getting-started/potato-skill-codex **Codex reads skills from `.agents/skills/` in your repository, and `npx skills add davidjurgens/potato-skill --agent codex` copies potato-skill there. Codex picks up the skill without a restart, and `$potato-skill` starts the skill in the CLI and the IDE extension.** [Quick Start with a Coding Agent](/docs/getting-started/quick-start-with-a-coding-agent) covers what the skill does with a study, and [Install potato-skill in Your Coding Agent](/docs/getting-started/install-potato-skill) covers the other agents. ## Before you install Potato goes in the Python environment that Codex's shell uses. The skill's helper scripts import Potato's registries and call the `potato` command, and the skill's first step checks that Potato imports. The browser extra and Chromium let `potato preview --screenshot` and the skill's walk-through script open the task in a headless browser: ```bash pip install 'potato-annotation[preview]' playwright install chromium ``` Codex's default sandbox has no network access. Install both before you start to save a round of approvals. The `skills` CLI runs through `npx` and needs Node.js 22.20 or later. ## Installing the skill Run the install in your repository: ```bash npx skills add davidjurgens/potato-skill --agent codex ``` The command copies `SKILL.md`, the 30 reference files, and the nine helper scripts into `.agents/skills/potato-skill/`, and records the source in `skills-lock.json` at the repository root. Codex scans `.agents/skills/` in every folder from the current one up to the repository root, so a session started in a subfolder of the repository finds the skill too. Committing `.agents/skills/` gives collaborators the skill with no install step. Add `-g` to install the skill for every repository on the machine instead: ```bash npx skills add davidjurgens/potato-skill --agent codex -g ``` A global install goes to `~/.agents/skills/potato-skill/`, the user-level skills folder Codex reads. Codex detects a new skill without a restart. If the skill does not appear, restart Codex. ## Checking the install In Codex CLI or the IDE extension, `/skills` lists the skills Codex found, and potato-skill appears in the list. In the ChatGPT desktop app, **Skills** in the sidebar shows the skills across your projects. ## Starting a study Mention the skill with `$potato-skill` and describe the study: ```text $potato-skill I want 200 tweets labeled for sentiment on a 5-point scale by three annotators each, with a consent page and two attention checks. Build it and start it. ``` Codex also loads the skill without a mention when a request matches the skill's description, such as a request to build an annotation task, edit a Potato `config.yaml`, or check on a running study. ## Sandbox and approvals Codex runs commands in a sandbox. The default mode, `workspace-write`, lets commands write inside the workspace with network access turned off. The setting behind the network is `sandbox_workspace_write.network_access`, which defaults to `false`. Under the `on-request` approval policy, Codex works inside the sandbox and asks before a command goes beyond the sandbox. During a build, the skill has Codex validate the config, start a Potato server on your machine, open the task in a headless browser, and run the skill's helper scripts. A `pip install` needs the network, and the default sandbox has none. Approve a command outside the sandbox when Codex asks, after reading the command. In the CLI, `/permissions` opens the permissions picker. In the IDE extension and the app, the same choices sit under the composer as **Ask for approval**, **Approve for me**, and **Full access**. Codex's [sandboxing documentation](https://learn.chatgpt.com/docs/sandboxing) describes each mode. ## AGENTS.md Codex reads `AGENTS.md` by filename. The potato-skill repository has an `AGENTS.md` at the repository root, a short summary of the skill's rules. Copying that file into your repository root gives Codex the rules with no install step. The file points at reference files and scripts that only a full install puts on disk. Use the `skills` CLI install for building a study. If your repository already has an `AGENTS.md`, add the summary to that file rather than replacing the file. ## Updating and removing the skill The `skills` CLI updates and removes the skill by name. Add `-g` to either command for a global install: ```bash npx skills update potato-skill npx skills remove potato-skill ``` `update` checks the installed copy against the GitHub repository and replaces the copy when the repository has changed. `remove` deletes `.agents/skills/potato-skill/` and the skill's entry in `skills-lock.json`. To turn the skill off without deleting the skill, add an entry to `~/.codex/config.toml` with the absolute path of the skill's `SKILL.md`, and restart Codex. Codex's [skills documentation](https://learn.chatgpt.com/docs/build-skills) lists the other settings an entry takes: ```toml [[skills.config]] path = "/path/to/your/repo/.agents/skills/potato-skill/SKILL.md" enabled = false ``` ## Further reading - [Quick Start with a Coding Agent](/docs/getting-started/quick-start-with-a-coding-agent) - [Set Up potato-skill in Claude Code](/docs/getting-started/potato-skill-claude-code) - [Set Up potato-skill in Cursor](/docs/getting-started/potato-skill-cursor) - [Configs your coding agent can check](/docs/guides/configs-your-coding-agent-can-check)