# Codex skills
Reusable Codex skills by Matija Grcic.
| Skill | Purpose |
| --- | --- |
| [steer-work](skills/steer-work/SKILL.md) | Carry sustained implementation and research through a persistent manager, runners, written steers, and scheduled checks. |
| [name-tasks](skills/name-tasks/SKILL.md) | Name or rename Codex chats with accurate, consistent titles that are easy to scan. |
**Install**
Ask Codex to install either skill from its folder:
```text
Use $skill-installer to install steer-work from
https://github.com/matijagrcic/codex-skills/tree/main/skills/steer-work
```
```text
Use $skill-installer to install name-tasks from
https://github.com/matijagrcic/codex-skills/tree/main/skills/name-tasks
```
Each skill works independently and will be available on the next turn after installation.
`steer-work` was previously named `manage-project`. The [old entry](skills/manage-project/SKILL.md) remains as a compatibility pointer for existing prompts and shared links; new installations and invocations should use `steer-work`.
**Use steer-work**
Start in the workspace where the work belongs. Give the agent an outcome and a way to verify it:
```text
Use $steer-work to carry this work through to verified completion.
Outcome: [desired result]
Constraints: [requirements and exclusions]
Done when: [observable acceptance criteria]
I authorize you to create and message separate runner chats for this
project, create isolated worktrees when useful, and create a scheduled
heartbeat in this manager chat every 15 minutes.
Keep at most two runners active in addition to the manager. Record my
steers and reconcile them with the full brief before dispatching work.
Verify results, keep unchanged monitoring quiet, and pause the
heartbeat when finished or when all progress needs my input.
```
The manager keeps a small durable record, delegates independent work, reviews evidence, and integrates results. The defaults are one manager, at most two active runners, and a 15-minute heartbeat; adjust them in your request.
This workflow uses ordinary chats and a saved automation rather than activating Goal mode. It requires a Codex environment with chat coordination and same-chat scheduling tools. Local scheduled work requires the host and desktop app to remain running. If a capability is unavailable, the skill reports the limitation instead of claiming background work is active.
The per-session subagent concurrency setting does not establish a global cap on independent runner chats. The manager enforces the requested runner limit. See the [optional subagent configuration](skills/steer-work/references/codex-runtime.md#optional-subagent-concurrency-setting) for Matija's clarification of the two-subagent setting.
If installed, `communicate-clearly` supplies additional handoff and reporting guidance. Essential communication principles are included in `steer-work`, so the companion is optional.
**Use name-tasks**
`name-tasks` gives chats a clear action and specific target, such as `Investigating checkout timeouts` or `Verifying checkout retries`. It follows Shreya Shankar's [article on readable AI text at scale](https://fsdatalab.github.io/blog/agent-text-at-scale/): consistent structure and vocabulary, distinctive words early, and variation that communicates real differences. Completion wording requires evidence.
Use it for the current chat, selected chats, or a project's titles:
```text
Use $name-tasks to rename the tasks in [project name]. Read enough context
to name each task accurately, keep the titles consistent and easy to scan,
and verify the saved changes.
Report any tasks you could not inspect or rename.
```
For suggestions without changes, ask: `Use $name-tasks to propose a clearer title for this chat.` Naming does not start runners or a heartbeat. The full workflow lives in [name-tasks](skills/name-tasks/SKILL.md); the earlier task-naming reference remains as a compatibility link.
See [sources and design](docs/sources.md) for the evidence behind the workflow. No claim is made that this approach universally outperforms Goals.
**License**
[MIT](LICENSE). Attribution for adapted delegation guidance is preserved in [third-party notices](THIRD_PARTY_NOTICES.md).