RoutineOSBeta
Learn

Claude skills in routines: put a skill on a schedule.

A Claude skill is a folder with a SKILL.md file that tells Claude how to do one kind of job, such as how your replies should read. A skill routine keeps that know-how in the skill and runs it on a schedule. This guide covers when a skill is the right home for the logic, how a scheduled run gets the skill, and what still sits outside it.

Get startedSkill routines in the docs
Free (MIT). Needs a paid Claude plan.from the docs
What it is

What a skill routine is

A skill is a Claude Code feature, not a RoutineOS one. Its SKILL.md has a short header that tells Claude when to use it, then plain instructions. A skill routine keeps judgment in that file: a house style, a rubric, a checklist, the shape of a good draft. The routine's own prompt stays short and names the skill at the step that needs it.

Choose a skill when the step is judgment, the guidance is longer than a few lines or used by more than one routine, and it changes on its own schedule. Do not choose one for a rule you edit in Settings, or for a step that must give the same answer every run. Those stay in code.

Source: Skill routines in the RoutineOS docs.

Delivery

Deliver a skill to a scheduled run: by repo or by account

A cloud run starts as a fresh session. It does not read the skills in your home folder, so the skill has to reach it one of two ways.

DeliveryWhere the skill livesHow we know
Repo (the default)Committed in the repo under .claude/skills/pilot-verified A repo skill loaded in a cloud run.
Account skillUploaded by you to your claude.ai account and switched ondocs-stated Not piloted.

RoutineOS never uploads a skill for you. For an account skill it gives you the steps and waits for you to say it is uploaded. A skill that is only in your own skills folder is not seen by a cloud run. The RoutineOS plugin does not run inside cloud scheduled runs either, so only skills committed to the repo travel with it.

What stays outside

What a skill does not replace

A skill changes how a step is written, not the rules around it. A skill can draft a send but cannot send it, so every send waits at a gate. Settings a person edits on the Desk are applied by code, not by the skill. Fetched mail and pages are data, not instructions. And the store is written by the same command as in any routine.

  • The prompt forbids sends that are not approved.
  • Pack validation fails a write connector that has no gate.
  • Lint refuses a skill description longer than 1,536 characters.

These are prompt and lint rules, not runtime guarantees. A shorter prompt that starts with a slash command to run the skill is unverified for a scheduled run, so do not rely on it.

Examples

Templates that ship skills

Inbox triage

Ships a reply-style skill for how drafts read. A reply is sent only after a person approves it.

unpiloted
Open

Content calendar

Ships a house-style skill for voice and source markers. It publishes only what a person approves.

unpiloted
Open

Only Jobs watch and Watch are piloted. Both skill templates have passed validation, not a real run. The plugins guide covers how the plugin differs from a skill.

Start from a skill template

See the content calendar, which ships a house-style skill and asks before it publishes.

Get startedSee content calendar