| name | setup-routines |
| description | Configure Claude Code routines for automated PR review, codebase health, issue triage, and docs drift detection. Use when setting up recurring automation, GitHub-triggered workflows, or API-triggered tasks that run on Anthropic cloud infrastructure. |
Setup Routines
Configure Claude Code routines -- cloud-hosted auto sessions triggered by schedule, GitHub events, or API. Routines clone repo, run as full Claude Code sessions. Hooks + CLAUDE.md rules enforce auto.
How it works
Routine fires -> clones repo -> SessionStart hooks -> CLAUDE.md loads
-> routine prompt executes -> PostToolUse hooks enforce on every edit
-> Stop hooks run quality gates -> session ends
Enforcement model
Hooks = enforcement layer | routine prompts = task layer. Standards evolve
in repo (hooks + CLAUDE.md), routine prompts stay stable. Every routine
session runs the same PostToolUse/Stop gates as an interactive dev session,
so a routine cannot ship code that a developer could not ship locally.
vs. Sandcastle
Sandcastle = local background agent in a tmux pane. Routines = cloud-hosted
sessions triggered by schedule/webhook/API. Pick Sandcastle for long-running
local work you want to supervise; pick routines for recurring automation
that must survive your laptop closing.
Available templates
| Template | Trigger | What it does |
|---|
| pr-review | pull_request.opened | Reviews PR vs standards, posts inline comments |
| pr-feedback-resolve | pull_request.review_submitted | Reads unresolved threads, fixes code, replies, resolves |
| issue-triage | issues.opened | Explores codebase, classifies, labels, posts investigation |
| weekly-health | Schedule: weekly | Runs quality checks, measures drift, opens health report issue |
| docs-drift | Schedule: weekly | Detects stale docs from recent changes, opens fix PR or issue |
Setup
1. Prerequisites
- Claude Code with web access (claude.ai/code)
- GitHub connected (
/web-setup in CLI)
- Pro, Max, Team, or Enterprise plan
2. Pick routines
| If you have | Recommended routines |
|---|
| Any hooks installed | pr-review |
| resolve-pr-feedback skill | pr-feedback-resolve |
| triage skill | issue-triage |
| Quality gate hooks/scripts | weekly-health |
| REFERENCE.md or other docs | docs-drift |
3. Create via web (recommended)
- claude.ai/code/routines -> New routine
- Name (example "PR Review -- [repo name]")
- Paste template from
routines/*.md -- customize OWNER/REPO placeholders
- Pick repo + environment
- Add trigger (GitHub event | schedule | API)
- Check connectors -- drop unneeded
- Create
4. Create via CLI
/schedule daily codebase health check at 9am
CLI = scheduled routines only. GitHub/API triggers -> use web UI.
5. Customize prompts
Templates = start point. Customize:
- Project-specific checks: reference patterns hooks enforce
- Labels: match issue label taxonomy
- Scope boundaries: "only review
src/" or "skip generated files"
- Connector actions: "post summary to #engineering Slack"
See REFERENCE.md for customization examples + API trigger setup.
6. Test
Run once by hand before trusting triggers:
- Web: Run now on routine detail page
- CLI:
/schedule run
- Watch session live at returned URL
- Check output -- tweak prompt if wandered
See REFERENCE.md: routine-vs-sandcastle decision, enforcement model, trigger/API/customization setup, troubleshooting.
