Analyzes this codebase and surfaces actionable Routine candidates across the three trigger types.
Usage: /routines-discover - no arguments needed. Run it in the root of any project.
A Routine is an autonomous Claude Code session running on Anthropic-managed cloud infrastructure, triggered in three ways:
| Trigger | How it fires |
|---|---|
| Schedule | Recurring cron cadence (min 1 hour) |
| API | HTTP POST to a per-routine endpoint with bearer token |
| GitHub events | Repository events: PR opened/merged, push, issue, workflow run, etc. |
Each run clones a fresh copy of the GitHub repository, runs a full Claude Code session with configured MCP connectors (Slack, Linear, GitHub, Google Drive…), and can create branches, open PRs, post messages, and call external APIs. No local machine required.
Daily limits: Pro 5/day · Max 15/day · Team/Enterprise 25/day.
Before generating any output, silently read:
README.md or CLAUDE.md to understand the project purpose and stackpackage.json, Cargo.toml, pyproject.toml, or equivalent for dependencies.github/workflows/ to understand existing CI/CD automationIf MCP connectors are configured in .claude/settings.json, note which external services are connected.
For each dimension, think concretely about this specific project before writing anything.
1. Scheduled maintenance What recurring work currently requires a human to run manually or remember to do? Think: dependency audits, stale issue/PR triage, test flakiness reports, coverage drift, TODO comment tracking, dead code detection, daily or weekly summaries.
2. Event-driven reactions What should happen automatically when a PR is opened, merged, or closed, but doesn't today because no one gets to it? Think: review checklists, changelog updates, cross-repo sync, Slack notifications with context, label enforcement, docs updates on API changes.
3. Alert and incident response What monitoring signals exist? When something breaks, what is the first thing a developer does? Think: correlating an alert with recent commits, triaging a failing build, drafting a postmortem skeleton, routing an error to the right team.
4. Cross-system sync What drifts today because the sync between two systems is manual? Think: keeping two SDKs in sync, updating a doc site when an API changes, syncing GitHub issues with Linear, keeping a README stats section current.
5. Release and deploy automation What steps happen before or after a deploy that a human runs by hand? Think: smoke tests, release notes, version bumps, stakeholder notifications, go/no-go summaries.
For each use case identified, produce a card in this format:
[Name] · schedule / api / github
Trigger: [what fires it: cron expression, which event, which external system]
Input: [what Claude receives: repo state, event payload, alert body]
Output: [what Claude produces: PR opened, message posted, file updated, issue created]
Value: [time saved or risk reduced, be specific]
Blockers: [missing connector, secrets needed, GitHub App required, etc., or "none"]
Sort cards by value-to-effort ratio, highest first.
After all cards, add a Quick Wins section: the two or three use cases that could be set up in under 15 minutes with the current repo and connector configuration.
Nightly stale PR report · schedule
Trigger: Every weekday at 8am
Input: Repo state: all open PRs older than 5 days
Output: Slack message to #engineering with list of stale PRs, assignee, and last activity
Value: Saves ~15min of manual triage each morning, reduces PR rot
Blockers: Requires Slack MCP connector