Configuring a Routine — Prompts, Repositories, and Connectors
A Routine is defined by three components: the prompt, the repository it operates on, and the connectors it can use. Getting each one right is what separates Routines that run reliably for months from Routines that stall, overshoot, or produce noise.
Quick Reference
- →Routine prompts must be self-contained — no back-and-forth, no asking for clarification
- →Always include conditional branches: if X found → action A; if nothing found → brief confirmation
- →Repository access is granted via GitHub connector; branch safety is enforced automatically
- →Default branch constraint: Routines can only push to branches prefixed with `claude/`
- →Connector configuration is per-Routine — each Routine manages its own integrations
- →Prompt should specify what tool to use for output: Slack channel, GitHub issue, PR, or webhook
- →Include explicit verification logic — what should Claude check to confirm success?
The Three Components of a Routine
Every Routine has exactly three required components: a prompt that tells Claude what to do during each run, a repository that provides the codebase context, and connectors that give Claude access to external systems. All three must be configured correctly for a Routine to work as intended.
| Component | What It Is | What Goes Wrong Without It |
|---|---|---|
| Prompt | Complete instructions for one run — what to check, what to do, where to report | Vague prompts produce inconsistent outputs; prompts requiring human input stall |
| Repository | GitHub repo(s) Claude can clone and modify during the run | No repository access = Claude can't read code, open PRs, or push changes |
| Connectors | External integrations: GitHub, Slack, custom webhooks | Without Slack connector, summaries can't be posted; without GitHub connector, PRs can't be created |
The three components are independent in configuration but interdependent in execution. A prompt that instructs Claude to 'post a summary to #engineering-alerts' requires the Slack connector. A prompt that instructs Claude to 'open a PR with the fix' requires the GitHub connector. Misalignment between prompt instructions and configured connectors is the most common source of failed Routine runs.
Before activating a Routine, read your prompt and highlight every output action — 'post to Slack', 'open a PR', 'create a GitHub issue', 'call webhook'. Each one requires a corresponding connector. If the action isn't covered by a configured connector, the Routine will fail at that step.