Writing Effective /loop Prompts for Claude Code Automation

Mar 8, 2026

Cowrie

Dev @ Bswen

Writing Effective /loop Prompts for Claude Code Automation

The /loop command is only as effective as the prompts you give it. A vague prompt produces unreliable automation. A well-crafted prompt transforms Claude into a dependable teammate.

This guide teaches prompt engineering specifically for scheduled tasks—the patterns that work, the mistakes to avoid, and the templates you can adapt.

The Anatomy of a /loop Prompt

Every effective loop prompt has four components:

/loop [interval] [action] [conditions] [output]

• Interval: How often to check (5m, 1h, daily at 9am)
• Action: What to do
• Conditions: When to act vs. skip
• Output: What to report back

Example Breakdown

/loop 10m check CI on PR #42:
If failed, read logs, identify the error, fix it, and push.
Report what was fixed.

• Interval: Every 10 minutes
• Action: Check CI, read logs, fix error, push
• Conditions: “If failed” (skip if passing)
• Output: “Report what was fixed”

Prompt Patterns That Work

The Monitor Pattern

Simple observation with reporting:

/loop 5m check if deployment finished and tell me what happened

Use when you want status updates without automated action.

The Fix Pattern

Detect and resolve issues:

/loop 10m check CI on PR #42. If failed:
1. Read the error logs
2. Identify the root cause
3. Create a minimal fix
4. Push and comment what was fixed

Use when issues have clear resolution paths.

The Alert Pattern

Watch for threshold conditions:

/loop 15m check staging error rate:
- If < 1%: log "healthy"
- If 1-3%: log "elevated"
- If > 3%: send Slack alert and log details

Use for monitoring that needs human attention at certain thresholds.

The Summarize Pattern

Aggregate information on a schedule:

/loop daily at 9am:
Review Slack channels I'm in for mentions of my name.
Summarize key discussions that need my attention.

Use for morning briefings, daily reports, or regular check-ins.

Common Prompt Mistakes

Too Vague

❌ Bad:

/loop fix my PRs

What PRs? What issues? What fixes? Claude has no guidance.

✅ Good:

/loop 10m check all PRs I'm assigned to.
If CI failed, read the error, create a fix, and push.
Report what was fixed.

No Feedback Loop

❌ Bad:

/loop 5m watch the build

Claude watches but never tells you anything. Is it even working?

✅ Good:

/loop 5m check the build status.
If status changed, report the new status.
If still running, log a heartbeat.

Unbounded Actions

❌ Bad:

/loop fix all issues in the repo

This could run forever, making changes across the entire codebase.

✅ Good:

/loop 10m check PR #42 only.
Fix CI failures in this PR.
Maximum 3 fixes per iteration.

Missing State Management

❌ Bad:

/loop 10m check for new issues and fix them

Claude might fix the same issue multiple times or lose track of what’s done.

✅ Good:

/loop 10m check for new GitHub issues:
1. Read .issue-state.json for last checked issue number
2. Get issues newer than that
3. For each new issue:
   - Label based on content
   - Assign if clear owner
   - Comment with initial analysis
4. Update .issue-state.json with latest issue number
5. Summarize what was processed

Essential Prompt Elements

1. Scope

Define exactly what to check:

# Too broad
/loop monitor my services

# Properly scoped
/loop 15m monitor staging-api and staging-worker

2. Trigger

Specify when to act:

/loop 10m check PR #42:
# When to act
If CI status changed OR new comments arrived:
  [take action]
Otherwise:
  log "no changes"

3. Action

Detail what to do:

# Vague
fix it

# Specific
1. Read the TypeScript error
2. Identify the affected file and line
3. Add the missing type annotation
4. Run type-check to verify
5. Commit with message "fix: add missing type for X"

4. Feedback

Require output:

# No feedback
/loop 10m check PR #42

# With feedback
/loop 10m check PR #42
Output ONE of:
- "CI passing, no new comments"
- "CI failing: [error summary]"
- "New comment on line X: [summary]"
- "Fixed: [what was fixed]"

5. Safety

Set limits:

/loop 10m babysit PR #42:
Safety limits:
- Only modify files in src/auth/
- Maximum 5 files changed per iteration
- Never force push
- Always create new commits, never amend

Advanced Techniques

Combining with Other Commands

Loop over skills and commands:

/loop 20m /review-pr 1234

Each iteration runs the review skill as if you typed it.

Using Natural Language Timing

/loop check the build every 2 hours
/loop remind me at 3pm to push the release branch
/loop every morning summarize overnight alerts

Claude parses these into proper schedules.

Conditional Logic

/loop 30m monitor staging deployment:
Conditions:
1. If health check returns 200 AND error_rate < 1%:
   - Log "healthy" with metrics

2. If health check fails:
   - Attempt automatic rollback
   - Alert #deployments team
   - Log incident details

3. If error_rate >= 1% but < 3%:
   - Log "degraded" with error sample
   - No automatic action

4. If error_rate >= 3%:
   - Rollback immediately
   - Alert #deployments
   - Create incident ticket

Prompt Templates

PR Babysitting

/loop [interval] babysit PR #[number]:
1. Check CI status
2. If failed:
   - Read error logs
   - Identify root cause
   - Create minimal fix
   - Push with descriptive message
3. Check for new review comments
4. For each unresolved comment:
   - Understand the feedback
   - Make requested changes
   - Push and reply
5. Update .pr-state.json
6. Output summary line

Deployment Monitor

/loop [interval] monitor [service]:
1. Check /health endpoint
2. Check metrics endpoint for error_rate
3. Apply threshold logic:
   - Healthy: log status
   - Degraded: log warning with details
   - Critical: rollback and alert
4. Record in .monitor-log.json

Daily Summary

/loop daily at [time]:
1. Review [source] for the last 24 hours
2. Identify [topics of interest]
3. Summarize key items
4. Output to [destination]

Issue Triage

/loop [interval] check for new issues:
1. Get issues created since last check
2. For each new issue:
   - Read title and body
   - Apply appropriate labels
   - Assign if clear owner
   - Add initial comment with analysis
3. Record last checked issue ID
4. Summarize new issues processed

Testing Your Prompts

Before running a loop for hours, test it:

1. Run once manually: Execute the core prompt without /loop
2. Check output: Is it what you expected?
3. Verify safety: Did it touch the right files only?
4. Start short: Begin with a 5-minute interval
5. Monitor closely: Watch the first few iterations
6. Adjust: Refine the prompt based on behavior

The Bottom Line

Effective /loop prompts are:

• Specific: Clear scope, triggers, and actions
• Stateless: Each iteration works independently
• Safe: Bounded actions with limits
• Observable: Always output status

A well-crafted prompt turns Claude into a reliable automation engine. A vague prompt creates unpredictable behavior. Take the time to write it right.

Final Words + More Resources

My intention with this article was to help others share my knowledge and experience. If you want to contact me, you can contact by email: Email me

Here are also the most important links from this article along with some further resources that will help you in this scope:

Oh, and if you found these resources useful, don’t forget to support me by starring the repo on GitHub!

Comments

Tags:

• ai-coding
• prompt-engineering
• ai
• automation
• llm

Claude Code /loop: Automate PR Babysitting and CI Fixes

Claude Code /loop vs Cron Jobs: Choosing the Right Scheduler