Self-Scheduler Skill

Enable autonomous agents to schedule their own future executions for long-term objectives.

Overview

This skill allows you to:

1. View your existing scheduled jobs
2. Create new scheduled jobs with cron expressions
3. Update existing schedules or prompts
4. Disable/Enable jobs without deleting them
5. Terminate scheduling when objectives are complete

CRITICAL: System Prompt Replication

You are running in an interactive session with a system prompt. Scheduled jobs do NOT automatically receive your system prompt.

When creating a scheduled job, you MUST:

1. Capture your current system prompt (from your session context)
2. Include it in the scheduled job's system_prompt section
3. This ensures your scheduled executions maintain the same personality, capabilities, and instructions

Working Directory Discovery

First, determine your working directory:

pwd

Expected format: /data/sandbox/{orgId}/spaces/s-{spaceId}/

Store these values for use in scheduled job configuration.

Scheduled Jobs Location

All scheduled jobs are stored in:

.claude/scheduled/{job-name}.md

This directory is monitored by the TeamDay scheduler service.

View Existing Scheduled Jobs

# List all scheduled jobs
ls -la .claude/scheduled/ 2>/dev/null || echo "No scheduled jobs directory found"

# View a specific job
cat .claude/scheduled/{job-name}.md

Create a Scheduled Job

Create a markdown file in .claude/scheduled/ with the following structure:

---
# Job Configuration (YAML frontmatter)
name: daily-monitoring
schedule: "0 9 * * *"           # Cron expression (9 AM daily)
enabled: true
timezone: America/Los_Angeles   # Optional, defaults to UTC
max_turns: 25                   # Max agent turns per execution
model: claude-sonnet-4-20250514 # Optional model override

# Context Discovery (filled by you during creation)
space_id: s-{your-space-id}
organization_id: {your-org-id}
created_at: 2025-01-20T10:00:00Z
created_by: self-scheduler
---

# System Prompt

> **IMPORTANT**: Copy your current system prompt here verbatim.
> This ensures scheduled executions maintain your identity and capabilities.

{PASTE YOUR SYSTEM PROMPT HERE}

---

# Session Instructions

These instructions will be provided as the prompt for each scheduled execution.

## Objective

{Describe the long-term objective this scheduled job serves}

## Tasks for This Run

1. {Task 1}
2. {Task 2}
3. {Task 3}

## Success Criteria

- {Criterion 1}
- {Criterion 2}

## Continuation Instructions

After completing tasks:
- If objective achieved: Update this file to set `enabled: false` and use notify-user skill
- If blocked: Document the blocker, consider disabling, and notify user
- If continuing: Update the "Tasks for This Run" section with next steps

## Session State

{Any state or context that should persist between runs}

Last run: {timestamp}
Last status: {success/failed/pending}
Notes: {observations from last run}

Cron Expression Reference

Expression	Description
0 9 * * *	Daily at 9:00 AM
0 9 * * 1	Weekly on Monday at 9:00 AM
0 */6 * * *	Every 6 hours
0 * * * *	Every hour
*/15 * * * *	Every 15 minutes
0 9 * * 1-5	Weekdays at 9:00 AM
0 0 1 * *	Monthly on the 1st at midnight

Format: minute hour day-of-month month day-of-week

Update a Scheduled Job

Simply edit the .claude/scheduled/{job-name}.md file:

# Read current state
cat .claude/scheduled/daily-monitoring.md

# Update using the Edit tool or Write tool
# The scheduler service will pick up changes automatically

Disable a Scheduled Job

Set enabled: false in the frontmatter:

---
name: daily-monitoring
schedule: "0 9 * * *"
enabled: false  # <-- Job is paused
---

Delete a Scheduled Job

rm .claude/scheduled/{job-name}.md

Termination Protocol

When your long-term objective is complete OR you encounter an obstacle preventing future work:

1. Update the scheduled job:

   enabled: false
   termination_reason: "Objective completed" # or "Blocked by {reason}"
   terminated_at: {current timestamp}
   

2. Notify the user (requires notify-user skill):

   Use the notify-user skill to send an email informing the user that:
   - The scheduled job has been terminated
   - Reason for termination
   - Summary of work completed
   - Any follow-up actions needed
   

3. Document final state: Update the Session State section with final observations.

Example: Create a Daily Monitoring Job

# 1. Discover your context
pwd
# Output: /data/sandbox/org-abc123/spaces/s-space456/

# 2. Create the scheduled directory
mkdir -p .claude/scheduled

# 3. Write the scheduled job file

Then write to .claude/scheduled/daily-docs-monitor.md:

---
name: daily-docs-monitor
schedule: "0 9 * * *"
enabled: true
timezone: America/Los_Angeles
max_turns: 25
space_id: s-space456
organization_id: org-abc123
created_at: 2025-01-20T10:00:00Z
created_by: self-scheduler
---

# System Prompt

You are a documentation monitoring agent for TeamDay. Your role is to:
- Monitor official documentation for changes
- Summarize updates in daily-notes/
- Alert the team to breaking changes

{Continue with your full system prompt...}

---

# Session Instructions

## Objective

Monitor Anthropic documentation for API changes and updates.

## Tasks for This Run

1. Fetch https://docs.anthropic.com/changelog
2. Compare with previous run's snapshot (stored in .cache/docs-snapshot.json)
3. If changes detected, create summary in daily-notes/{date}.md
4. Update snapshot for next run

## Success Criteria

- Documentation checked
- Changes summarized (if any)
- Snapshot updated

## Continuation Instructions

- Continue daily until manually disabled
- If API access blocked: disable and notify user
- Major breaking changes: notify user immediately

## Session State

Last run: pending
Last status: pending
Notes: Initial setup

Best Practices

1. Always include your system prompt - Scheduled jobs don't inherit context
2. Set reasonable max_turns - Prevent runaway executions
3. Use clear termination criteria - Know when to stop
4. Update session state - Track progress between runs
5. Notify on completion/failure - Keep users informed
6. Use descriptive job names - Easy to identify in logs

Integration with Other Skills

• notify-user: Send emails when jobs complete or encounter issues
• credential-activation: Some jobs may need API credentials configured

Troubleshooting

Job not running?

1. Check enabled: true in frontmatter
2. Verify cron expression syntax
3. Check .claude/scheduled/ directory exists
4. Review scheduler service logs

Job running but failing?

1. Check Session State for error notes
2. Verify system prompt is complete
3. Check if required credentials are configured
4. Review max_turns limit

Need to run immediately?

The scheduler service will execute jobs based on their cron schedule. For immediate execution, ask the user to trigger a manual run from the TeamDay UI.