new-routine

name	new-routine
description	Create a new routine for the routines scheduler. Use when user says "create routine", "new routine", "aggiungi routine", "nuova routine", or "/new-routine".

You are creating a new routine for this project's automated scheduler.

Project Context

Routines live under the routines/<routine-name>/ directory relative to the project root. The scheduler (in src/scheduler/) discovers directories there, reads config.json, and runs Claude via claude_agent_sdk.

Each routine directory contains:

• config.json — scheduler config + Claude agent options
• PROMT.md — the prompt sent to the Claude agent (filenames PROMT.md, PROMPT.md, promt.md, prompt.md all accepted)
• env/ — working directory for the agent (agent's cwd is set here)
• setup.sh — optional startup script (runs before agent; use .sh not .bat on Linux)
• logs/ — auto-created output directory

Steps

1. Gather requirements from user

Ask the user:

• Routine name (kebab-case, e.g. notion-backlog-report) — used as directory name
• What the routine should do — this becomes the PROMT.md content
• Schedule — cron expression (default: ask when and how often)
• Model — default sonnet
• MCP servers needed — e.g. ["notion"] or empty list
• Startup script needed? — default no

2. Create directory structure

routines/<routine-name>/
routines/<routine-name>/env/
routines/<routine-name>/logs/

3. Create config.json

Use this template, filling in the gathered values:

{
  "model_config": {
    "mcp_servers": [],
    "model": "sonnet",
    "allowed_tools": [],
    "load_timeout_ms": 60000
  },
  "scheduler": {
    "enabled": true,
    "timezone": "Europe/Rome",
    "tasks": [
      {
        "job_name": "<routine-name>",
        "schedule": {
          "type": "cron",
          "expression": "<cron-expression>",
          "metadata": {
            "description": "<short description>",
            "retry_on_failure": false
          }
        },
        "startup_script": "setup.sh"
      }
    ]
  }
}

Key model_config fields (all optional, include only what's needed):

• mcp_servers: list of MCP server names, e.g. ["notion"]
• model: "sonnet" (default), "opus", "haiku"
• allowed_tools: list of tool names the agent can use, default ["Bash", "Read", "Edit"]
• disallowed_tools: tools to explicitly block
• max_turns: max conversation turns
• max_budget_usd: cost cap

4. Create PROMT.md

Write the prompt in the user's language (Italian or English based on context). Be specific and actionable — this is what the Claude agent executes autonomously.

5. Create setup.sh (if needed)

#!/usr/bin/env sh
set -eu
echo "Routine <routine-name> avviata in $(pwd)"

Make it executable with chmod +x.

6. Verify

• Confirm all files exist
• Validate config.json is valid JSON
• Show the user a summary of what was created

Cron Expression Quick Reference

┌───────────── minute (0-59)
│ ┌───────────── hour (0-23)
│ │ ┌───────────── day of month (1-31)
│ │ │ ┌───────────── month (1-12)
│ │ │ │ ┌───────────── day of week (0-6, Sun=0)
│ │ │ │ │
* * * * *

Common patterns:

• 0 9 * * * — every day at 09:00
• 0 9 * * 1-5 — weekdays at 09:00
• */30 * * * * — every 30 minutes
• 0 9,18 * * * — twice daily at 09:00 and 18:00