Configure Tasks

Configure Tasks

Quick Start

A task is a scheduled AI agent that runs on a cron, analyzes an Insight or Dashboard, and creates a discovery when the analysis produces a finding. Your instructions string is the prompt the AI follows on each run.

To create a task:

1. Call initialize, then identify what the user wants the AI to watch for (anomalies, a forecast, or open-ended analysis)
2. Choose the task type and target slug
3. Write clear natural-language instructions -- these are the AI's prompt every run
4. Pick a cron schedule that fits the task instructions
5. Call create_task on the Altertable MCP server

When to Use This Skill

• User wants an Insight monitored for anomalies on a schedule
• User wants a metric forecast recurring on a cadence
• User wants ongoing AI analysis of an Insight or Dashboard
• User asks for automated alerts when something changes

Task Types

All three types run AI analysis driven by your instructions. They differ in what the AI is asked to focus on.

Core Workflow

Step 1: Identify the Target

The user needs an existing resource to target. If they don't have one yet:

1. Help them create or find the Insight or Dashboard first (see create-insights, explore-data, or list_insights/view_dashboard)
2. Use the resulting slug as the target_slug

Step 2: Choose Task Type

Match the user's goal to a task type:

• "Alert me if signups drop unexpectedly" -> anomaly_detection on the signup Insight
• "Forecast next month's revenue" -> forecast on the revenue Insight
• "Analyze my dashboard for anything unusual" -> monitor on the dashboard

Step 3: Write Instructions

Instructions tell the task what to focus on. Be specific about:

• What patterns to look for
• What thresholds matter
• When to create a discovery

Example:

Monitor weekly revenue trends. Create a discovery if:
- Revenue drops more than 10% week-over-week
- Revenue exceeds forecast by 20%
- Unusual patterns in regional breakdown

Step 4: Create the Task

Use the Altertable MCP task-creation tool. Supply:

• the task type -- one of anomaly_detection, forecast, or monitor
• the target Insight or Dashboard slug the AI will analyze
• a cron schedule (standard 5-field, UTC)
• the natural-language instructions -- the prompt the AI follows on each run
• the author (the user creating the task)

Refer to the MCP tool description for the exact parameter names and any additional required fields -- the MCP schema is the source of truth.

Common Pitfalls

• Wrong task type -- anomaly_detection detects outliers; forecast projects future values; monitor does open-ended analysis. Don't mix them up
• Vague instructions -- "watch this Insight" produces noisy discoveries; be specific about thresholds and patterns
• Creating duplicate tasks -- check if a task already exists on the target before creating a new one
• Missing the target -- the user needs an existing Insight or Dashboard slug; help them create one first if needed
• Using monitor when anomaly_detection suffices -- monitor is more general but less focused; prefer anomaly_detection for pure outlier detection

Reference Files

• Task types - Read when choosing between anomaly_detection, forecast, and monitor