magic-need

Capture what your AI agent wishes it had. Let your agent spec your product for you.

Overview

When an AI agent is executing a task and hits a wall because it lacks data or tools, instead of just failing or working around it, this skill lets the agent register exactly what it's missing. Over time, this builds a prioritized roadmap of integrations and features.

Inspired by Sonarly's magic_fetch concept — give the agent a "tool that does nothing" and let it tell you what it actually needs.

Usage

As an Agent (During Task Execution)

When you realize you need something you don't have:

CODEBLOCK0

The CLI will:

1. 1. Save the need to INLINECODE1
2. Auto-categorize it (integration, observability, devops, auth, database, storage)
3. Return a confirmation with the need ID

As a Human (Reviewing Needs)

CODEBLOCK1

Auto-Categorization

Needs are automatically categorized based on description keywords:

Data Format

Needs are stored as JSON in ~/.magic-need/needs.json:

CODEBLOCK2

Report Format

The report command outputs a formatted summary:

CODEBLOCK3

Best Practices

Good Need Descriptions

Be specific about what you need:

• - ✅ "API endpoint for deploys in the last 2 hours, filtered by service name"
• ✅ "CPU and memory metrics for upstream auth-service pods"
• ✅ "Feature flags that changed in the last 24h for api-gateway"
• ✅ "Sentry errors grouped by affected user segment"

Bad Need Descriptions

Avoid vague descriptions:

• - ❌ "need more data"
• ❌ "can't do this without tools"
• ❌ "would be nice to have logs"

Integration Roadmap

Periodically review the generated reports to:

1. 1. Identify patterns (which categories have the most needs?)
2. Prioritize integrations (which needs block the most tasks?)
3. Build the most impactful tools first

CLI Reference

See scripts/cli.js for the full implementation.

Commands

Cron Integration

To receive daily reports, set up a cronjob:

CODEBLOCK4

Or use OpenClaw's cron system to send reports to a Discord channel.