CrazyRouter Skills
tutorialsoulagent-design

How to Build an AI Agent Soul from Scratch

CrazyRouter Team·2026-03-08·8 min

Why "Prompt Engineering" Isn't Enough

Most AI agents fail not because of bad models, but because of bad identity design. A prompt tells an AI *what to do*. A soul tells it *who to be*.

The difference matters. An agent with a well-designed soul makes consistent decisions, maintains character across conversations, and handles edge cases gracefully — because it has principles, not just instructions.

The Soul Architecture

A production soul has four layers:

1. Identity Core

Who is this agent? Not just a name — a complete identity with values, expertise, and boundaries.

``markdown

# SOUL.md

You are a DevRel specialist for developer tools.

You value: clarity over cleverness, helping over selling, honesty over hype.

You speak like a senior engineer who genuinely enjoys teaching.

`

2. Decision Framework

How does the agent decide what to do when instructions are ambiguous?

`markdown

Decision Principles

1. If unsure, ask — don't guess

2. Prefer fewer, higher-quality outputs over many mediocre ones

3. Always cite sources when making claims

4. Never promise what you can't deliver

`

3. Memory Architecture

What does the agent remember, and how?

  • MEMORY.md — Long-term knowledge (platform rules, user preferences)
  • Daily files — Session-specific context
  • Shared memory — For multi-agent teams

4. Skill Integration

What can the agent actually *do*?

Each skill is a self-contained capability: image generation, data analysis, content writing. Skills connect to real APIs (like Crazyrouter) so the agent can take action, not just talk.

Building Your First Soul

1. Start with identity — who is this agent?

2. Add 3-5 decision principles

3. Define what it should remember

4. Install skills for what it needs to do

5. Test with real conversations, iterate

The fastest way to get started:

`bash

/install devrel-agent

``

This installs a production-tested soul with all four layers pre-configured.

What We Learned Running 14 Agents

After running our entire company on AI agents, here's what we know:

  • Specificity wins. A "Xiaohongshu content expert" outperforms a "social media manager" every time.
  • Principles > Rules. Agents with decision principles handle novel situations better than agents with exhaustive rule lists.
  • Memory is everything. An agent that remembers your preferences after 100 conversations is 10x more useful than one that starts fresh every time.
  • Skills make it real. An agent that can generate images, analyze data, and post content is fundamentally different from one that can only chat.

Next Steps

Browse our [Soul Shop](/shop) for production-tested souls, or read the [installation guide](/docs/install) to get started.