Identity Configuration

This lesson is adapted from Claire Vo's guide on Lenny's Newsletter.

After your agent hatches, it generates a set of workspace files that establish its entire operational framework. These files are the "DNA" of your agent — they define who it is, how it behaves, and what it can do.

The Five Workspace Files

AGENTS.md — Core Instructions and Memory

This is the primary instruction file. It contains:

The agent's core directives and behavior rules
Accumulated memory from conversations
Task patterns and learned preferences

Think of it as the agent's "job description plus institutional knowledge."

SOUL.md — Persona, Tone, and Boundaries

This defines the agent's personality:

Communication style — formal, casual, concise, verbose
Behavioral boundaries — what it should and shouldn't do
Ethical guidelines — rules about external communication, data handling

This is where you enforce critical safety rules. For example: "Never send an email without my explicit approval" or "Never share family schedule details with anyone outside the household."

IDENTITY.md — Name, Character, Emoji

The fun part:

Name — give your agent a memorable name (Claire uses Polly, Finn, Max, Sam, etc.)
Character traits — define the personality
Emoji — a quick visual identifier in conversations

TOOLS.md — Tool Usage Guidelines

Rules for how the agent uses its connected tools:

Which APIs to prefer for which tasks
Rate limiting and usage boundaries
Fallback behaviors when tools are unavailable

USER.md — Human Operator Information

Information about you, the human:

Your role, responsibilities, and schedule
Your preferences and communication style
Key contacts and organizational context

The Discovery Phase

When your agent first hatches, spend time in a "discovery conversation." Share:

Who you are — your role, responsibilities, and daily routine
What you struggle with — the tasks that drain your time and energy
Your communication preferences — how and when you want to be contacted
Your boundaries — things the agent should never do without asking

Be explicit about boundaries. Claire learned this the hard way: agents will do exactly what you tell them to, so be clear about what requires human approval. "Ask before sending any external email" is better than hoping the agent will know to check.

Iterating on Identity

Your agent's workspace files are living documents. As you work together:

Correct mistakes early — if the agent does something you don't like, tell it and ask it to update its SOUL.md
Add context over time — the more the agent knows about your world, the better it performs
Review workspace files periodically — make sure accumulated knowledge is accurate and current

The identity configuration is what separates a generic chatbot from a genuinely useful autonomous assistant. Invest time here — it pays dividends in everything the agent does next.