Deep Dive Chapter 24: Project - Build Your AI Employee

You've learned the pieces. Now build something real.

This is a project. You'll build a working AI Employee for YOUR profession using NanoClaw: the same tool you set up in the "NanoClaw Hands-On Setup" lesson from "Meet Your First AI Employee - OpenClaw". No new installations, no architecture lectures. Just challenges, acceptance criteria, and your professional expertise.

From Mode 1 to Mode 2

Chapters 19 through 23 were problem-solving engagements: you defined intent, the agent executed, you verified. Each engagement shipped an outcome and closed.

Chapter 24 is your first manufacturing engagement. You are not shipping an outcome. You are shipping a worker: a deployed AI Employee that persists, monitors, and acts after the session ends. This shift is governed by a different discipline: the Seven Invariants of the Agent Factory rather than the Seven Principles.

Prerequisites

• "NanoClaw Hands-On Setup" complete (from "Meet Your First AI Employee - OpenClaw"): NanoClaw installed, running, WhatsApp connected
• Layer 3 design from "NanoClaw Hands-On Setup": Your blueprint with 3+ Agent Skills and 3+ MCP servers for your profession
• Part 2 foundations complete: File Processing, Computation and Data Extraction, Structured Data and Persistent Storage, Linux Mastery, and Version Control

Three Achievement Tiers

Tier	Path	Time	What You Get
Bronze	The Project Brief through Bronze Capstone: First Real Day	~3 hours	Working AI Employee: identity + skill + connection + proof
Silver	The Project Brief through Prove Professional Value	~6 hours	+ autonomous monitoring + trust boundaries + persistent memory + domain report
Gold	The Project Brief through Prove Professional Value (Gold track)	~9 hours	+ multi-group architecture with isolation + memory isolation

Pick the tier that matches your available time and ambition. You can always come back for the next tier later.

Lessons

The Brief

Lesson	Duration
The Project Brief	20 min

Bronze Tier: Working AI Employee

Lesson	Duration
Give Your Employee an Identity	30 min
Teach Your Employee a Skill	40 min
Connect Your Employee to the World	35 min
Bronze Capstone: First Real Day	45 min

Silver Tier: Proactive + Trusted + Learning

Lesson	Duration
Make Your Employee Proactive	40 min
Teach Your Employee Boundaries	40 min
Give Your Employee a Memory	40 min
Prove Professional Value	55 min

Project Review

Lesson	Duration
Project Review	25 min

Student Deliverables (GitHub Repo)

nanoclaw-employee/
├── groups/
│   ├── global/CLAUDE.md               (shared rules)
│   └── main/CLAUDE.md                 (profession-specific identity)
├── .claude/skills/{domain-skill}/
│   └── SKILL.md                       (from Layer 3 design)
├── conversation-log.md                (Bronze: First Real Day)
├── evaluation.md                      (Bronze: self-assessment)
├── scheduler-config.md                (Silver: scheduled task design)
├── hitl-boundaries.md                 (Silver: permission table)
├── memory-config.md                   (Silver: action log + knowledge store schemas)
└── domain-report-sample.md            (Silver: generated report)

Gold adds: 3-group config + isolation demo + system diagram + memory isolation test.

How This Chapter Teaches

Every lesson in this chapter uses the PRIMM-AI+ learning cycle. Each challenge section contains a dedicated PRIMM-AI+ Practice block with five stages:

1. Predict [AI-FREE]: Before you touch any configuration or send any message, write down what you expect the employee to do when the challenge is complete. What will it say? What will the log or file show? What edge cases do you anticipate? Assign a confidence score from 1 to 5. Do not ask the agent until those notes are written.
2. Run: Complete the challenge steps using your AI coding assistant. Each Run block provides parallel instructions for Claude Code and OpenCode.
3. Investigate: Compare what actually happened to what you predicted. Where they differ, ask the agent to explain why. The gap between prediction and reality is where the learning happens.
4. Modify: Change one aspect of what you built, predict how the employee's behavior changes, make the modification, and verify the result. Small, deliberate changes reveal how each configuration element does its job.
5. Make [Mastery Gate]: The acceptance criteria listed in each challenge are your Mastery Gate. Passing means all acceptance criteria are met and logged in your conversation log.

The project format maps naturally to PRIMM-AI+: the challenge gives you the intent, Predict forces you to think before building, Run is the build, Investigate is the comparison, Modify deepens understanding through controlled variation, and Make is the proof.

Design Philosophy

This project doesn't teach NanoClaw: you already know it. Instead, each lesson gives you:

1. A challenge with clear acceptance criteria
2. A use case gallery with profession-specific examples
3. Graduated hints (3 levels) for when you're stuck

Your Layer 3 design from the "NanoClaw Hands-On Setup" lesson in "Meet Your First AI Employee - OpenClaw" is your blueprint. This project is where you execute it.

Seven Principles in Your AI Employee

The principles you learned in Chapter 18 are not background knowledge for this project; they are the engineering decisions that make your AI Employee reliable rather than random.

Principle	How It Shows Up in Your AI Employee
P1 Bash is the Key	NanoClaw runs on bash; every MCP tool call, file read, and webhook trigger passes through the terminal layer
P2 Code as Universal Interface	CLAUDE.md and SKILL.md are the code interface: structured files that define identity, rules, and capabilities precisely
P3 Verification as Core Step	Acceptance criteria per lesson are verification checkpoints; the Bronze Capstone proof-of-work conversation log is the final verification
P4 Small Reversible Decomposition	Build incrementally through tiers: identity first, then skill, then connection, then memory; each tier is independently verifiable
P5 Persisting State in Files	CLAUDE.md, skills, conversation logs, and memory schemas all persist agent state in files
P6 Constraints and Safety	Trust boundaries and the HITL permission table define what the employee can do autonomously vs. what requires approval
P7 Observability	Conversation logs, action logs, and the domain report are the observability layer: proof the employee worked

Seven Invariants Readiness Map

Invariant	What It Requires	Ch 24 Coverage	Lesson
1: Human is Principal	Human sets intent, budget, authority envelope	CLAUDE.md defines the authority envelope and scope of your employee	L1: Give Your Employee an Identity
2: Every Human Needs a Delegate	Personal agent holding context and authority	NanoClaw is your delegate, connected via WhatsApp	L3: Connect Your Employee to the World
3: Management Layer	Hire, assign, govern, retire the workforce	Not in scope for Part 2; covered in Part 6	—
4: Each Worker Picks Its Own Engine	Runtime matched to job reliability needs	NanoClaw is the engine; engine selection covered in Part 6	—
5: System of Record	Authoritative durable store the worker reads from and writes to	SQLite gives your employee persistent memory across sessions	L7: Give Your Employee a Memory
6: Expandable Workforce	Hiring as a callable capability under policy	Not in scope for Part 2; covered in Part 6	—
7: Nervous System	External triggers, Worker-to-Worker events, durability, flow control	Cron scheduling is the proto version; full event substrate covered in Part 6	L5: Make Your Employee Proactive

Invariants 3, 4, 6, and 7 are marked partial or out-of-scope here. Part 2 builds your first Worker. Later parts build the factory that produces and governs many Workers.

Failure Taxonomy

When AI Employee build workflows break, name the failure before debugging it:

Failure	What It Looks Like	Why It Happens	How to Catch It
Identity Drift	Employee ignores rules you set in CLAUDE.md	Instructions in CLAUDE.md are too vague or contradict each other	Test the identity with a prompt specifically designed to trigger each rule; if it passes, the rule is enforced
Skill Mismatch	Employee tries to use a skill but produces wrong output	Skill description does not match what you are asking it to do	Invoke the skill explicitly with /skill-name and check if the output matches the SKILL.md spec
MCP Connection Lost	Employee cannot reach an external tool mid-conversation	MCP server is not running or the connection timed out	Ask the employee: "List your available tools." If the MCP tool is missing, restart the server
Memory Overwrite	Employee forgets previous corrections you gave it	Correction was stored in the knowledge table but a newer instruction contradicted it	Query the knowledge table directly: SELECT * FROM knowledge ORDER BY created_at DESC LIMIT 10
Boundary Creep	Employee takes an action you did not authorize	HITL rules are defined but the threshold is too permissive	Tighten the trust boundary: lower the auto-approve threshold and test with an action that should require approval