# The AI Architect — Companion Pack (Book 4: Advanced Systems)

This free download comes with the book. It holds the living, hands-on materials — the parts that
change *fastest* as Claude Code's subagents and orchestration evolve, plus the practice files you'll
design and coordinate a system from.

**Hybrid on purpose.** The evergreen essentials — the Systems-Architect Cheat-Sheet, the
Safety-at-Scale Checklist, the prompt quick-reference, the glossary, and the troubleshooting guide —
are **printed inside the book's appendix**, so every reader has them without a link or a download.
This pack is the *rest*: the volatile, version-specific detail and the ready-to-use practice assets.
If a link and a paper page ever disagree, trust this download — it's the one we keep current.

**Why the download matters most in Book 4.** Subagents (specialist helpers) and the ways to
coordinate them are under active development — this is the fastest-moving corner of the whole series.
**The current walkthroughs for delegating to a specialist and for defining your own — the exact shape
of an agent-definition file, and how you ask Claude to write one — live here, not in the printed
pages**, so a detail that shifted never leaves you stuck. Buttons change; principles don't.

## What's inside
- **systems-architect-cheat-sheet.md** — the one-page approach: decompose → delegate → coordinate →
  verify → observe → guard → know the limits. Print it and keep it by your keyboard.
- **safety-at-scale-checklist.md** — the short list to run before you build a system and again before
  you trust it with real work. Least privilege per specialist, human checkpoints, a verification
  stage, observability, a cost cap, a kill switch, and knowing when to bring in a human engineer.
- **prompt-library.md** — copy-paste prompt shapes for every stage: decompose a goal, delegate to a
  built-in specialist, make a reusable specialist (describe it → Claude writes the definition),
  coordinate a pipeline, add a verification specialist, read a run, guard the whole system, and
  diagnose/recover — each with a filled-in example. Includes a prompt that scopes a specialist to
  least-privilege, read-only tools.
- **sample-data/** — small, fake, safe practice files so you can learn the shapes before you point a
  system at anything real:
  - `sample-system-spec.md` — a worked decomposition: one multi-step goal broken into specialists,
    each with its job, inputs, output, and handoff. This is what "designing a system" looks like on
    paper *before* you build it.
  - `sample-specialist-definition-EXAMPLE.md` — the *shape* of a `.claude/agents` definition file
    (frontmatter: name, description, tools, model + a short system-prompt body). Clearly an EXAMPLE,
    with minimal read-only tools, so you recognize the parts before you ask Claude to write your own.
  - `sample-data.csv` — a small, fake data file to practice feeding a specialist.

## Keeping current (read this — it matters most here)
Claude Code's subagents and orchestration change often — how delegation is triggered, how a specialist
is defined, what runs in the background, and the edges (an agent view, an experimental teams mode, and
dynamic workflows where Claude writes code) all shift as the software grows. The book teaches the
**durable ideas**; for anything version-specific (exactly how to delegate, the current shape of an
agent-definition file, how to see a run, the current model, prices, commands), check the official
sources:
- Docs: https://code.claude.com/docs
- Pricing: https://claude.com/pricing
- In the app, type `/help`

If a walkthrough or screen in this pack doesn't match yours, that's expected — follow the *idea*, and
lean on the official docs for the current click-path. **Custom specialists are no-code:** you
*describe* the specialist you want and Claude writes its definition for you — you never hand-write it.

## Pricing, honestly
Coordinated systems multiply the work. Running several specialists — especially in parallel, or as an
experimental team — multiplies token use (agent teams run roughly several times the cost of a single
session). We never print prices or limits here because they change; set a monthly limit, prefer a
cheaper model on a stage that doesn't need the strongest one, and check the current numbers at
https://claude.com/pricing.

## The Book 5 boundary
Everything in this pack keeps systems **supervised** — you drive, Claude's specialists do the parts,
you review at the risky joints, and a kill switch is always in reach. Running your whole
practice or business on these systems — the org-level "put it all to work across everything you do"
integration — is the optional **Book 5, *The AI-Powered Practice***, not something to bolt on here.
And once a job needs an SLA, audit logs, automatic retries, guaranteed uptime, or long unattended
overnight runs, that's a **human engineer** wrapping it in a real harness — not a no-code system.

## How to use it
Start with the cheat-sheet and the Safety-at-Scale Checklist — read them once before you design
anything. Then, as you reach each chapter's "Do it with me," open `prompt-library.md` and the matching
file in `sample-data/`, and follow along. When you reach the decomposition chapter, open
`sample-system-spec.md` to see a goal broken into specialists on paper; when you reach the custom-
specialist chapter, `sample-specialist-definition-EXAMPLE.md` shows the *shape* of a definition (you'll
have Claude write yours, not type it); `sample-data.csv` is safe practice data to feed a specialist.
For the exact, current delegation and definition walkthroughs, follow this download and the official
docs.

*You are the architect. Claude is the team. You keep your hand on the switch.*