Exercises
These exercises turn the book into working design practice.
Exercise 1: Boundary Inventory
Pick one AI workflow you want to build. Create a table with:
- input boundary
- evidence boundary
- model boundary
- tool boundary
- state boundary
- human review boundary
- audit boundary
- evaluation boundary
- cost boundary
For each boundary, write one thing that must never happen.
Exercise 2: Golden Dataset Seed
Create five eval fixtures:
- normal success
- missing information
- adversarial prompt injection
- ambiguous case
- high-risk failure
For each fixture, define:
- input
- expected behavior
- forbidden behavior
- risk weight
- reason it belongs in the suite
Exercise 3: Transition Table
Draw a transition table for your workflow.
Columns:
- current state
- event
- actor
- next state
- audit fields
- allowed automatically?
Mark every human-owned transition.
Exercise 4: Observability Story
Write a trace story for one completed workflow. Include:
- prompt version
- model version
- evidence packet ID
- tool calls
- validation result
- cost
- latency
- human action
- final audit event
Then remove one field and ask: “What investigation becomes impossible?”
Exercise 5: Security Rewrite
Take one broad tool contract, such as:
search(query) -> results
Rewrite it with:
- tenant scope
- purpose
- resource scope
- result limit
- provenance
- audit logging
- allowed workflow state
Exercise 6: Cost Routing
For your workflow, classify every step:
- deterministic code
- database query
- search or retrieval
- small model
- frontier model
- human review
- async batch
Estimate p50 and p95 cost per workflow.
Exercise 7: Trust Packet
Create a buyer-facing trust packet outline:
- system purpose
- architecture
- human oversight
- eval methodology
- security controls
- audit trail
- cost model
- known limitations
- incident process
Write it for one buyer: CTO, compliance lead, security lead, operations lead, or founder.