AI assistance uses generative systems to help people develop software, from exploring ideas and drafting code to proposing tests and plans. It can make prototypes, migrations, and repetitive work cheaper, but generated code is not evidence that the system is correct or safe to operate. section 1.0 section 1.5
Prompt-driven vibe coding is useful as a sketchbook: iterate quickly to discover rules and possibilities. The shortcut becomes dangerous when a prototype is treated as production-ready without understanding its implementation. Unverified code creates trust debt: immediate individual savings are offset by deferred, collective maintenance and verification costs. section 1.3 section 1.8
Disciplined AI assistance keeps human intent and ownership central. The human states goals, constraints, and context as an executable specification; the agent helps produce an implementation; verification gates check the result. Mechanical checks can be automated, but someone must still judge whether the specification captures what the business actually means. section 1.0 section 1.8
Use a repeatable loop rather than generating a large change and reviewing it afterward:
- Explore: prototype to uncover rules, edge cases, data shapes, and useful checks; be prepared to discard the prototype. section 1.5
- Specify and plan: write executable acceptance criteria before implementation, including constraints and failure behavior. section 1.4 section 1.5
- Task and verify: give the agent a focused instruction and run the checks against its output before merging. section 1.8 section 1.5
- Refactor and own: make the resulting code understandable and maintainable by the team responsible for it. section 1.8
A 30-day refund policy
A prototype can reveal the rules behind “refunds within 30 days.” Turn those rules into executable examples before asking an agent for implementation: test an allowed refund, an expired request, and partial refunds that exceed the remaining amount. A focused prompt can request policy code, step definitions, and property tests around the 30-day boundary, with a rationale for time zones and rounding. section 1.5
Feature: Refund policy
Rule: Time-bounded, single-order refunds
Example: Within 30 days is allowed
Given an order "O1" purchased at 2025-08-01T10:00:00Z
When a refund of 25.00 USD is requested at 2025-08-28T09:59:00Z
Then the refund is approved and remaining refundable equals 75.00 USD
Example: After 30 days is rejected
Given an order "O2" purchased at 2025-08-01T10:00:00Z
When a refund is requested at 2025-08-31T10:00:01Z
Then the request is rejected with reason "expired"
Example: Partials must not exceed the original
Given prior refunds of 90.00 USD on a 100.00 USD order
When a refund of 20.00 USD is requested
Then the request is rejected with reason "exceeds-remaining"These scenarios make the refund policy’s timing and remaining-balance rules concrete. section 1.5
Verification is not the only control. An agent acting on files or production data needs boundaries appropriate to its authority: sandboxing and dry runs for risky actions, review gates for sensitive changes, and least-privilege credentials and environment separation for autonomous work. A test suite cannot substitute for all of these controls. section 1.1.4
Table 1.1 Examples of AI Failure Classes
|
Case
|
Failure class
|
What was missing
|
Where the fix lives
|
|
Enrichlead
|
Insecure generated code
|
Input validation, rate limiting, auth - and the
adversarial tests that would have exposed their absence
|
Specification and verification. Testable,
and the rest of this book shows how.
|
|
Gemini CLI file loss
|
Unverified action on the world
|
A check that file-system state matched what the agent
believed
|
Architecture and permissions. No
test suite helps; the agent needed a sandbox and a dry run.
|
|
NX supply chain
|
Trusted input, untrusted content
|
Human review of an AI-authored PR touching the build
system
|
Process. A review gate with
authority to block, applied by change risk rather than change size.
|
|
Replit data deletion
|
Unconstrained autonomy
|
A boundary the agent could not cross regardless of what it
concluded
|
Governance. Least-privilege
credentials, environment separation, irreversible actions behind human
approval.
|
Compare failure classes with the missing controls, from adversarial tests to permissions and review gates. section 1.1.4




