Module 20 — AI-Assisted Requirements, Development, and Testing
This is where we can make the process significantly more systematic.
The important design principle is:
AI may derive, organize, implement and test requirements—but it must never silently invent missing business requirements.
That's Rule #1.
20.1Think of AI as several specialized rolesDon't give one giant agent:20.2Recommended repository structureI would actually put the integration knowledge in the repository:20.3Add a requirements state modelEvery requirement should be:20.4The most important AI rulePut something like this in your project's agent instructions:20.5Second critical rule: distinguish facts from assumptionsThis is extremely useful for integrations.20.6Mapping-agent skillCreate a specialized skill/prompt:20.7Q&ABusiness-requirements interview skillPrompt:20.8Requirements-gap agentAfter every workshop, give AI:20.9Automatically produce a decision logAfter Product answers:20.10Generate acceptance scenarios automaticallyFrom:20.11codeRequirement → tests traceabilityCreate IDs:20.12AI architecture reviewer skillPrompt:20.13AI implementation skillThen implementation agent gets:20.14DataWeave-review skillVery Mule-specific and useful:20.15Salesforce integration reviewerPrompt:20.16Error-handling reviewerVery high-value automated review.20.17Test-generator skillDo not prompt:20.18Regression-selection agentThis is particularly useful.20.19codeAI should verify implementation against mappingsThis can be automated surprisingly well.20.20Schema-drift agentOn a schedule or CI:20.21Requirements-change impact agentPM changes:20.22AI-generated test dataGive it Salesforce constraints:20.23Production-log feedback loopA powerful later-stage system:20.24But AI should not automatically turn every production error into a testRule:20.25AI pull-request reviewerI would have a PR-level rule like:20.26Security-review agentPrompt:20.27Observability-review agentEvery new integration operation should answer:20.28Definition of Ready agentBefore implementation:20.29Definition of Done agentAfter implementation:20.30I'd put a compact rules file in the repoSomething like:20.31Then add specialized “skills”I'd create these:20.32A possible AI workflowHere's the workflow I'd actually want:20.33Where humans stay authoritativeI would explicitly reserve these for humans:20.34Where AI can be highly autonomousAI can safely do much more with:20.35A great “requirements meeting → engineering” master promptAfter a meeting:20.36And the implementation master promptOnce approved:20.37My preferred AI test strategyHave AI think in layers:•cheat sheetModule 20 Cheat SheetThe interview-worthy sentence tying both modules together is: