Skip to content

Case Studies

These are reconstructed engagements: invented teams, real dynamics. Every study follows the same arc — the vague prompt someone actually typed, what it produced or would have produced, the spec that replaced it, and at least one mistake that cost real hours. The mistakes are the point. A case study where everything worked teaches nothing you couldn't get from a playbook.

How to read them: steal the intervention points, not the stack choices. Whether a team used Fastify or Express is incidental. Where the human stopped the agent, what they checked before merging, and which claim they failed to verify — that transfers directly to your codebase. Each study ends with numbered lessons; if a lesson feels generic, we failed and you should tell us.

StudyScenarioStackHeadline lesson
SaaS MVPTwo founders build a gym-equipment maintenance SaaS in 3 weeksNext.js, Fastify, Postgres, StripeA 6-hour demo is not 20% of an MVP — it's 2%. Restart with specs, not momentum
Node Backend RefactorUntangling a 45k-LOC, 5-year-old Express billing backendNode, Express, MySQL, JestPin current behavior — including the bugs — before letting an agent "improve" anything
Add AuthenticationRetrofitting real auth onto a product with 30 paying customersReact, Fastify, PostgresAuth is the worst feature to hand an agent vaguely; happy-path tests hide security holes
Admin Dashboard12 CRUD/ops screens for support and operationsReact, TanStack, FastifyBuild one exemplar screen carefully, then mass-produce — and regenerate, don't patch
Prototype to ProductionHardening a deliberately dirty prototype before 200 users arriveExpress, BullMQ, PostgresVerify cross-cutting claims per surface, not per PR — "added error handling" rarely means everywhere
CI/CD PipelineStanding up a full pipeline with agents writing the workflow codeGitHub Actions, Docker, TerraformPipeline code is code: it needs review gates just like the app it ships
Internal PlatformA platform team builds golden-path tooling with agent leverageGo, Kubernetes, BackstageInternal tools tolerate agent speed better than external products — but conventions must come first
Architecture RecoveryReverse-engineering an undocumented system into decision recordsJava, Spring, OracleAgents are excellent archaeologists if you make them cite file-level evidence for every claim
Multi-Agent FeatureParallel subagents deliver one feature across API, UI, and testsTypeScript monorepo, Claude Code subagentsParallelism pays only when interfaces are frozen before agents fan out
AI-Built ProductA product where ~90% of code was agent-written, end to endNext.js, Postgres, Claude Code in CIThe human's job shifts entirely to specs, review, and verification — and that job gets harder, not easier

A field manual for AI-native software engineering.