Building 3 apps with AI: lessons learned, part one

The numbers.

We produced 3 independent apps in parallel, for internal consumption but at large scale.

• 2 people only, for all 3 apps.
• ~120 hours of calendar work (240 hours of total effort).
• Used Kiro + Claude Code + AI-DLC from AWS + a live, continuously evolving set of custom steering rules.

Lessons learned, part one.

The repo must be designed for AI-assisted development.

You cannot just bring an agent into an old repo designed by a human and expect it to replace the human. This means almost inevitably that even small apps should be composed from a number of small repos, instead of one single repo for the entire app.

AI-DLC from AWS works very well.

Seemed slow at first — going through the spec-driven process — but was worth it for the speed and quality returned. Gets faster once you start trusting the agent(s).

Prompting experience eliminates the need to read specs end-to-end.

More experience with prompting almost completely removes the need for the human to review every detail of generated specs and task breakdowns.

AI really wants to document everything.

Claude almost feels eager to jump into documentation — very different from how docs were produced by human devs in past years. Add steering instructions on how to handle this and when to write documentation, not just how.

Commits can grow fast — and that becomes a problem.

Compared to slower human-based development, AI-assisted changes can pile up quickly and a commit can easily start to lose focus. Small repos help narrow it down but keep in mind the outcome so this doesn’t become a challenge for collaborators.

Walk backwards from the commit message.

As a human doing prompting and reviewing, make the desired commit message your mental model. That keeps commits small, focused, and easy to follow.

Part Two (coming soon)