AI across the software development lifecycle
For each stage: what a person does, where AI helps, and how I check the result. The colour code is the same on every page.
- A person decides or directs
- AI produces
- How it's verified
| Stage | What a person does | Where AI helps | How it's verified |
|---|---|---|---|
| Requirements | Work out the real problem, the future state and the objectives with the business. Define the what and the why. | Drafts user stories, acceptance criteria and other deliverables. My SDLC agents each have their own skills and assemble a package of deliverables as a project moves between stages. | I check drafts against what the business actually said. Stakeholders sign off. |
| Design | Decide the user experience and the logic. | Builds working mockups of layouts and logic, so people react to something real before development starts. | Walk the mockup through with the people who will use it. |
| Build | Set the direction and scope, direct the work, review what comes back. Developers keep the how. | Claude Code writes the code under my direction. | Console logs while developing, and reading the code instead of accepting large files unread. |
| Test | Define what must be true, and which cases matter. | Drafts tests alongside the code. My personal finance app has about 160 automated tests over its core logic. | Automated tests, and Sentry to track a problem down instead of asking the AI to guess. |
| UAT | Plan and run user acceptance testing with real users. Make the training. | Drafts test scripts and training content. | Real users confirm it solves the problem it was built for. |
| Release | Write the release plan and go or no-go criteria. Make the call. | Drafts release notes and communications. | Go or no-go against criteria agreed before the release. |
| Operate | Check whether the change made a difference. Find the next improvement from users and data. | Summarizes usage and feedback data. | Sentry for errors, dashboards and a control plan with an owner. |
The loop I run on every change
This cycle repeats inside the Build and Test stages. I direct, the AI generates, tests and logs verify, and then I decide whether it ships or goes around again.
The verify step is where most of my value sits. Anyone can use AI to build a website. What I do differently is understand what got built, so when something breaks I find out why instead of telling the AI to "make it work."