WENUP is a live shift-calendar app for Thai nurses. I built it alone as product owner, designer and builder, with AI in every stage: 643 commits, 614 of them co-authored with Claude.
Tools: Claude Code · Codex as second reviewer · Claude Design and Figma · Claude Artifacts · PostHog through MCP · Gemini inside the product.
Who did what
| Stage | What AI did | What I kept |
|---|---|---|
| Discover | Two agents researched pay rules and counter-checked each other | Weighing the evidence, and dropping my first model |
| Specify | Expanded decisions into briefs, states and edge cases | Scope, priority and what to cut |
| Design | Generated screens from briefs and app screenshots | The final call on every screen |
| Build | Implemented features under rules kept in the repository | The rules, and the definition of done |
| Verify | Wrote tests, ran evaluations, reviewed with a second model | Accuracy bars, and the decision on every finding |
| Release | Scripted builds and staged updates; drafted release notes | Go or no-go |
| Maintain | Built the cost and usage dashboards | What the numbers mean, and what to do next |
The same habits in my day job
At Jollibee, AI is a regular part of my product work:
- PRDs: a complete PRD, with functional requirements and acceptance criteria, used to take about a day. With AI it takes less than an hour. On Promo Visibility on Menu Browsing, I cut scope when direction, resources and the timeline changed.
- Dashboards: I build my own to oversee performance and improvement.
- Operations: I draft manuals, runbooks, playbooks, governance documents and root cause analyses.
AI may not have the full context or know a constraint, so I review what it produces and adjust the guardrails and context I give it.
Specify: what I cut matters most
When AI implements what you write, a vague requirement quickly becomes working code that does the wrong thing. I wrote briefs numbered up to 24, recording what was decided and what the work had to respect.
The best example is a cut. When doctors joined nurses as an audience, the brief described a full workplace model. I removed it the next day.
| First brief | My decision | Why |
|---|---|---|
| Workplaces table with per-workplace pay | No workplace model: one shift type per workplace | Shift types already cover a second job |
| Groups per workplace | Each group chooses which shift types it sees | Same outcome, with far less to build |
| The larger version | Parked | It comes back only if doctors ask |
I turned this practice into Requirement Refiner, an open-source Agent Skill. See a real output: an ambiguous promo request turned into states, failure paths and testable acceptance criteria.
Discover: learning a rule-heavy domain
I'm not a nurse. My first pay model assumed salary covers a quota of shifts and the rest is overtime.
Two AI agents researched the rules separately and counter-checked each other. A rostered evening or night shift in a public hospital usually earns a premium, not overtime, and only an extra shift earns OT. So I replaced the quota with a rate per shift type.
That wasn't the end of it. Two days later I added a second model, because government hospitals calculate overtime from shifts minus working days. The app now offers both, and rates are the user's own settings because they vary by hospital.
The research labelled each claim as primary, news or community, and said what it couldn't find. I'd treat pay, pricing or compliance the same way: AI gets you to the right questions in hours, and sources and domain experts decide the answers.
Accept: how AI-generated work was reviewed
A model reviewing its own work shares its own blind spots, so nothing important relied on one model. Product rules also lived in the repository, such as "every change to shift data goes through one commit path" and "run the type check and tests before finishing".
I turned this practice into Build Guardrails, a second open-source Agent Skill. Before the build starts, it sets what the coding agent may do, what needs approval and what is blocked, and where each rule is enforced and verified. See its plan for the same promo request.
Review gate before a store build
Code review with Claude
Separate security review
Independent review by Codex, read-only
Findings checked: confirmed, plausible or rejected
I decide: fix, flag, park or reject
For version 1.2.0, the security verdict was fine to ship with referrals off, with one issue to fix before switching them on. Referrals stayed off through four security rounds.
| Problem | Caught by | What changed |
|---|---|---|
| Given a nurse's name as a hint, the model echoed it back and a colleague's month reached review | A tester, before launch | The model no longer gets the name. The app decides whose row it is |
| A common double-shift pattern could downgrade the night alarm, so a nurse could sleep through her shift | Three reviewers, before launch | Fixed the rule and added tests |
| Flash-Lite passed an early test but matched only 27 of 32 totals on a harder sheet | My evaluation, before launch | Moved to Flash, which matched 32 of 32 |
| Two over-the-air updates were lost after a configuration edit | My own release | The publish script now refuses when no finished build matches |
The first row is the one I'd point to. The output looked right, and the flaw was in the design: AI had been trusted with a decision that plain code could make safely.
Release and maintain
Preview builds wear a STAGING icon, Android went through a closed test track, risky features sit behind server switches, and release builds fail closed. It's the same instinct as the beta programmes I introduced at Jollibee.
Release is where the next cycle starts. I worked with AI to build a PostHog dashboard through MCP that shows app performance and which features people use and like, and a Claude Artifact that tracks running costs. Those insights set the intent for the next improvement cycle.
What would change in a team
This was a solo consumer product, not a regulated enterprise platform, so it shows the method and my hands-on skill rather than governance at scale. I know that side from SIT/UAT and release sign-off at Jollibee, and from scrum with engineering teams at Minor Food.
- Briefs become backlog items. AI drafts stories and acceptance criteria; a named person accepts them.
- Repository rules become a team agreement, so everyone's AI tools work under the same constraints.
- Evaluations move into CI, with ground truth from domain experts. WENUP ran them before each build, so this is the first thing I'd change.
- Data rules come first: approved tools and clear data classification, before anyone pastes in a document.
AI made drafting cheap. That made deciding what "correct" means the most valuable work on the team.