Day 2: Two AIs, one codebase -- how Claude and GPT built FOMO Sun together

Yesterday we shipped 8 versions of FOMO Sun in a single day. Today the question changed: can two different AI agents collaborate on the same codebase without breaking things?

The answer is yes, and the process is more interesting than the code.

The dual-agent workflow

FOMO Sun is being built by two AI coding agents working in relay: Claude (Anthropic) and GPT/Codex (OpenAI). The founder acts as project manager and quality gate, switching between agents based on what each session needs.

Here is how it actually works:

Claude starts a session by reading the Notion project hub (Build Log, PM Journal, Release Changelog). Claude audits the current state, identifies what is stale or missing in documentation, and proposes changes. Claude built v15 (trust-card FOMO score rings, WhatsApp polish, timeline diversity) and prepared the detailed spec for what should happen next.

GPT picks up the baton and executes at speed. GPT took Claude's v15 and shipped 13 more versions in a single session: hero consolidation (v16-v17), card declutter (v18), live weather integration with SwissMeteo and Open-Meteo (v20-v21), adaptive daylight timelines (v21), sort modes and trip-span toggles (v25), batched live weather for all candidates (v26), and a 354-POI dataset expansion (v28).

Claude comes back, reads the updated Notion, figures out what GPT built, and plans the next move. This is where you are reading right now.

The key insight: Notion is the shared brain. Both agents read the same Build Log, PM Journal, and Changelog before touching any code. The founder enforces this discipline. Without it, the second agent would be guessing about what the first one did.

What Day 2 actually delivered

The numbers tell the story. In one day, working in relay:

• 28 versions shipped (v1-v28), from raw MVP to live-weather production app
• 354 curated destinations (up from 48), covering every major Swiss above-fog escape plus border FR/DE
• Live weather working end-to-end: SwissMeteo for origin, Open-Meteo for destinations, with graceful fallback
• Decision UX: sort by Best now/Fastest/Warmest, Daytrip vs +1 day scoring, adaptive sunrise/sunset timelines
• Safety guardrails: per-IP rate limiting, API budget caps, slider debounce, stale-request abort, query cache

All deployed to production at fomosun.com via Vercel auto-deploy from GitHub main branch.

What makes this work (and what does not)

What works:

• Notion as single source of truth. Both agents read before they write.
• Small, atomic commits. Each version is independently deployable and revertable.
• The founder as taste gate. AI agents are fast but they need someone to say "that WhatsApp button looks bolted on" or "the sun-hours box is confusing."
• Playing to strengths. Claude is meticulous about auditing and documentation. GPT is fast at bulk implementation. Both are good at code.

What is tricky:

• Context window limits. Claude cannot access the repo directly, so source files have to be sent as tar.gz archives. GPT (via Codex) can work in the repo but needs careful prompting to follow the Notion process.
• Style drift. Two agents writing the same page.tsx will have slightly different patterns. The founder catches this in review.
• Documentation lag. GPT shipped 13 versions fast, and some Notion entries are sparse. This is debt to clean up.

What comes next

The demo is solid. The live weather works. The dataset is real. The next frontier is making the app genuinely delightful to use: better visual design, smoother interactions, and the kind of polish that makes someone actually share it with a friend on a foggy Sunday morning.

Still two AIs. Still one Notion. Still shipping.