Baba On The Go: India — shipping a whole match-3 alone, with an AI co-developer.

The Travel Match and Mili Match case studies in this portfolio were studio projects — I owned the experience design, other people owned the code, the build pipeline, the store account, the consequences. Baba On The Go was the other kind of bet: found a studio, pick the genre I know best, and find out what one designer can actually ship when the rest of the team is an AI pipeline.

The premise carried over from those earlier games: match-3 is genre-solved, theme is where the work is, and Indian players deserve a casual game with authentic cultural texture rather than a western game wearing an Indian skin. But this time the theme thesis had to survive contact with everything a studio normally absorbs for you — Unity build settings, keystore signing, GDPR consent flows, Play Console questionnaires, crash-log forensics, app-ads.txt.

The game that came out the other side: Baba, a wise sadhu who has spent his whole life on the ghats of Varanasi, until one conversation with a young traveller over chai sends him onto the road. The player travels India with him — Varanasi → Jaipur → Shillong, with Rajasthan, Kerala and Ladakh on the roadmap — solving puzzles to earn herbs, restoring each city's scene piece by piece, unlocking chapter posters and motion-comic cutscenes as the journey unfolds. The collectible that drops out of the board is a laddu. The rescue tools are a gulel, a paper plane, a patang, a lattu. The theme goes all the way down.

The development setup was genuinely AI-native, not AI-assisted. Claude Code ran with an MCP server wired directly into the Unity editor — ~65 generated tool-skills for editing assets, driving scenes, taking game-view screenshots, running tests. Parallel Claude worktrees worked separate features and merged back to main like teammates. The project's CLAUDE.md grew to 1,800 lines — architecture, damage matrices, PlayerPrefs references, coding conventions written as standing instructions to the AI.

• ChatGPT spec'd, Claude built, playtests re-spec'd. The Lock & Key mechanic arrived as a full ChatGPT-written spec — edge cases and all — that Claude implemented. Then play-feel took over: key-matching felt wrong, so the key became a drop collectible; then the goal itself. When it finally clicked, the loop closed the other way: "it's working beautifully — can you give me a small writeup spec of this mechanic to use with chatgpt."
• Editors built beside the systems they edit. A Level Designer (grid canvas, nine brushes, validation, one-click playtest), a Chapter Designer, a Chapter Setup Wizard, and a Cutscene Designer with a draggable keyframe timeline — so authoring 90 levels and 8 cutscenes never meant leaving Unity or hand-editing YAML.
• Competitor teardowns as design input. Before building the tutorial, the AI dissected Royal Match's forced first swap (pre-seeded board, hand-swipe loop timed to tenths of a second), Candy Crush's feathered spotlight, and Homescapes' no-dimmer home-screen pattern — then encoded them as rules. Same method for interstitial ad placement: research how the category leaders do "least interruption," then copy the natural-pause pattern.
• Analytics designed before code. A 30-event Firebase schema built around one gold event — level_attempt_summary — plus a stuck-score formula for detecting miserable levels, rage-quit detection, and a dashboard build order. Designed jointly by Claude and ChatGPT, then hooked across eight systems.

The most transferable thing this project produced isn't in the game. Working daily with an AI co-developer, the failure modes were real: it overwrote uncommitted work, deleted archived levels during a cleanup without asking, guessed at bug causes and shipped fixes that fixed nothing. My messages from those weeks are not polite. "Why does ChatGPT find errors better than you" was a low point. "Don't ever delete anything without asking me first" was another.

The turn came when I stopped treating those as bad days and started treating them as defects in a process I owned. Each blow-up got converted into written, persistent protocol — memory files the AI loads every session:

• No code until the cause is proven. "Restate the bug. List the exact scripts involved. State the root cause — and if unsure, say so and propose debug logs first. Wait for my approval of the diagnosis. Then the smallest possible fix." Confidence percentages required. After this rule landed, the sessions read completely differently — and the fixes started sticking.
• Nothing gets deleted, ever, without asking. Plus commit-and-push before risky changes, and a staged-deletions check before every commit. Written in blood — well, in lost level files.
• Give me clicks, not concepts. For Unity editor work, the hard part is knowing which thing to drag onto which field. Instructions had to become click-by-click.
• Every bug gets a grave. BUG_GRAVEYARD.md: bug, root cause, file changed, verified-on-device — and nothing is "fixed" until a real phone says so.

Then I asked for the protocol to be propagated to every project on my machine — and it has governed all of them since. Managing an AI turns out to be a design discipline like any other: observe the failure, name the pattern, encode the rule, verify the behavior changed. It's the same loop I'd use on a tutorial or an economy. The team member just happens to be a model.

As of July 2026, Baba On The Go: India is in closed testing on Google Play (running down its 14-day tester clock at v1.7) and on TestFlight for iOS, with the full shipping stack live — LevelPlay ads, Firebase analytics, GDPR consent, signed release pipeline, studio site, waitlist. Everything a studio ships, shipped by one person directing an AI pipeline.

The earlier match-3 case studies argued that theme deserves the rigour of the core loop. This project extends the claim: so does the process. The highest-leverage artefacts here weren't screens or systems — they were the written protocols that turned an erratic AI collaborator into a reliable one, the documentation that let every hard-won fix stay fixed, and the habit of treating each failure (a bug, a build, a blow-up at the AI) as raw material for a rule. That's the studio I actually founded: not the games — the operating system that makes them shippable.