AI workflow in building ImperiaQuiz

I am trying to go all-in. This project (ImperiaQuiz) started first as an experiment - how far can I get with just AI usage. Slowly, I found out that it works pretty well and the results are satisfactory.

The big picture

I act more as a ‘product owner’ than a developer. While I give a look at the code changes, I found that I care less and less about reading line-per-line what is going on and more broadly what is going on.

The backend is Go, frontend is React app. I have written go before, but just as a learning/side project. Read half a book, did a few courses, wrote a sudoku solver and all that. So I can at least get the basics. But there’s much more going on which is way above my head. And I won’t attempt to hinder the AI output by having a strong opinion. My bet is that I know less. I am not a backend developer after all.

Instead, I put my energy towards the product itself. I want to build something to use. I honestly don’t care too much about per-line issues or “this pattern vs that pattern” as long as it’s consistently organized and makes sense.

Linear -> Browser

First I add tasks to Linear. (Subscription #1) Tasks are usually just titles + descriptions of what I want in the app. They are a mix of user request (as a player) and technical request (as my web-dev experience guides me).

At v0.7.0 I’ve had about 500 linear tickets. My workflow requires each ticket to contain a PR though sometimes I am merging them and some tickets got deleted.

Important in a linear ticket - it must have clear finish state/requirement. Even if the ‘how’ isn’t there, a ‘when done’ title indicates what is expected to complete it. This is in track with the recent post by Anthropic on what Claude opus 5.5 prefers.

LLM integrations and skills

I am trying to keep initial instruction size as small as possible. In fact, ideally it is just a short bullet point list of ‘global requirements’ + links to files if context requires them.

The project integrates:

• MCP to linear
• CLI to cloudflare
• CLI to render.com (hosting)
• CLI to upstash

Here’s a quick generated preview of the setup:

Important bits:

1. learnings - one line per fact/finding that claude had to struggle with in it’s sessions. This file is a quick search whenever it needs to do any work in order to grab ready-made solution instead of wasting tokens to solve it again. Example - if it attempts to work with cloudflare but errors with auth issues, know to ask me quickly before attempting other approaches/tools (made-up example)
2. commands - the more important skill/workflow is to /add-question - this is AI automated workflow for building the questions bank. I will have a full post on it, it’s fun! Just heads up - the last step is my eyes manually checking and approving questions anyways.
3. Some of the skills folders repeat, this screenshot is a current snapshot but have room for improvement/cleanup.

Testing

There are many automated unit tests and few e2e ones that run per PR. 600+ Go tests, 500+ client tests + 200 or so other in admin/scripting.

Further, I have created small ‘debugger’ view of the game that helps me quickly jump to game states to visually preview the work or regression test. AI uses that as well. I have a small ‘debug bar’ with quick links to jump to complete game or UI test:

This also helps narrow down the reported bugs in number and then point out specific problem in a prompt. If not, I would have to play a full game just to reach a specific spot. Which also has a timer so it’s easy to miss!

Some thoughts

I think it would be easy to mark this game as “AI Slop” by some because basically 100% of the code (minus minor edits I made on few lines) are AI generated. If this is the term sure. I don’t care honestly.

My workflow focuses on prompting what I want until I get it. And I am still not there. I am not happy yet with the visuals. There’s much more to perfect.

• Even better map design. I’ve been thinking of hand-drawing the map completely on procreate and placing it as an asset
• Rework to capital attacks as right now it’s just set of placeholder icons/text
• Better labels on the map itself
• Polishing / redesigning the home and game-start form as well as submenus/modals and more

A lot of what’s in the game is there so it’s playable. There is some design direction but it’s not final for sure. It’s an alpha!

I had to fight against AI-isms like ‘all caps’, ‘letter spacing’, ‘blinking dot’, ‘border radius’ and countless such defaults. I also have it in my plan to work on carefully written skill set that defines a brand language once I complete that too (likely to design in figma by hand, no ai there).

Claude/AI in general unlocked me to actually build what I want to play. To create.

There is some feeling of loss to the time I’ve spend perfecting my css skills, building interactive UIs and much more yes. It’s true. But I will not use that as a wall to overcome, it’s just a fond memory of the struggles I had over the years. For some this would be like throwing away a skill you’ve been building for a decade. In a way that is true. I have not really written any code on the game even though I had many inputs and comments on it. Lots of small reworks, breaking components down, reusing the right parts, changing approaches and defining rules.

This project proved to me that I can embrace this new workflow. Use all the skills I had and apply them where it fits. If it doesn’t, well, that’s progress for you - things change. I will try to be adaptive and so far it’s working well.