I stopped letting AI write my code blindly. Here is the framework I use instead.

If you missed my first post, I got tired of not understanding my CTO, so I built a game to learn tech jargon, I recently built TechLexicon.app, a gamified app to help developers master technical jargon.

I built the app from scratch using Antigravity IDE starting on December 26th. Antigravity is an incredibly powerful engine—it has terminal access, can run commands, create files autonomously, and allows you to set skills to thwart prompt over-bloating. To be completely fair, I am sure I wasn't even maximizing all of its features at first.

The issue was my own fragmented workflow.

I found that the experience of high-level planning and brainstorming was just more powerful and acute using Claude Pro on the web. So, my workflow became: Brainstorm, plan, and move through my Implementation Roadmap in Claude Pro ➡️ Copy and paste the decisions and code into Antigravity IDE to execute.

But this created the ultimate developer nightmare: Context Syncing.

Even though I manually uploaded my "Source of Truth" docs to a Claude Project every week, the disconnected workflow meant the AI was constantly losing context, losing information, hallucinating, and making the exact same mistakes over and over again.

The breaking point happened while building my real-time 1v1 Speed Round game. The game was suffering from a brutal React Native bug where the component was remounting 30+ times per session, prematurely ending the game timers.

I realized that if I let the AI just blindly rewrite my code without context, it was going to break the entire battle system. I had to step in.

I explicitly prompted the AI to operate as a Senior Engineer, and I took on the role of CEO. I stopped asking for quick code fixes and instead forced it to follow a strict "4 Mandatory Pre-Fix Requirements" framework.

Claude followed the framework and spit out a code fix. Most developers would have just copied and pasted it into their IDE. But because I was operating as the CEO, I looked at the fix and realized it felt wrong. So, I pushed back and challenged the AI to verify its own work and check the web to ensure it wouldn't break my other game modes.

And this is exactly why you never blindly copy-paste. Because I challenged it, Claude actually stopped, researched the React Navigation documentation, and realized its first fix was dangerous. It admitted fault and came back with the actual, researched solution.

While my strict prompting framework and pushback strategy saved my app, the manual copy-pasting to sync my browser and IDE was still a massive drag on my productivity.

Within the last week or so, I finally made the jump to using Claude Code (the CLI tool) directly in my terminal.

It completely eliminated my manual syncing workflow. But instead of just chatting with it, I took my strict "CEO/CTO prompt" from the web and broke it down into modular custom skills.

Here is the exact skill framework I built to manage my codebase:

By standardizing my habit into a strict loop (/session-start → /preflight before changes → /session-close), I finally have an AI that documents, edits, and maintains my "Source of Truth" directly in my workspace without the copy-paste bottlenecks.

You can test out the newly stabilized 1v1 Speed Round and the rest of the app at techlexicon.app.

I'm curious about your workflows: For those of you bootstrapping apps with AI right now, how are you handling context loss? Have you started building custom CLI skills yet? Let me know below!