I’ve been experimenting with building side-projects where AI isn’t just “helpful”—it’s doing most of the heavy lifting.
One recent project made me pause: I got a working prototype up in under 10 days, barely touching the keyboard myself.Here’s what I’ve learned (my words, my scars 👇):⸻1️⃣ Talk in Blueprints, Not SnippetsAI struggles with half-baked prompts.
Instead of “write me a login page,” I start with the bigger map—what the system is, the entities, the flows—then zoom in.
It’s like explaining the whole building before asking it to draft a single room.⸻2️⃣ Turn Errors into GuidanceRather than fixing bugs line-by-line, I paste the failing output or stack trace and say: “make this work.”
The AI doesn’t just fix it—it learns what “working” means in my context.⸻3️⃣ Codify Your Preferences EarlyA tiny set of “project principles” (naming rules, API conventions, version choices) saves hours later.
Think of it as training wheels for the AI—it stops it from wobbling all over the place.⸻4️⃣ Stay in One UniverseKeeping everything—UI, API, docs, infra—in a single repo makes AI exponentially more useful.
Context lives where the code lives. Split things apart, and you cut its brain in half.⸻5️⃣ Tools Are Just Signal BoostersI used fewer integrations than I expected. The right three tools felt like superpowers.
The wrong ten tools would’ve been noise.⸻6️⃣ Models Are Like TeammatesSome are better planners, others better doers.
I’ve stopped treating them like one magical brain and started treating them like a squad.⸻7️⃣ Feed It Quality, Get Quality BackGarbage in, garbage out never felt so real.
When I refactor before handing code to AI, the results multiply in quality. When I don’t… chaos.⸻The takeaway?
AI isn’t here to replace me—it’s here to multiply what I can do in short bursts of focus.
The faster I adapt my habits, the more I can ship without burning out.
