1. Define “good” before you build. Beginner output comes from vague targets. Before a project, collect 3–5 concrete refe
2. Go narrow and deep. Impressive projects are usually one small thing polished far past what’s reasonable. A single button that feels incredible beats a whole
Saved context
1. Define “good” before you build. Beginner output comes from vague targets. Before a project, collect 3–5 concrete references (a specific awwwards site, a specific game’s feel) and write down exactly what makes them good: timing, sound, easing, feedback. The AI can only hit a target as specific as the one you give it. 2. Go narrow and deep. Impressive projects are usually one small thing polished far past what’s reasonable. A single button that feels incredible beats a whole site that’s okay. This fits your button-reel idea well. 3. Let the AI see what it made. It can’t feel its own output. Close the loop with screenshots, screen recordings, or Playwright captures fed back in for critique. Your zero-context review already works; aim it at the actual experience too, not just the code. 4. Turn retrospectives into rules. This is where your system will compound. A lesson that lives in a doc only helps if you remember it. A lesson turned into a skill, checklist, or project instruction file gets applied automatically every time. If a mistake shows up in two retros, it becomes a rule. 5. Layer the craft passes separately. Build it working first, then do dedicated passes for motion, then sound, then micro-feedback (game devs call this “juice”). Doing it all in one pass gets you average at everything. 6. Benchmark your process, not just your projects. Every couple of cycles, rebuild the same tiny thing, like one scene or one card, with your current process. Different projects vary too much to tell whether you’re improving. The same one shows it clearly. 7. Reverse-engineer what impresses you. When a post blows you away, find their process thread or ask them. Many share their tools, prompt count, and hours. That’s the fastest way to learn what’s actually behind the result, including how many attempts failed first. Opus 5.5 advice on making better projects.