What do you do when your amazing community members submit 14,715 high res images to you during a mega event? 😱 Build a fast image viewer with @claudeai and @rustlang of course! 🧠 github.com/bitcoin-verse/Sup… Claude learning points 🧵👇 1/12 The community event that started it all. 14,715 images, 2.5k+ contributing members: t.me/GetVerse/177601/392398 Just a fraction of it:

2/12 Tried using @windows default image viewer. UI flashes. Loading screen. It's already 2026 guys. I just can't.... 🤮

3/12 @claudeai Team plan to the rescue. One of the many amazing benefits of being at @BitcoinCom 🥰

4/12 Main Prompts: 1. I have 14k images of different formats. The default windows image viewer is too slow. Find out what makes a good image viewer for windows. Focus on the best practices to optimize for speed and startup time. Draft a PRD. 2. Implement the MVP, test and fix any bugs, make sure it works then continue with the rest. 3. Find out who are the competitors in this space. propose how to benchmark our app against theirs then run it. 4. Using the results from the benchmark, think hard and propose a plan how optimize for startup time and for viewing images quickly. implement the plan. 5. Write a readme, a retroactive, and learning points on how to use claude code better.

5/12 Key learning points for new claude users: A. Start with a PRD, not code requests. The very first prompt produced a product requirements document: goals, non-goals, architecture, module breakdown, success metrics. Every subsequent prompt had a frame of reference. Without the PRD, each prompt would have needed more context and there would have been more back-and-forth about scope.

6/12 B. Ask questions before implementation "What happens if we reduce cache to 128 MB?" informed a design decision without touching code. Use Claude as an analytical tool, not just a code generator.

7/12 C. Benchmark-driven optimization, not intuition. Navigation went from 92ms to 63ms through 5 iterations of measure→hypothesize→implement→measure. The biggest surprise: moving resize to worker threads made things *worse*. Without benchmarks, that "obvious improvement" would have shipped.

8/12 D. External edits break Claude's model. I'm GUILTY. Part of the README was rewritten outside Claude, and Claude didn't know until told to re-read. Two prompts were spent rediscovering and re-applying a change. Lesson: if you edit files manually mid-session, mention it.

9/12 E. The highest-ROI prompts were the boring ones. The installer script, the test checklist, the benchmark suite: none of these are exciting code. But they're what turned a prototype into a shippable product. Claude is especially good at this kind of thorough, systematic work that humans tend to skip or rush.