I built HN Showcase as a weekend project in 2011 (https://news.ycombinator.com/item?id=2843490) - a thumbnail gallery for Show HN posts. It got some love (https://news.ycombinator.com/item?id=4053755, 158 pts), then went offline like all side projects do. Fourteen years later, Show HN gets hundreds of posts per day and it's harder to find the interesting stuff. So I rebuilt it from scratch. Every post gets a Playwright screenshot, then Claude Haiku analyzes the screenshot + page content + GitHub README to classify it into tiers (Gem → Banger → Solid → Mid → Pass) with a one-sentence editorial take and vibe tags like "Wizardry" or "Dark Horse." The whole rating system was built by iterating with AI — I gave it the objective (surface interesting projects, don't inflate scores) and had it design the tier rubric, pick real posts as calibration benchmarks, and tune itself until the distribution had teeth. I didn't hand-label a single example. Out of a couple thousand posts analyzed, under 1% got "Gem." Open source: https://github.com/InsipidPoint/showhn Would love feedback — are the ratings reasonably calibrated? Any projects rated unfairly? |