read-trans: Tanglish explainer for YouTube transcripts, with eval

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

tests/evals/read_trans/fixture_make_it_better.txt

@@ -0,0 +1,15 @@
+0:00 0 seconds Today I gave Claude Code a horrible coffee shop website and told it make it better, 100 times in a row.
+0:15 15 seconds I did this with Sonnet 5.5 on medium effort and Opus 5.5 on medium effort.
+Chapter 2: The prompt
+0:25 25 seconds The prompt was: look at the site first, then innovate. Think about user experience, think outside the box. Make one bold change. You must ship a change every time.
+0:50 50 seconds I had to force it to change every time because earlier it stopped at round 20 or 25 and said nothing else can be done.
+Chapter 3: Setup
+1:20 1 minute, 20 seconds Every round ran in a clean room: fresh session, no skills, no CLAUDE.md, no plugins, no memory. Only the site code carried over from the last round.
+Chapter 4: Sonnet
+2:10 2 minutes, 10 seconds Sonnet cost about $22 and ran almost 2 hours. Early rounds were good: sticky order bar, receipt style order, an AI video, an order form. After round 32 there were no big milestones. Lines of code kept rising linearly but quality did not.
+Chapter 5: Opus
+7:40 7 minutes, 40 seconds Opus cost $137 and ran 8.5 hours. Round 1 was already better. It added a QR code that texts your order, and a mood-based drink picker. By round 100 it had 21,000 lines: leaves, pumpkin carving, caffeine tracker, music, a Spanish toggle that did not work. Some sections broke. It was AI slop.
+Chapter 6: Surprises
+12:30 12 minutes, 30 seconds Neither model used web search, even though I said innovate. They never looked at award winning sites for inspiration.
+Chapter 7: Final thoughts
+13:15 13 minutes, 15 seconds The model sets the ceiling and the prompt sets the direction. Innovate without a vision turns into adding. It tried to upsell everything, a hard sell instead of a soft close. Please like and subscribe.