Stop Dumping Your Entire Repo Into Claude Code

• TECHNICAL FLAW: 'Model allocates attention budget across all tokens equally' is wrong. Transformers use learned, softmax-weighted attention — not equal allocation. The dilution effect is real (lost-in-the-middle research backs it), but the stated mechanism is imprecise. Fix: 'Every token competes for finite attention; irrelevant files shift weight away from your bug.'
• INVENTED STAT: 'Signal-to-noise: 3%' in s4 warning — this number is made up. Either derive it from the example (4 signal files / 214 total = 1.9%, which is actually more damning) or cut it. Fabricated precision is slop.
• STRONG: parseInvoice() → formats.ts:locale undefined → parser.ts:17 null short-circuit is a tight, believable bug trace. Specificity here earns trust.
• STRONG: 'Cut until it hurts' CTA is the best line in the piece — memorable, action-oriented, slightly uncomfortable.
• MINOR: 'Root cause identified in 1 hop' in s6 reads as a soft performance claim. It's fine but edges toward hype. Consider removing.
• s3 kinetic highlight on 'attention is finite' / 'competes' is the right emphasis — these are the load-bearing claims.
• Hook is counter-intuitive in exactly the right way. No revision needed there.

Storyboard / 45s

Script