Every line from Anthropic's Opus 5.5 prompting guide, word for word, with where to paste it and what actually happened when we ran it

The companion to the video. Anthropic published 'Prompting Claude Opus 5.5' and most of it says delete. We ran the five lines you can actually paste on Opus 5.5 itself, at least six runs each. INSIDE: (1) FIX 1, GOLD — delete 'think carefully' and set effort on purpose: time to Claude's first word went from 5.5s to 2.6s at medium and 0.7s at low, same answer 12 out of 12, 21% to 45% cheaper. (2) FIX 2, GOLD — Anthropic's named design ban list, word for word, plus our round-two line: websites went from 0 of 6 clean to 3 of 3, then a completely different look. (3) FIX 3, SMALL WIN — the settled-answer line for chat apps (follow-ups 24% shorter) and Anthropic's warning. (4) FIXES 4 AND 5 — explore-before-acting and the don't-stop paragraph: why they made no difference in our tests and when you still want them, plus Anthropic's simpler nudge. (5) BONUS — pasted-content tags, the /claude-api prompt-audit command, and Boris Cherny on deleting 80% of Claude Code's prompt. Every line quoted from Anthropic's docs; every number from our runs. Independent sheet from Hyperautomation Labs — not affiliated with Anthropic.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — three practical AI emails a week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.