Anthropic's own lines for Claude Fable 5.1, verbatim, with where to paste each one and what it changed when we tested it
The companion to the video. On September 1, 2026 Anthropic published a 16-section guide called Prompting Claude Fable 5.1. This sheet is every paste line from that guide, character for character (the lines are pulled from the doc itself, not retyped), with where each one goes (CLAUDE.md, the top or end of your message, the system prompt, or the end of the user message) and what happened when we pasted it into Fable 5.1 inside Claude Code across 26 real runs. THE 7 FIXES WE TESTED: (1) the progress-updates line that ended a 170-second silence, (2) effort as a setting, with our table showing the same 27/27 test result at every level from 27.4s and $0.35 at low to 388.2s and $2.49 at xhigh, (3) the batching nudge that goes at the end of the request (1 to 4 parallel tool turns), (4) 'Please remove all mannered prose' plus the long version, with the honest note that it invented statistics once so give it the numbers, (5) the search-the-name-as-written line for low effort, (6) both finish-the-whole-task blocks and (7) the keep-changes-and-tests-to-the-task line, the last two marked honestly as insurance because they did not reproduce on our small tasks. THEN THE REST OF THE DOC for API builders: the formatting-in-chat rule, the quoting example, the compaction summary instruction, the targeted-edits line, the long-outputs note with the [max_tokens] placeholder, the three refusal false-positive phrasing tips, the append-only history rule, the subagent setup, and the vision crop tool. Honesty box included: tested inside Claude Code headless with a replaced system prompt, 1 to 2 runs per arm, costs are list-price estimates. Independent test sheet from Hyperautomation Labs; not affiliated with Anthropic.
Free. No spam. Unsubscribe anytime.