Get The Fable 5 Loop Recipe

One week after "stop prompting, start designing loops" went viral with no definition attached, an engineer inside Anthropic — Lance Martin (@RLanceMartin, ex-LangChain) — published the actual recipe, with experiments attached. This guide is the decode: two loops plus one rule. Loop 1, self-correction inside the session — the real /goal setup (Claude Code v2.1.139+: you write a checkable finish line and a separate small model, Haiku by default, checks it after every turn until it passes) and its API twin, Outcomes in Claude Managed Agents (a grader agent in an independent context window). Loop 2, memory across sessions — the 5-rung memory ladder (fail, investigate, verify, distill, consult) with the receipts from Continual Learning Bench: Sonnet 4.6 exits at rung 1, Opus 4.7 exits at rung 3 verifying only 7-33% of its answers, and Fable 5 tends to complete the ladder (73% verification in its strongest runs). The one rule that makes both loops trustworthy: the maker is never the grader. Inside: the two-loop diagram, the /goal setup with copy-paste finish lines, a copy-paste verifier rubric and skeptical-grader subagent prompt, the memory-entry format (FAILED / WHY / VERIFIED / RULE), the Parameter Golf receipt (Fable 5 improved a training pipeline ~6x more than Opus 4.7 — on OpenAI's own challenge), and the four honest catches the post does not put in bold, including the meter: a self-correcting loop is the most expensive way to use an already expensive model, so design the loop and cap it.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.