AIMM Hub

Mastermind recap

Haiku-First, Sonnet-Mostly, Opus-Rarely

· AIMM · 90 min

Facilitators: Lou D'Alo

AI Mastermind | Knowledge Entrepreneurs Edition

This Week in 30 Seconds

  • Lou shaved 11TB off his life — Claude wrote a self-checking dedupe script and turned a two-year-postponed drive cleanup into a two-day background job. The tasks you keep avoiding are the ones to hand to Claude.
  • Killing the $100/month content tax — Lou’s new Cloudflare + GitHub + Astro + Quartz stack replaces three SaaS platforms with one repo and one git push. Member-gated access via Cloudflare magic-link.
  • Donald turned every Claude session into a language tutor — A global CLAUDE.md skill that captures spontaneous mistakes mid-conversation and writes a daily lesson file. Works for any language.
  • Joanna’s tax avalanche became a live workflow demo — 900 receipts, returning member, and a four-step reset: scan only cash receipts, pull bank CSVs, point Claude Code at a folder (never the conversation), spawn parallel agents.
  • Haiku-first, Sonnet-mostly, Opus-rarely — The room has converged on a model-selection discipline that inverts the default. Plus Kasimir’s rewind discipline: don’t correct Claude forward, rewind and re-instruct.

Haiku-First, Sonnet-Mostly, Opus-Rarely

A model-selection discipline crystallized across the call that contradicts the default. The default is “use the best model.” The room has converged on something sharper: match the model to the cognitive load, not the importance.

Scott Delinger named the recurring mistake. He’d been doing his receipt-extraction in Sonnet and burning through limits. He should have started in Haiku — “very unlikely you’re going to hit limits in Haiku at the top speed.” For batch parsing, image extraction, format conversion, and other mechanical work, Haiku is the right tool. Reserve Sonnet for the work that actually involves judgment. Scott does Python coding against 90M-record datasets in Sonnet because “the Python does a lot of the heavy lifting for me” — the model isn’t the bottleneck, the code is.

Don Back’s rule of thumb: “Opus only for the deepest tasks.” Most members had been treating Opus as the default and hitting limits within hours.

Scott’s complementary discipline: “I start in chat, because then I can review what it’s doing. Rather than doing a batch upload in Cowork, and then finding out that I wasted half a day’s worth of tokens because my approach was wrong.” Validate in chat, then promote to batch.

Kasimir’s rewind move is the discipline that pairs with the model choice. Token efficiency isn’t just which model you pick — it’s how you recover from the AI going sideways. “If Claude does something that you don’t want it to do — don’t correct it. There is a rewind. It will delete everything from the context that is below that.” Most users dig the hole deeper by trying to course-correct in the same thread. The smarter move is to back up to the last good fork and re-instruct.

💡 What This Means for You

Today, before your next Claude session: open your settings, look at which model is set as your default, and ask whether you’re paying the Opus tax on Haiku work. Then practice the rewind keystroke twice on a throwaway conversation so the muscle memory is there before you need it. Both moves are free. Both compound.

Next session: Thursday, May 21, 2026