AIMM Hub

Mastermind recap

Context Hygiene and the Handoff Skill

· AIMM · 90 min

Facilitators: Lou D'Alo

AI Mastermind | Knowledge Entrepreneurs Edition

“We have a very highly informed resource, but not a very smart one to make judgments. It’s very eager to be helpful. If you want to hold it back, that’s where our judgment comes in.” — Lou

This Week in 30 Seconds

  • The deliverable is the folder, not the article — Lou is shipping his entire working folder (chats, skills, articles, .claude/ agents) so members can replay the messy middle, not just consume the polished output.
  • Three layers of memory — Kasimir’s three-skill Obsidian stack and Lou’s hot-cache → wiki → semantic vector hierarchy give a working blueprint for a personal knowledge base that scales past 900 entries.
  • Context hygiene at 50% — Lou flagged the emerging research that Claude’s reliability drops past half-full context. The /handoff skill is the antidote; the /goal command is the bound.
  • Constraints, not better prompts, end the improvement spiral — Dirk surfaced the trap every operator falls into. Lou’s answer: audit prompts, ROI stop-conditions, and a target-persona critic.
  • Skill vs. MCP vs. API — Three layers for exposing your expertise to clients’ AI. The skill leaks your IP; the API hides it; the MCP translates. Design APIs for agents, not humans.

Context Hygiene: The 50% Rule

Scott discovered the technique organically months before the skill existed. “This chat is getting long, and there’s a new effort to base off of this — summarize this chat for a new chat in this same project.” It works beautifully for serial work on a single thread. It breaks down when, as Don Back described, you’ve spawned six parallel chats off a central theme and now need to reconcile them. “It’s just a mess, Lou.”

Lou’s answer is the /handoff skill in the AIMM GitHub. It summarizes the current context to a file, clears the conversation, and points the new conversation at the file. The trigger is not “when the chat feels long” — it’s a structural threshold. The emerging research Lou cited puts Claude’s reliability degradation at around 50% context fill. Past that, instruction-following starts to slip. The discipline is to compact before you feel the degradation, not after.

💡 What This Means for You

Pick one chat you have open right now that’s drifted past three rabbit holes. Don’t try to salvage the thread. Run the handoff: ask Claude to summarize “what we actually decided and what’s still open” into a fresh markdown file, then start a new chat with that file as the only input. The conversation you’ve been having for two days will compress into ten minutes of grounded work.

Next session: Thursday, May 28, 2026