Developers make Claude talk like a caveman to save tokens. I made it talk like a status report to save my attention.
Working with AI agents can dramatically increase your cognitive load, especially if you run a few agents in parallel, as I do.
I don't use the caveman plugin, and I don't ask Claude to write in ASD-STE100, the Simplified Technical English that Karpathy suggests. To simplify my life, I asked Claude, my primary agent, to save a global CLAUDE.md rule:
## Reporting completed work
- One block per piece of work: title + status (✅/⚠️/❌), one-line result, "Your step:" (or "none").
- Details (tables, reasoning, caveats) only on request. "Watch": brief, only non-actions.
- End with "What are the next steps?": numbered decisions/actions for me, each with your recommendation, so I can answer "1 yes, 2 no".
- Carry forward every unanswered question from earlier turns until I answer it; a skipped item is not a "no" and not "unimportant".
- Next steps = only what's decided or done today. Don't list items scheduled for tomorrow or later; record them as a dated reminder (memory or `bd defer --until`) and raise them at the start of the session on that day.
- Questions still open when the session ends: if the project uses beads (`bd`), record the ones that affect work (decisions, not minor clarifications) as beads so the next session picks them up; otherwise list them explicitly in the final message.