Feedback from running PAI/LifeOS day to day.
I added a TRIAGE GATE at the top of my MODES section in CLAUDE.md. Before any mode or tool selection, it classifies the incoming message as TRIVIAL / SIMPLE / COMPLEX, with a hard rule: TRIVIAL (greetings, acknowledgments, a bare rating) → reply in one sentence, zero tools, then stop. Plus a per-bucket tool budget and an explicit stop condition.
Why: when I swap the DA onto a non-frontier model, trivial prompts like "are you there" made it over-execute — it would grind the whole Algorithm pipeline and fire a cascade of tool/hook calls instead of just answering. Frontier models (e.g. Opus) short-circuit that implicitly; weaker models don't, and the existing mode rules assume that judgment. Making the triage explicit and mechanical fixed the over-execution and made mode selection robust across model tiers.
Might be worth considering for core — a capability-robust front gate so the scaffolding degrades gracefully on smaller models.
For reference, the gate I'm using:
## TRIAGE GATE (runs FIRST, before any mode or tool call)
- TRIVIAL — greeting / check-in / acknowledgment / bare rating → one sentence, zero tools, stop.
- SIMPLE — answerable now or a single reversible action → NATIVE mode, minimal tools.
- COMPLEX — multi-step build/debug/refactor/investigate → ALGORITHM mode.
Tool budget: TRIVIAL 0 tools; SIMPLE stop when answerable; COMPLEX only calls that change the next step.
Stop condition: when done, end the turn — no re-running checks to fill space.
Thanks for PAI/LifeOS.
— Jim (@jshingler)
Feedback from running PAI/LifeOS day to day.
I added a TRIAGE GATE at the top of my MODES section in
CLAUDE.md. Before any mode or tool selection, it classifies the incoming message as TRIVIAL / SIMPLE / COMPLEX, with a hard rule: TRIVIAL (greetings, acknowledgments, a bare rating) → reply in one sentence, zero tools, then stop. Plus a per-bucket tool budget and an explicit stop condition.Why: when I swap the DA onto a non-frontier model, trivial prompts like "are you there" made it over-execute — it would grind the whole Algorithm pipeline and fire a cascade of tool/hook calls instead of just answering. Frontier models (e.g. Opus) short-circuit that implicitly; weaker models don't, and the existing mode rules assume that judgment. Making the triage explicit and mechanical fixed the over-execution and made mode selection robust across model tiers.
Might be worth considering for core — a capability-robust front gate so the scaffolding degrades gracefully on smaller models.
For reference, the gate I'm using:
Thanks for PAI/LifeOS.
— Jim (@jshingler)