question-economy

Every question Claude asks has a cost — it interrupts your flow (especially a focus-stealing modal in the middle of voice dictation). And every answer you give is a latent standing default: next time the same situation arises, Claude should already know the answer and just proceed, stating the choice instead of asking.

*The goal is not zero questions — it's zero avoidable questions.* Genuinely novel decisions that need your judgment still deserve a question. What should decay toward zero is re-asking things you've effectively already answered. Run this every session and the avoidable-question count should trend down, session over session.

Steps

1. Extract the question→answer ledger. Run the extractor on this session's transcript:

   python3 question_economy.py --print-trend

It prints JSON: hard_questions (AskUserQuestion modals — the costliest), soft_questions (an assistant turn ending in ? + your next reply), counts, and trend (the last 8 sessions' question rates). It also appends one metric row to question_economy_log.jsonl in your memory dir.

1. Classify each pair. For every question→answer, decide:

• Standing default — your answer generalizes ("whenever X, do Y"). → becomes a memory.
• One-off — context-specific, doesn't generalize. → skip.
• Avoidable — Claude could have inferred the answer from existing memory or a sensible default and shouldn't have asked. → flag it; this is the number to drive down.
• Good question — genuinely needed your judgment. → leave it; not every question is waste.

1. Write the new defaults to memory. For each standing default, write a memory entry that leads with the generalized rule (not the one instance), followed by Why: and How to apply: lines. If your setup curates memories through an inbox/review step, write proposals there rather than straight to live memory. Check your existing memories first — update a related one instead of duplicating.

1. Report the economy. One short readout:

• This session: N hard + M soft questions; K were avoidable (and why).
• New standing defaults learned → so Claude won't ask these again: (list).
• The trend: this session vs the last few — is the avoidable count actually falling?
• If the rate is NOT falling, say so plainly and name what's still being re-asked — a signal a memory isn't being honored or is mis-scoped.

The loop it serves

Question asked → you answer → answer distilled to a standing default → memory → next session Claude applies the default instead of asking → avoidable-question count falls. The question_economy_log.jsonl metric is the proof; the memories are the mechanism. Over many sessions this converges Claude onto your working style, so the conversation is about novel decisions, not re-litigating known preferences.

Notes

• Hard (modal) questions are weighted heaviest — a modal mid-dictation is the worst interruption. If the log shows hard questions not falling, that's the priority fix.
• Distinguish workflow defaults (how you want work done) from technical defaults (architecture/tooling choices you keep making the same way) — capture both, tagged, so the memories stay findable.
• Complements broad session-close memory capture; this skill is narrow on purpose: the Q&A ledger and the question-rate metric.

Requirements

Python 3.9+. No dependencies — standard library only. Reads Claude Code session transcripts from ~/.claude/projects/*/.