Summary

The agent should proactively capture emsenn’s statements, corrections, and ideas as plans, policies, or decisions — not wait for explicit instructions to do so. This is both a process improvement and a policy amendment.

Motivation

emsenn’s correction: “If you reviewed this session I bet you could find like 50 plans that you currently are making me repeat in various contexts.”

The current interpret-message skill classifies messages and encodes insights, but it does not aggressively mine every statement for plan-worthy content. When emsenn says “X should happen,” that IS a plan. When emsenn says “don’t do Y,” that IS a policy. The agent should not require emsenn to explicitly say “make a plan for this.”

This is the meta-governance problem: the planning system exists, but the agent doesn’t use it aggressively enough. Every conversation should produce plans proportional to the ideas expressed.

Steps

  1. Amend policy 008 (prompts become artifacts) to explicitly include plans: “Every prompt that describes work to be done produces a plan before (or instead of) doing the work.”
  2. Update the interpret-message skill to:
    • Scan every message for imperative statements (“X should…”, “we need to…”, “make sure to…”, “don’t forget…”)
    • For each, create a plan (or add to an existing plan)
    • Report back what plans were created/updated
  3. Write a mine-conversation skill that reviews a full session transcript and extracts all uncaptured plans, policies, and decisions.
  4. Run the mine-conversation skill on this session as a test case.
  5. Add a session-end checklist: before ending, review what was said and ensure everything plan-worthy is captured.

Done when

  • Policy 008 amended to cover plans
  • interpret-message skill scans for imperative statements
  • mine-conversation skill exists
  • This session’s uncaptured items are all captured
  • Session-end checklist documented

Dependencies

None — this is the most important process improvement.

Log

2026-03-07 — Created. emsenn explicitly frustrated that the agent makes them repeat ideas across contexts. The mining agent found ~20 uncaptured plan-worthy items from this session alone.