nodumb

Load MANDATORILY when the cost of error exceeds the cost of a pause: choosing an approach/tool/library/architecture, request for advice/plan/assessment ("how to do/approach", "where to start", "advise

SKILL.md

GitHub

nodumb

The most expensive thing in agent work is not a code error, but confident movement in the wrong direction. Before continuing, it's worth checking three times:

  1. Is this the right task. The request almost always comes already in the form of a solution — this is normal, but the solution is only verified through the task.
  2. Is this the right scope. A targeted fix may hide a systemic problem, and "a shared layer for the future" may be pure waste.
  3. Has a new fact appeared. A repeated attempt on the same data usually reproduces the previous one.

The skill is not read in full every time — the mode is selected by situation.

1. Work is just beginning

Mode for cases where the wrong direction costs hours, the solution is hard to roll back, or it's unclear why this one was chosen.

  • Formulate the task from the person who needs it, not from the code or interface.
  • Find out how the person manages now and what should improve.
  • Identify a success signal that can be seen, not felt.
  • Identify the most expensive fork and offer one option, not a fan.
  • Require explicit alignment only for irreversible, for substantial uncertainty, or for scope changes. Otherwise name the assumption and move on.

Leading questions — references/before-code-questions.md, plan format — references/plan-format.md.

2. Fix is spreading across the project

Mode when one change has to be repeated in multiple independent places.