Good AI collaboration should not only deliver this answer; it should also preserve “how to judge next time.”
Raymond’s explanation
When asking an expert to put out a fire, fixing it once is not enough. The truly compounding approach is to ask them to leave behind four things: what state counts as done, when to ask a person, what signs mean the approach should change, and what evidence should be used for final verification.
This is externalizing judgment: writing down experience that originally existed only in Raymond’s or a higher-tier model’s mind as conditions, counterexamples, and acceptance checklists that people and AI can execute later. The point is not to hard-code every decision, but to keep the next executor from having to guess everything from scratch.
When to use it
- The same kind of task will recur, rather than being a one-time emergency fix.
- Work will be handed among multiple people, AIs, or devices.
- A mistaken judgment is costly, or the same kind of mistake has already happened repeatedly.
- A higher-tier model has just solved a difficult problem, and the next step is to give it to a lower-cost model for reliable execution.
Four minimum fields
| Field | Question to answer | Example |
|---|---|---|
| Stop | What conditions must all be met for the work to count as complete? | The file exists, behavior passes a real test, and acceptance criteria align item by item |
| Ask | When must the executor not decide independently? | Irreversible actions, conflicts between two goals, final aesthetic judgment |
| Reroute | What signals mean the current direction is wrong? | The same error recurs after two small adjustments; there is no verifiable root-cause hypothesis |
| Verify | What evidence is needed before declaring completion? | A command actually run, live status, real behavior, file and line number |
How to do it
- First ask the problem-solver to explain the real judgment points from this task, not only record the operating steps.
- Write each judgment as “situation + condition + action + counterexample,” and place it in the relevant Skill, rubric, or checklist.
- Have an executor who did not participate in the original discussion try it, and check whether two people would read it as two different procedures.
- Use actual misjudgments to revise the rule rather than continually adding another process.
Counterexamples and boundaries
- “Use a stronger model this time” is not externalization because no one will know when to upgrade next time.
- “It looks about right, so call it done” is not an acceptance rule because it offers no reproducible evidence.
- One-time aesthetic choices, public release, and irreversible decisions cannot be replaced by a checklist; a rule can only specify when to return the decision to Raymond.
- Facts about changing external conditions should not be written as permanent conclusions; rules should require “checking live status first.”
Origin
This card comes from the AI-collaboration procedures Raymond organized from 2026-07-04 to 2026-07-07: MODEL_DISPATCH defines model routing and delivery boundaries; JUDGMENT_RUBRICS turns stopping, asking, rerouting, and verification into a checklist; DELEGATION_TEMPLATES places these judgments in the assignment contract; and MAINTENANCE_PROTOCOL specifies where to write back after a misjudgment.
Where it has been discussed
Governance documents
- JUDGMENT_RUBRICS: full judgment checklist, with positive and negative examples.
- MODEL_DISPATCH: downgrade after a higher-tier model solves a pattern so a workhorse can repeat it.
- DELEGATION_TEMPLATES: fixes goals, scope, acceptance, and reporting format as the assignment interface.
- MAINTENANCE_PROTOCOL: turns repeated errors into reusable rules.
Related concept pages
- AI Tool Applications: turning AI from a one-time conversation into a handoff-ready collaboration system.
- Retrospectives: upgrading the outcome of review from “knowing what happened last time” to “knowing what rule to follow next time.”
Related concepts
Tool Neutrality and Context Compact, The Problem Analysis Method, Retrospectives, Learning Methods, Shared Upstream + Personalized Entry