Write agent actions as procedures a stranger could follow — aspirations and principles are not executable steps

Why This Is a Rule

The action component of a behavioral agent fails most often at the specificity boundary. "Practice mindfulness" is an aspiration. "Be more present during conversations" is a principle. "When the trigger fires, close your laptop lid, take three breaths, and write one sentence about what you were just thinking" is a procedure. Only the procedure is executable without real-time interpretation — and real-time interpretation is exactly what fails under cognitive load, stress, and fatigue.

The "stranger test" is the quality gate: could someone who has never seen this agent before execute the action steps without asking clarifying questions? If the steps require domain knowledge, contextual judgment, or interpretation of ambiguous terms, they're too vague. The trigger already handled "when" and the conditions handled "whether" — the actions must handle "what, specifically, step by step."

This matters because the whole point of a behavioral agent is to eliminate deliberation at the moment of action. If the action steps require deliberation to execute ("what does 'practice mindfulness' actually mean right now?"), the agent hasn't offloaded the decision — it's just deferred it to a moment when your executive function may be even more depleted.

When This Fires

• When writing the Actions component of any agent documentation (Document every agent with five components: Name, Trigger, Conditions, Actions, Success Criteria — undocumented agents degrade silently)
• When reviewing existing agent documentation for executability
• When an agent fires but the action step "doesn't feel clear enough to do" in the moment
• When coaching someone whose behavior change commitments are vague

Common Failure Mode

Writing principles instead of procedures: "Engage deeply with the material" instead of "Read one page. Write one sentence summarizing the main claim. Write one question the claim raises." The principle sounds good in planning mode but is unexecutable in execution mode because "deeply" and "engage" require real-time judgment that the agent should have pre-committed.

The Protocol

(1) Write the action steps in numbered order. (2) For each step, apply the stranger test: could someone unfamiliar with your context execute this step without asking you what you mean? (3) If any step fails the stranger test, decompose it further or replace abstract terms with concrete ones. "Review your goals" → "Open the goals document, read each goal aloud, and rate current progress 1-5." (4) Include physical actions where possible: "open," "write," "say," "move to" — these are unambiguous. (5) Test by executing the steps literally, following only what's written. Where you fill in gaps from memory, the documentation is incomplete.