R023 · llm-actionability

LLM: instructions executable


What	LLM-tier: instructions should be executable — the agent can act without inventing missing steps or APIs.
Why	Trivia does not ship work. Over-specified micro-workflows also fail when reality diverges from a few test prompts.

How to fix

Lead with what success looks like, then constraints.
Give one concrete example (R017).
Explain why a constraint generalizes (“model Y is deprecated and returns errors”).
Cover failures in scenarios.yaml (R027).
Fragile multi-step order → script, not a long procedure.