A research coordinator dispatches subagents with prompts like: "Step 1: search Google Scholar. Step 2: read the top 5 papers. Step 3: summarize each. Step 4: identify common themes." When Google Scholar returns sparse results that don't justify reading 5 papers, the subagent rigidly reads 5 anyway — including weak sources — because that's what the prompt said.
What's the better dispatch pattern?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
**Coordinator prompts should specify what to achieve, not how to achieve it.** Procedural step lists fail the moment reality doesn't match the steps — sparse results, paywalled sources, irrelevant top hits, all of which the rigid procedure can't adapt to. A goal-and-criteria prompt ("find the strongest 3–5 sources, evaluated on relevance, recency, methodological rigor") lets the subagent exercise judgment about how while staying accountable for what. Predictability comes from criteria, not from procedure.
There's a combinatorial explosion of edge cases (sparse results, irrelevant top hits, paywalled sources, language barriers). A procedural prompt that handles all of them becomes brittle and unreadable. Goal-and-criteria handles the long tail by ceding the procedure to the subagent.
"Use your judgment" is a non-criterion — it tells the subagent nothing about which judgment to apply. Quality criteria are the judgment guidance.
Predictability and effectiveness aren't the same thing. A rigidly procedural subagent reliably produces output that doesn't match the situation. The goal is reliable outcomes, which sometimes requires adaptive procedures.