R0020/2026-03-25/Q003/H3¶
Statement¶
Guides recommend clarity and specificity in ways that encompass constraint language, and some explicitly demonstrate imperative directives, but the field is actively evolving — with the latest guidance from major vendors suggesting that aggressive enforcement language may be less necessary with newer models, favoring context and explanation over imperatives.
Status¶
Current: Supported
This hypothesis best captures the complex, evolving picture. Constraint language is documented and demonstrated, but Anthropic's latest guidance explicitly recommends moving from "CRITICAL: You MUST use this tool" to "Use this tool when..." — suggesting that imperative constraints are a current practice that may be transitioning toward more nuanced instruction styles as models improve.
Supporting Evidence¶
| Evidence | Summary |
|---|---|
| SRC01-E01 | Anthropic uses constraint examples but frames them within context-giving |
| SRC01-E02 | Explicit advice to replace "CRITICAL: You MUST" with normal prompting |
| SRC02-E01 | Constraint-based design recommended but evolving toward prompt scaffolding |
Contradicting Evidence¶
| Evidence | Summary |
|---|---|
| SRC03-E01 | Some industry sources still strongly recommend imperative language |
Reasoning¶
The evidence reveals an inflection point. Pre-2026 guidance heavily recommended imperative constraints ("MUST," "NEVER," "ALWAYS"). Post-2026 guidance from Anthropic suggests newer models respond better to contextual explanation than to enforcement language. This does not mean constraints are being abandoned — it means the implementation is evolving from imperative (MUST) to explanatory (here's why, and what to do).
The most diagnostic finding is Anthropic's specific guidance: "Where you might have said 'CRITICAL: You MUST use this tool when...', you can use more normal prompting like 'Use this tool when...'" This directly describes the transition from imperative to explanatory constraints.
Relationship to Other Hypotheses¶
H3 synthesizes the valid elements of H1 (constraints are discussed) with the evolutionary trend that neither H1 nor H2 captures — the field is moving from enforcement to guidance.