C002 — Claim Definition¶


Research	R0053 — Prompt Claims
Run	2026-03-31-02
Claim	C002

Claim as Received¶

Any requirement stated to an AI without enforcement language will be treated as a suggestion — you must tell the AI what it is not allowed to do, not just what to do.

Claim as Clarified¶

This claim makes two assertions: (1) AI treats non-enforced requirements as suggestions (i.e., optional guidance), and (2) negative constraints ("must not") are necessary for reliable compliance — positive instructions alone are insufficient. This is a claim about AI instruction-following behavior and prompt engineering best practices.

Embedded assumptions: The claim assumes a binary between enforcement and non-enforcement language. It assumes negative framing is the enforcement mechanism. It conflates "enforcement language" with "negative constraints."

BLUF¶

The claim is partially correct but oversimplified. Evidence confirms that AI frequently treats weakly-stated requirements as suggestions. However, the specific prescription — that you must use negative constraints — is contradicted by research showing negative instructions ("do not") are often less effective than positive reframing. The core insight (requirements need enforcement) is sound; the specific mechanism proposed (tell AI what not to do) is not the best approach.

Scope¶

Domain: AI prompt engineering, LLM instruction following
Timeframe: Current as of March 2026
Testability: Research on AI instruction compliance, enforcement language effectiveness

Assessment Summary¶

Probability: Roughly even chance (45-55%)

Confidence: Medium

Hypothesis outcome: H2 (partially correct) prevailed. The observation that AI treats requirements as suggestions is well-supported, but the prescribed remedy (negative constraints) is contradicted by research.

[Full assessment in assessment.md.]

Status¶

Field	Value
Date created	2026-03-31
Date completed	2026-03-31
Researcher profile	None provided
Prompt version	prompt-snapshot.md (2026-03-31-02)
Revisit by	2026-09-30
Revisit trigger	New research on LLM instruction-following effectiveness; changes to RLHF training approaches