C002 — Claim Definition¶


Research	R0054 — Prompt Claims v2
Run	2026-03-31
Claim	C002

Claim as Received¶

Descriptive guidance alone — telling the AI what to do — is not sufficient for complex, multi-step analytical processes. Detailed positive instructions produced inconsistent results until complemented with explicit constraints on what the AI could not do.

Claim as Clarified¶

This claim asserts: (1) positive instructions alone ("do X") are insufficient for complex multi-step AI processes, (2) adding negative constraints ("do not do Y") was necessary to achieve consistency, and (3) this complementary relationship between positive and negative instructions is a real and documented phenomenon. The claim contains an embedded assumption that complex multi-step processes require different prompting strategies than simple tasks.

BLUF¶

The claim is well-supported by current research on prompt engineering and LLM behavior. Multiple studies confirm that positive instructions alone produce inconsistent results for complex tasks, that LLMs have documented weaknesses in processing negation, and that the most effective approach combines positive guidance with carefully structured negative constraints. The 80/20 ratio (positive to negative) recommended by practitioners aligns with the claim's description of complementary instruction types.

Scope¶

Domain: AI prompt engineering, LLM behavior
Timeframe: 2024-2026
Testability: Verifiable through prompt engineering research, practitioner experience, and LLM behavioral studies

Assessment Summary¶

Probability: Very likely / Highly probable (80-95%)

Confidence: Medium-High

Hypothesis outcome: H1 (claim accurate) prevailed, with strong support from both academic research and practitioner experience. The underlying mechanism (RLHF-trained helpfulness overriding instruction compliance) is well-documented.

[Full assessment in assessment.md.]

Status¶

Field	Value
Date created	2026-03-31
Date completed	2026-03-31
Researcher profile	Phil Moore
Prompt version	ai-research-methodology v1 research.md
Revisit by	2027-03-31
Revisit trigger	Publication of systematic studies specifically comparing positive-only vs. positive+negative prompting for multi-step analytical tasks