Skip to content

R0054/2026-03-31/C002/SRC01/E01

Research R0054 — Prompt Claims v2
Run 2026-03-31
Claim C002
Source SRC01
Evidence SRC01-E01
Type Reported

Research-backed recommendation for 80/20 ratio of positive to negative instructions in AI prompts.

URL: https://www.vibesparking.com/en/blog/ai/prompt-engineering/2025-08-14-prompt-do-not-to-do-playbook/

Extract

Key findings from the search results summary:

  • An empirical ratio of 80% positive to 20% negative instructions is recommended, "prioritizing positive guidance first by telling the AI what it should do"
  • Research from KAIST found that "larger AI models actually perform worse on negated prompts — instructions telling them what not to do"
  • Converting "don'ts" into "do's" is recommended: replace "don't write ungrammatical sentences" with "Self-check grammar and cohesion sentence by sentence and revise to fluency"
  • "Models have systematic weaknesses in handling negation; positive instructions reduce ambiguity and off-topic drift"
  • Despite this, negative constraints serve a distinct function: they establish hard boundaries that positive instructions cannot replicate

Relevance to Hypotheses

Hypothesis Relationship Strength
H1 Supports Directly confirms that positive alone is insufficient — the 80/20 ratio implies both types are needed
H2 N/A Does not suggest constraints are merely helpful rather than necessary
H3 Contradicts Directly contradicts the idea that positive instructions alone suffice

Context

The 80/20 ratio recommendation suggests that while positive instructions should dominate, negative constraints are a necessary complement — not optional. This aligns with the claim's assertion that positive instructions alone produced inconsistent results.