Zenaique

You're tuning a customer service LLM to stop suggesting refunds in cases where refunds aren't allowed. Which directive is more reliably effective?

MCQ·Medium·4.0 · 0·~1 min·Asked atWhylabsWorkdayZepto·Relevant atAnthropicMicrosoftOpenAI
Attempt it
TL;DR

Positive 'do this' beats negative 'don't do that' because the model has an action to execute instead of an action to suppress.

Memory aid
Sign in to see the mnemonic that makes this stick.
Easy to grasp

Imagine telling a toddler 'don't think about a red balloon.' What do they think about? Models behave the same way when you say 'do not suggest refunds.' To follow the rule, the model has to model the prohibited action first, then try to suppress it. Sometimes it works, sometimes the prohibited thing slips through. The fix is to give the model something to do instead. 'For refund requests outside policy, redirect the user to the support portal.' Now there is a clear action attached to the trigger, and the model just executes it. No suppression required.

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Positive vs negative directive framing is one of the few prompt-engineering moves with consistent, replicated evidence behind it. The instruction-tuning literature finds again and again that models follow 'do this' instructions more reliably than 'don't do that' instructions, across model families and task domains. The effect is large enough that a positive rewrite of a negation-heavy system prompt is often the single highest-leverage prompt change a team can make.

The goal of this deep dive is to give a senior-engineering grasp of why the asymmetry exists, where the failure modes show up, and how to reframe prohibitions as substitute behaviors in production. The interview-worthy framing is to name the mechanism (suppression is harder than execution), the failure mode (stacked negations and priming), and the fix (policy as action map).

The sections below cover the mechanism, the priming risk of repetition, the production reframe pattern, and the failure modes a senior engineer should call out from a prompt review.

Why suppression is harder than execution

A positive directive of the form 'when X, do Y' establishes a clean trigger to action path in the model's attention pattern. The model can attend to X in the input, attend to Y in the directive, and emit Y. There is no suppression step; the generation is direct.

A negative directive of the form 'do not Y' is structurally asymmetric. The model has to first attend to Y in the directive, infer what Y would look like in the current context, then ensure the generation path through Y is suppressed. Suppression is fragile because it competes with smoother completion paths through Y that the base model's prior makes locally probable.

The practical consequence is that negative directives have higher variance in compliance. Sometimes the suppression fires cleanly. Sometimes the prohibited behavior leaks through, often partially: the model produces a softened version of the prohibited action while claiming it is not. Positive substitutes do not have this failure mode because there is no prohibited action to model in the first place.

Stacked negations and the priming backfire
The reframe pattern: policy as action map
When negative directives are still justified
Failure modes a senior engineer should call out
Sign in to unlock the full deep dive.

Situations where this technique stops working.

Sign in to see when this approach fails.

2–4 min · Everything important, quickly.

Sign in to see the quick scan of the deep dive.

Real products, models, and research that use this idea.

  • Anthropic's Constitutional AI and Claude Opus 4.7 prompting guide both recommend positive framing over prohibitions for behavioral steering.
  • OpenAI's GPT-5.5 prompt engineering docs advise replacing 'do not' instructions with 'instead' patterns to improve compliance.
Sign in to see more production examples.

What an interviewer would ask next. Try answering before peeking at the approach.

QHow would you reframe a stack of six negative directives without losing coverage?
A

Pair each prohibition with a recognition condition and a substitute action; merge overlapping conditions; produce a single positive policy map that covers the same surface.

2 more follow-ups an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Strengthening a negative directive with capitalization or repetition instead of replacing it with a positive substitute action.

Sign in to see all red flags and common mistakes.

60 second bullets to scan on the way to the call.

  • Why positive directives outperform negative ones

  • The trigger plus substitute action pattern

Sign in to unlock the revision sheet.

Primary sources. Browse if you want the original framing.

Similar questions

Same topic, related formats. Practice these next.

4 curated
Next question
Flashcard: what is a stop sequence in an LLM API call and what is it used for?
Flashcard·Easy