Why do six short, specific instructions usually beat one long paragraph of guidance?

Flashcard·Easy·4.0 · 0·~30s·Asked atAnthropic OpenAI Cohere

Attempt it

A bulleted list is dense: each line is one rule the model can follow or violate, and verification is local. A paragraph buries rules inside connective tissue, makes priorities ambiguous, and lets the model pattern-match to the tone instead of the rules. The token cost of a bulleted version is usually smaller too, the connective words are gone. The bigger reason is testability: with bullets you can write evals that probe each rule independently. With a paragraph, you can only test the gestalt, which makes regressions hard to localise.

Your answer

A bulleted list works better than a paragraph for system-prompt rules for three concrete reasons. **Bullets are locally verifiable.** Each line is one rule. The model can satisfy or violate it independently of the other lines. With a paragraph, the rules are tangled in connective sentences and the model often pattern-matches to the overall tone instead of extracting individual rules. **Bullets are token-cheaper.** The connective tissue,"and", "meanwhile", "in addition", "furthermore", is gone. Six short bullets typically use fewer tokens than a paragraph saying the same thing. **Bullets are testable.** You can write an eval for each rule: does the output respect this one constraint? With a paragraph, you can only test the gestalt: does the output feel right? Gestalt tests catch the big regressions but miss the subtle ones, and they cannot localize which rule was violated when something fails. The one case where a paragraph genuinely helps is when the guidance is a single coherent narrative (a persona description, an explanation of context) rather than a list of rules. There the connective tissue is the content, not overhead.

**Six short, specific instructions beat one long paragraph because they are denser, more testable, and easier for the model to extract as discrete constraints.** A system prompt is read by a model that processes tokens through attention rather than by a human reader who scans for structure. Bullets exploit two facts about how this reading happens. First, each bullet is a local rule. The model can recognise it, weight it, and check the output against it independently of the other bullets. A paragraph buries the same rules inside narrative flow; the model is more likely to pattern-match to the tone of the paragraph than to extract the individual constraints. Production teams routinely see this empirically: converting a paragraph to bullets without changing the content improves rule adherence on the eval set. Second, bullets are testable. With six bullets you can write six independent evals (one per rule) and localize failures. With a paragraph you can only test whether the output "feels right" and the failure modes are diffuse. Diffuse failures are the worst kind because you cannot regression-test them. There is also a token-cost argument. The connective tissue in a paragraph ("and", "meanwhile", "in addition", "that said") costs tokens and contributes nothing to the rule set. Six bullets typically run shorter than the paragraph that says the same things. The one exception is when the system prompt is a narrative rather than a rule set, a persona description, a context explanation, a tone-setting passage. There the connective tissue is the content and bullets would be a worse fit.

Quick reference

TL;DR

Bulleted instructions are dense, testable, and cheaper in tokens; paragraphs hide rules inside connective tissue and reduce evals to gestalt checks.

Memory aid

Easy to grasp

Picture a kitchen with two sets of recipe instructions on the wall. One is a paragraph in flowing prose with the steps embedded inside connective sentences. The other is six short bullet points. New cooks doing the same dish under time pressure mess up the paragraph version more often, not because they cannot read English but because each step is hard to find inside the prose. The bullet version has one rule per line. You can point at the one that got missed. You can train someone by walking through each line. You can put a checkmark next to each one. The model running your system prompt is in the same situation as the cook under time pressure.

Key concepts

Context engineering Prompt engineering LLM evaluation

Detailed answer & concept explanation

~5 min read

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

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

Quick scan · 2–4 min

Everything important, quickly.

3 minutes: why bullets are denser, why they are testable, the token-cost argument, when prose is right, the recommended 2026 layout.

Real-world examples

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

Anthropic's published system-prompt patterns for Claude Opus 4.7 use a short persona paragraph followed by a numbered rule list followed by output-format hints.
OpenAI's 2026 prompt-engineering guide for GPT-5.5 explicitly recommends bullet-list rule sections over flowing prose for instructions.
LangSmith, LangFuse, and Phoenix all expose per-rule pass-fail evaluators that work cleanly on bullet prompts and poorly on paragraph prompts.
DSPy's instruction-optimization pipeline routinely converts paragraph instructions into bulleted constraint lists during its compile step.
Cursor and Cody both ship default system prompts in 2026 with explicit numbered rules rather than narrative guidance, after iteration showed it works better.

Interviewer follow-ups

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

QWhat does a good per-rule eval look like in practice?

QHow does instruction ordering inside the bullet list matter?

QWhen does an explicit deny-list of forbidden outputs beat a bullet rule?

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Writing system prompts as flowing English because it reads well to humans. The model is not the audience that benefits from connective tissue; bullets are denser, more testable, and usually cheaper per call.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

References

Primary sources. Skim if you want the original framing.