Click any words you think contain an error. Click again to unmark.
Every directive is a vague adjective with no observable target. Replace 'helpful', 'thorough', 'accurate', 'good' with concrete actions or measurable properties the eval suite can check.
Imagine giving a new chef the instruction 'cook good food'. They cannot tell what good means, so they guess based on their training. Now imagine telling them 'use fresh ingredients, cook to medium-rare, and serve in under fifteen minutes'. They know exactly what to do and you can check whether they did it. The prompt here is the first version, three times. Every adjective sounds caring but tells the model nothing it can act on.
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.
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.
This eighty-character prompt is a compressed catalog of the most common production prompt-engineering anti-patterns. Three sentences, three unfalsifiable directives, zero behavior the model can act on, zero metrics the eval can check. Every adjective sounds like care and contains no information.
The question matters because prompts like this end up in production systems through inheritance. Someone writes a quick starter prompt for a prototype, the prototype graduates to a production service, and the vague directives travel with it. Six months later the team is debugging quality regressions on a prompt that never had any operational content to begin with.
The deep dive walks through each directive, names the specific anti-pattern, shows the operational rewrite, and connects each rewrite to an eval assertion you would write alongside it. The point is to make the refactor mechanical so it becomes a habit rather than a special-case effort.
Anti-pattern one, vague role
You are a helpful assistant looks like a role definition. It is not. It is a register cue with no scope, no audience, no domain, no behavioral envelope. The model interprets it as a tone hint (be polite, be cooperative) and otherwise falls back to its training defaults.
The operational fix names three things at once. The role itself (Python tutor, customer-support agent, legal research assistant). The audience (beginners learning their first language, enterprise users on a paid plan, contract attorneys). And the implicit goal (explain concepts, resolve tickets, surface case law citations). Each of those constrains the model's prior over what kind of output is appropriate.
A worked rewrite. You are a Python tutor explaining concepts to beginners learning their first programming language. Use concrete code examples. Avoid jargon without definition. This sentence gives the model a behavioral envelope it can latch onto, and gives reviewers something to check the output against.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- OpenAI's prompt engineering guide for GPT-5.5 leads with 'be specific' precisely because vague adjectives like 'helpful' and 'accurate' degrade compliance reliably.
- Anthropic's Claude Opus 4.7 docs explicitly recommend naming concrete behaviors over asking for quality in the abstract.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you write the eval suite for the rewritten prompt?
Map each directive to an assertion: citation-rate metric for the citation directive, word-count assertion for the length cap, structural-regex check for the definition then example then pitfall format.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Reading this prompt and thinking it is just thin; the real failure is that every directive is unfalsifiable, so neither the model nor your eval suite can act on it.
60 second bullets to scan on the way to the call.
Why 'helpful assistant' is a register cue, not a role
The two operational shapes (action, measurable property)
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.