Most advice about prompting AI focuses on templates: give the AI a role, provide context, specify a format, add examples, and so on.
But what actually makes a prompt work better?
Recent research offers a useful way to think about it.
A 2026 ACL paper, Beyond Prompt Engineering: A Systematic Analysis of Prompt Lexical Sensitivity and Its Impacts on Quality, analyzed more than 132,000 prompt variations to study how changes in language affect LLM outputs.
Among the patterns identified by the researchers, two are particularly useful for everyday AI users:
1. Domain-Specific Terminology — which the researchers describe as helping to "anchor semantic boundaries."
2. Explicit Action Directives — which help "formalize reasoning trajectories."
In simpler terms:
Ground it. Direct it.
1. Ground the AI in the right domain
Consider this prompt:
What is the definition of Yoga? Answer in 2 sentences.
This is a perfectly reasonable question.
But "Yoga" can refer to many different things: physical postures, breathing practices, meditation, wellness, spirituality, Indian philosophy, and more.
The model therefore has a large space of potentially relevant information from which to construct its response.
Now consider:
What is the definition of Yoga according to Patañjali's Yoga Sutras? Explain it using the central concepts of citta, vṛtti, nirodha, and draṣṭā and connect how they define the purpose of Yoga. Answer in 2 sentences.
We've now grounded the question within a much more specific domain.
Terms such as Patañjali, citta, vṛtti, nirodha, and draṣṭā establish the conceptual territory in which we want the model to operate.
This is what domain-specific terminology helps accomplish.
Why does grounding work?
At their core, Large Language Models generate responses by repeatedly predicting the next token based on the context that precedes it.
A broad word such as "Yoga" is compatible with many possible continuations.
But adding terminology associated with a particular philosophical framework changes that context. Continuations consistent with that domain become more relevant to the request.
The researchers describe this effect as anchoring semantic boundaries.
The lesson isn't that you should fill every prompt with jargon.
It's that when you care about a particular domain, giving the model the language of that domain can help communicate what kind of knowledge and conceptual framework you want it to use.
2. Tell the AI what to do
Grounding establishes what the AI should be thinking about.
But there's another source of ambiguity: what should it actually do with that information?
Consider instructions such as:
Analyze this.
Tell me what you think.
Explain this.
These aren't necessarily bad prompts. Modern models can often produce useful answers from them.
But they leave the model considerable freedom to determine what kind of response you wanted.
Compare that with directives such as:
Identify the three major risks.
Compare the two approaches and explain the trade-offs.
Rank the alternatives and justify the ranking.
Identify the assumption, evaluate the evidence, and recommend the next step.
The researchers refer to this second pattern as Explicit Action Directives.
Instead of leaving the structure of the task implicit, we specify the actions we want performed.
A simple mental model
You don't need to memorize complicated prompting frameworks.
Before sending an important AI request, ask yourself two questions:
Have I grounded it?
Does the model know which domain, concepts, terminology, or context matters?
Have I directed it?
Does the model know exactly what I want it to do with that information?
That's it.
GROUND IT → DIRECT IT
Better prompting isn't necessarily about writing longer prompts.
It's about reducing unnecessary ambiguity between what you mean and what the model thinks you mean.
Research reference: "Beyond Prompt Engineering: A Systematic Analysis of Prompt Lexical Sensitivity and Its Impacts on Quality," Findings of ACL 2026.
This piece also runs as a post on LinkedIn.
Want this running in your business instead of on your reading list? We build it with you — first call is free.