Why Most Prompts Underperform
Most people write prompts the same way they write a text message: short, casual, and vague. "Summarize this." "Make it better." "Write an email." These work fine for casual conversations. They fail for anything where quality actually matters.
The model is not failing. The communication is failing. The model cannot guess what "better" means to you. It does not know the audience for the email. It does not know how long the summary should be or what format you expect. When you give it a vague prompt, it fills in those gaps with its best guess, which may have nothing in common with what you actually needed.
The CLEAR Framework
CLEAR is a five-part structure for building prompts that get consistent, useful results. Each letter stands for one component.
| Letter | Component | What it does | Example |
|---|---|---|---|
| C | Context | Background the model needs to do the task well | "This is for a client in the healthcare industry who is not technical." |
| L | Length | How long or short the output should be | "Keep it under 150 words." or "Write at least three paragraphs." |
| E | Examples | Show the model what good output looks like | "Here is an example of the tone I want: [paste example]" |
| A | Audience | Who will read or use this output | "The reader is a CFO who reads quickly and values bullet points." |
| R | Role | Who the model should be when responding | "You are a senior copywriter with ten years of B2B experience." |
You do not need to use all five in every prompt. A simple task might only need two or three. But knowing all five means you know exactly what to add when a prompt is not working.
Seeing CLEAR in Action
Here is a weak prompt and its upgraded version.
Weak prompt:
Upgraded with CLEAR:
[C] I am following up with a prospect who attended our product demo last Tuesday but has not responded to my first follow-up email sent three days ago.
[A] The recipient is a VP of Engineering at a mid-size software company. They are busy and skeptical of sales emails.
[L] Write three subject line options, each under 50 characters.
[E] Good examples of subject lines I like: "Quick question about the demo" and "One thing you mentioned Tuesday"
The second prompt will almost always produce better output. Not because the model is smarter, but because you gave it everything it needs to do the job.
The Role Component: More Powerful Than It Looks
Setting a role is one of the highest-leverage moves in prompting. When you say "you are a cardiologist," the model activates patterns from medical literature, clinical communication, and specialist terminology. When you say "you are a patient educator," it activates patterns from plain-language health communication instead.
Roles work best when they are specific. "You are an expert" is weak — an expert in what? "You are a senior tax attorney specializing in corporate restructuring" is specific. It gives the model a clear pattern to match.
A useful trick: set the role first, before any other instruction. Research into prompting has found that model responses shift meaningfully when the role is established before the task is presented, rather than after (Wei et al., arXiv:2201.11903 examines this in the context of structured reasoning, where role framing affects output quality).
The Format Component: Often Overlooked, Always Worth Adding
The model will choose a format if you do not specify one. Sometimes that guess is fine. Often it is not. If you need bullet points, say so. If you need a table, describe the columns. If you need a numbered list with headers, show that structure.
Format specification also helps with consistency across many prompts. If you are using AI to process a batch of documents, specifying a fixed output format means every output has the same structure, which makes downstream processing much easier.
Constraints: Telling the Model What Not to Do
Constraints are a sixth component not in the acronym, but important enough to mention. Telling the model what to avoid is often as valuable as telling it what to do.
Examples of useful constraints:
- "Do not use jargon."
- "Do not include information about competitors."
- "Do not make any claims you cannot support from the text I have provided."
- "Do not use bullet points — write in prose."
Constraints are especially important for high-stakes tasks where a certain type of error is particularly costly. If you are using AI to draft a legal summary, "do not speculate about intent" is a constraint that can prevent serious problems.