Being clear and direct

The baseline prompt scored 2.32 out of 10. The first technique for getting that number up is also the simplest: rewrite the opening line so it's clear and direct. This single change will take the score from 2.32 to 3.92 — a 1.6-point jump from one line of text.

The opening line carries the request

The first line of a prompt is its single most important element. It sets the context for everything that follows. Claude reads the whole prompt, obviously, but the opening sentence disproportionately shapes how the model interprets the rest — what kind of task this is, what tone to use, what the output should look like.

Two principles apply to that opening sentence: clarity and directness.

Clarity: simple language, no ambiguity

Being clear means using language anyone could parse on the first read, with no room for misinterpretation. State exactly what you want.

Before:

I need to know about those things people put on their roofs that use sun — those solar panel things, I think they're called.

After:

Write three paragraphs about how solar panels work.

The before version is how people talk when they're thinking out loud. The after version is how people talk when they know what they want. Both communicate with a human; only one communicates cleanly with a language model. Claude can still handle the vague version, but you're spending context budget on translation that could have gone into the actual task.

Directness: instructions, not questions

Being direct means giving Claude an instruction rather than asking a question, and leading with an action verb.

Before:

I was reading about renewable energy and geothermal energy sounds neat. What countries use it?

After:

Identify three countries that use geothermal energy. Include generation stats for each.

Questions invite open-ended answers. Instructions pin down exactly what to produce. The action verb — Write, Generate, Identify, Create, List, Analyse, Extract — is the single word doing the most work. It tells the model what shape the output should take before any detail fills in.

Apply it to the meal plan prompt

Here's where we were:

What should this person eat?

That line is doing none of the work it could be doing. It's a question, not an instruction. It doesn't say "meal plan." It doesn't say "one day." It doesn't specify that it's for an athlete. Claude has to infer all of that from context.

Rewriting with clarity and directness:

Generate a one-day meal plan for an athlete that meets their dietary restrictions.

One line. Look at everything it communicates:

• Action verb: Generate — the model is now in "produce a structured artifact" mode, not "answer a question" mode.
• Output shape: a one-day meal plan — named, concrete, known vocabulary.
• Audience: for an athlete — calibrates the expected caloric range and composition.
• Constraint: that meets their dietary restrictions — promotes the restriction data from "input field" to "hard requirement."

All of that was implicit before. Now it's explicit, which means Claude doesn't have to guess and you don't have to hope.

The score jumps from 2.32 to 3.92

Run the eval again — same dataset, same grader, same extra_criteria — and the average score goes from 2.32 to 3.92. That's a 1.6-point lift from rewriting four words of the first line. Read the reasoning strings on the new report and you'll see the difference directly: the grader now credits the prompt with more complete meal plans and better restriction compliance, and penalises it less for "too vague."

Is the prompt good yet? No. 3.92 is still on the wrong side of the midpoint. But you've proved the technique moves the number, and that's the whole point — the technique isn't a feeling, it's a measurable improvement on a stable eval.

The take-away habit

Every time you write a new prompt, start from the opening line. Lead with an action verb. Kill questions. Replace conversational hedging with direct specification of the task. Even before you add anything else — guidelines, examples, XML structure — a clear and direct opening is the cheapest win in prompt engineering, and it compounds with everything else you do in the rest of this module.