Domain 1 — Prompting and Task Execution (14%)

14% of the exam, roughly 8 questions. The smallest of the "doing the work" domains — the exam assumes you can prompt and spends its weight on what you do with the result.

1.1 What makes a prompt work

A weak prompt is almost always a prompt missing context that only lives in your head. The four things a colleague would need before starting the task are the four things to write down:

Beyond that, three techniques carry most of the remaining value:

• Give examples. One or two samples of the output you want — a previous email, a well-formatted summary — do more for consistency than any amount of description. This is the single strongest lever available to a non-technical user.
• Provide the source material. Attach the document, paste the data, upload the transcript. A prompt that supplies the facts outperforms a prompt that asks Claude to recall them.
• Ask for reasoning on analytical work. "Walk through your reasoning before giving the recommendation" produces better analysis and, just as usefully, lets you see where the logic went wrong.

1.2 Breaking a task into steps

Large requests fail in a characteristic way: the output is uniformly shallow because attention was spread across everything at once. "Analyse this 40-page report and write a board summary, an email to staff, and a slide outline" produces three mediocre things.

The fix is to sequence: extract the findings first, confirm they are right, then write each deliverable from the confirmed findings. This also gives you a checkpoint — if the extraction is wrong, you catch it before it has been propagated into three documents.

Exam pattern — when a stem describes a big multi-part request producing disappointing output, the answer is usually to decompose it into ordered steps with a check between them, not to write a longer prompt or switch models.

1.3 Iterating rather than regenerating

When an answer misses, the instinct is to try again. That is the weakest available move: the same prompt with the same gaps will usually produce the same class of answer.

Instead, name what was wrong and give Claude the missing piece. "Too formal — rewrite it for an internal Slack audience." "You assumed we sell to consumers; we sell to hospitals." "Keep the structure, cut it to half the length." Each correction is information the first prompt did not contain, and each one is worth saving if the task will recur.

1.4 Matching the approach to the task type

← Previous: The mental model  |  Next: Domain 2 — Output Evaluation and Validation →