Refining Claude Outputs for Your Audience

1.Refinement is a loop, not a one-shot fix

Treat a first draft from Claude as a starting point. Anthropic's AI Fluency courses describe the working pattern as a loop: you describe what you want, the model responds, you discern what's useful, and you describe again, sharper this time. Two competencies drive it. Description is how you tell Claude what you need: who the audience is, what the piece is for, and any constraints. Discernment is how you judge what comes back against those needs. Editing for an audience happens where the two meet. You find where the draft misses its reader, then describe the fix more precisely than you did the first time.

The courses say plainly that the first pass is rarely the last one. In the K-12 example, the useful move is to tell Claude up front that the students read below grade level and need key vocabulary defined in context. A missed draft often shows you that this kind of detail was absent from the first request. In the same way, a grant draft that wanders may reveal that you never said what the funder cares about. Each round teaches you something about your reader that the next description can include.

Sources1

2.Write revision instructions Claude can act on

The most common way refinement goes wrong is a vague revision request. The nonprofit writing lesson spells out the fix: be specific in your revisions, and instead of "make this more compelling," say "add these specific details about our community relationships". The K-12 lesson draws the same line. It asks teachers to write one concrete revision prompt, not "make it better," and gives the example of shortening the steps in a set of directions and adding a picture cue for each one. The builders course is shorter still: “Make it look good” is a wish, not a spec.

Specificity matters because a vague adjective such as compelling, better or friendlier leaves Claude to guess which property of the draft is wrong. A specific instruction names the change: which details to add, which steps to shorten, which cue to include. The builders course separates the reaction you might give a colleague from the instruction Claude can carry out, saying that a good critique and an actionable AI description are different artifacts, and you have to learn to translate between them.

Specific revisions often need facts that only you have. The writing lesson tells you to inject the details AI can't know: your track record, partnerships, staff expertise, and actual community relationships. Making a draft fit a particular audience often means supplying exactly this material, because no amount of rephrasing will produce it.

Sources23

3.When Claude follows the words but misses the point

Even a concrete instruction can fail in a typical way. The AI Capabilities and Limitations course calls this letter-over-spirit (the instruction was honored but the intent wasn't). It happens because Claude follows instructions by continuing a pattern. A short instruction like "make this shorter" can be carried out exactly while the real problem stays untouched. The course gives two cases: "Make this shorter" on a draft whose real problem is its structure, and "Make this more professional" on an email where the real problem is that it's burying the ask.

The fix is to change what you say, not how loudly you say it: when an instruction is followed literally but uselessly, restate the goal. Repeating the instruction with more force won't close the gap. The course separates goals from formats: "Convince my team this timeline is realistic" is a goal. "Three bullet points" is a format. When you adapt a draft for an audience, the goal is usually about the reader: what they must notice, believe or do once they've read it.

Sources4

4.Steer register and depth with role, context, and examples

Anthropic's prompting best practices describe three ways to shape tone and depth before a draft is even written. The first is a role: setting a role in the system prompt focuses Claude's behavior and tone for your use case. The second is context and motivation. Providing context or motivation behind your instructions, such as explaining to Claude why such behavior is important, can help Claude better understand your goals. The third is examples, which are one of the most reliable ways to steer Claude's output format, tone, and structure.

The AI Fluency courses make the same point for everyday writing. One says to upload past materials, because your unit plans, rubrics, or strong examples help AI match your voice and rigor. Another says to upload past successful work to help AI understand your organization's voice and approach. If you need a register Claude didn't produce, show it a sample of that register.

client = anthropic.Anthropic()

message = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=1024,
     ? ="You are a helpful coding assistant specializing in Python.",
    messages=[
        {"role": "user", "content": "How do I sort a list of dictionaries by key?"}
    ],
)

print(message.content)

Sources51

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.Asking Claude to "make it better" or "more compelling" gives it enough direction to fix a draft for its audience.Why is that wrong?

   Vague adjectives leave Claude guessing. The guidance is to name the specific change, such as the details to add or the steps to shorten.

   Covered in Write revision instructions Claude can act on

2. 2.If Claude follows a tone or length instruction but the draft still misses the point, repeat the instruction more firmly.Why is that wrong?

   A literal-but-useless result means the intent never got across. Restate the goal next to the instruction. More force won't help.

   Covered in When Claude follows the words but misses the point