The quality of what Claude gives you is almost entirely determined by the quality of what you ask. This isn't theory. These are the techniques that actually make a difference.
Most people write prompts the way they'd text a mate: brief, context-free, and hopeful. Claude doesn't read between the lines. It responds to exactly what you give it. The techniques below fix that, in order of impact.
Start here
The core formula: use this every time
RoleWho it is
+
TaskWhat to do
+
ContextWhy + who for
+
FormatHow to output it
+
ConstraintsWhat to avoid
Example: "You are an experienced tradesperson's bookkeeper who understands cash flow pressures in small service businesses in New Zealand. Write a 3-paragraph explainer on why automating invoice follow-ups saves time and protects revenue. Use short sentences, no jargon, and a warm but practical tone. Do not use the word 'streamline'."
The techniques that actually make a difference
Starting with "You are a..." shifts how Claude frames its entire response: vocabulary, depth, assumptions, tone. A prompt with no role gets a generic assistant voice. A prompt with a specific role gets a subject-matter expert.
Not thisSummarise this client brief for me.
ThisYou are an experienced operations consultant. Summarise this client brief, flagging any scope risks or unclear deliverables.
Showing Claude an example before asking your real question is the single most reliable way to control format, tone, and style. Paste in a piece of writing you liked. Claude calibrates from examples much faster than it calibrates from adjectives like "professional" or "friendly."
How to use it"Here's an example of the tone I want: [paste sample]. Now write one in the same style about [topic]."
Be precise: bullet list, numbered steps, table, email, plain paragraph, three headings, 150 words max. Match the format to how you'll actually use the output. Don't leave Claude to decide.
Not thisGive me some ideas for this month's newsletter.
ThisGive me 5 newsletter topic ideas as a numbered list. Each should have: a working headline (max 8 words) and one sentence explaining the angle.
For anything analytical or multi-step, tell Claude to reason through the problem before giving you the final output. This surfaces the logic so you can spot errors, and produces more accurate results. Especially useful when you're working through something you're not sure about yourself.
Add this phrase"Think through this step by step, then give me your recommendation." or "Before answering, explain your reasoning."
AI models will fill gaps with plausible-sounding content, which is how you get confident nonsense. Explicitly instructing Claude to flag uncertainty dramatically reduces this. This is especially important for anything factual, research-based, or NZ-specific.
Add to any factual prompt"If you're uncertain about any specific facts or figures, say so rather than guessing."
Claude has defaults: bullet points, formal tone, lengthy preambles. These aren't bad, but they're not always what you want. Ruling things out closes off a whole set of unhelpful patterns before they appear. Especially useful for tone, word choice, and format.
Examples"Do not use bullet points." / "Don't use the word 'leverage'." / "Don't include a preamble. Start with the content."
For complex tasks like proposals, strategies, and plans, ask Claude to identify what information it needs before it starts writing. This catches gaps before you get a mediocre first draft. One of the more underused techniques, and one of the most effective for anything high-stakes.
How to use it"Before you draft this proposal, ask me the 3–5 questions you need answered to do it well."
Claude specifically handles structured prompts well. When your prompt has distinct parts (background, task, examples, constraints) wrapping them in XML tags keeps things clear. This matters most for long or complex prompts where Claude might blend sections together.
Worked example<context>I run a small bookkeeping firm in Auckland. My clients are trade businesses — builders, electricians, plumbers. Most are behind on their invoicing.</context>
<task>Write a short email I can send to a client who has 3 unpaid invoices. The tone should be firm but not aggressive. I want to preserve the relationship.</task>
<format>Plain text email. Subject line included. Max 120 words.</format>
<constraints>Do not use the word "unfortunately". Do not threaten late fees unless I ask. Do not use a formal sign-off like "Yours sincerely".</constraints>
Without the tags, a long prompt like this often produces something generic. With them, Claude knows exactly what each section means.
If Claude's first response is fundamentally wrong or misses what you actually needed, don't keep going in the same conversation trying to fix it. Go back and edit your original prompt instead. In Claude.ai you can hover over your message and click the edit icon to rewrite it and resend.
This matters for two reasons. First, it's faster: you're correcting the source of the problem rather than layering instructions on top of a bad start. Second, it produces a cleaner result: a fresh run from an improved prompt beats a long back-and-forth trying to steer Claude back on course.
Less effectiveYou send a vague prompt, get a generic response, then send three follow-up messages trying to narrow it down. Each message pulls the conversation further from what you actually needed.
BetterYou read the first response, realise the prompt was missing context, go back and edit the original prompt with that context added, and resend. One better response instead of four mediocre ones.
Prompts that produce reliable output are repeatable systems. Save them, label them by use case, and update them when you find something better. You'll likely use the same 15–20 prompts 80% of the time.
Practical move: Start a simple "prompt library" (even a Google Doc) with three columns: what it's for, the prompt itself, and notes on what you tweak each time. In Airtable this is a five-minute setup.
What consistently produces bad results
No context about the audience
"Write a social post about my new service." For who? What platform? What do they already know?
Vague outcome words
"Make it engaging" or "make it professional" mean nothing. Describe exactly what engaging looks like for your use case.
Stacking multiple tasks in one sentence
One prompt, one job. If you need three things done, write three prompts or break it into clear numbered steps.
Assuming it remembers
Each conversation starts fresh unless you're in a Claude Project. Don't reference "what we discussed last time". Restate the context every time.
Accepting the first output
The first response is a starting point, not the final answer. Refine it: "make this shorter", "change the tone", "remove the second point and expand the first."
Over-prompting simple tasks
For a quick question or a simple rewrite, a long structured prompt makes things worse. Match prompt complexity to task complexity.
When to use which technique
Match the technique to the task
Writing content (posts, emails, proposals)
Role + example + tone constraints. Paste in a sample of the style you want.
Analysis or decision-making
Chain-of-thought + "if uncertain, say so". Ask it to reason before concluding.
Complex brief (strategy, plan, proposal)
Ask it to ask you questions first. Don't let it fill gaps with guesses.
Repeatable business task
Full formula + save it. Role, task, context, format, constraints. Then store it.
Getting a specific format (table, list, JSON)
Specify format explicitly. Don't leave it to Claude to decide.
Anything factual or research-based
Add "if uncertain, say so" and verify key facts before using them.
Long or complex prompt
Use XML tags to separate context, task, format, and constraints clearly.
The honest bit
Prompting is a skill, not a trick
These techniques genuinely improve results. But there are a few things worth saying plainly about how Claude actually works, so you're not surprised when something doesn't land the way you expected.
Ambiguous prompts get confident guesses. If your prompt could mean two things, Claude picks one and runs with it without flagging the ambiguity. Re-read your prompt before sending and ask: could this mean something else?
Better prompts don't fix bad information. If the content you're working with is thin or wrong, a well-structured prompt produces a well-structured wrong answer. Claude works with what you give it.
NZ-specific knowledge is thinner than you'd expect. Local regulations, market context, cultural nuance: Claude has less reliable training data here. For anything NZ-specific, verify before you rely on it.
There's no magic prompt. No formula produces perfect output every time. The goal is reducing the gap between what you asked and what you needed, and that gets easier with practice, not shortcuts.
One habit worth building
After you get a response, especially on anything strategic, financial, or client-facing, don't just edit it. Test it first:
"Tell me why this wouldn't work, and what would need to change so that it could."
That question surfaces the gaps in the reasoning before they become your problem. I use it more than any other prompt I have.
The bottom line
A lot of people try a generic AI tool, get underwhelming results, and write the whole thing off. In my experience, the gap is almost always in the prompt, not the model. The formula, the role, the example, the constraints: these aren't advanced techniques. They're just clear communication.
Start with the formula. Pick one technique from the list that fits something you do regularly. Try it this week. That's enough to notice a difference.
Prompting well is a low-stakes skill to practise. You can't break anything, and the cost of a bad prompt is a few seconds and one more try. The upside is responses you can actually use without spending twenty minutes editing them.
If you want to go further and use Claude inside a proper system built around your business, that's exactly what I help with. Get in touch.