The anatomy of a good prompt
A good prompt has four elements:
**Context** — Who are you, what are you working on, why does this matter? 'I'm a marketing manager at a B2B software company. I'm writing an email to enterprise prospects who have gone quiet after an initial demo.'
**Task** — What specifically do you want? 'Draft a three-paragraph follow-up email.'
**Format** — What should the output look like? 'Tone: professional but warm. No bullet points. Under 150 words.'
**Examples or constraints** — What should it avoid? What's the target? 'Don't mention pricing. Don't use the phrase 'just checking in'.'
Most people only provide the task. The other three elements make the difference between output you have to rewrite from scratch and output you can edit and send.
Using Claude well
Claude (by Anthropic) is particularly strong at tasks involving long documents and careful reasoning. A few things that work well:
Long document analysis: Paste an entire contract, report, or transcript and ask specific questions. 'What are the three main risks in this contract? What are the payment terms? Are there any unusual clauses I should flag for legal review?'
Iterative writing: Don't ask for a finished piece in one shot. Ask for an outline. Review it. Then ask for a section at a time. You'll get much better results and maintain more control.
Explicit reasoning: For complex tasks, ask Claude to 'think through this step by step before giving your answer'. It sounds simple but consistently improves output quality.
Feedback loops: If the output isn't quite right, say exactly what's wrong. 'This is too formal — make it sound like a colleague, not a report.' 'The second paragraph misses the point — the key issue is X.'
What Claude is *not* good at: real-time information (it has a training cutoff), precise arithmetic, and anything that requires genuinely novel reasoning rather than sophisticated pattern-matching.
Using Microsoft Copilot well
Copilot's advantage over other AI tools is its access to your Microsoft 365 data. That's also what makes it worth being thoughtful about.
In Outlook — what works: 'Summarise this thread and list the action items for me.' 'Draft a response agreeing to the meeting but noting I'll need to leave by 3pm.' 'What emails from last week mentioned Project Phoenix?'
In Teams — what works: 'Recap what was decided in the meeting I missed.' 'What action items are assigned to me from today's calls?'
In Word — what works: Drafting from an outline. Rewriting in a different tone. Summarising long documents into executive summaries. Do *not* expect factual accuracy on things outside your documents — it will hallucinate, especially on specific numbers.
In Excel — what works: 'Create a chart showing revenue by quarter.' 'Find the rows where the value in column C is more than 2x the average.' 'What trends do you see in this data?'
Key limitation: Copilot's quality varies significantly depending on how well-organised and labelled your Microsoft 365 data is. If your SharePoint is a mess, Copilot's answers about your company data will reflect that.
Using Cursor (for technical teams)
If you manage developers or work in a technical environment, understanding how Cursor changes work patterns is increasingly important.
Cursor changes the economics of technical work in a few specific ways:
**Routine implementation is faster.** Creating a standard database query, writing a test, building a form — tasks that previously took 30-90 minutes can take 5-10 minutes. This is already the norm for teams using Cursor.
**The bottleneck shifts to design and review.** If AI can write code quickly, the scarce resource becomes knowing *what* to build and reviewing *whether* the AI-generated code is correct and secure. Technical leadership matters more, not less.
**New risks emerge.** AI-generated code can be subtly wrong in ways that aren't obvious. Teams need review practices specifically designed for AI-assisted code — not just the same practices applied faster.
What this means for managers: Your developers using Cursor aren't cheating. They're using a genuinely productivity-enhancing tool. The question to ask isn't 'should they use it' but 'what does good code review look like when AI wrote the first draft?'
Building your personal AI workflow
The people who get the most value from AI tools are the ones who've developed a consistent workflow — not those who try every new tool.
**Start with one tool and one task.** Pick one repetitive task in your week that involves a lot of writing or analysis. Use AI for that task for 30 days. Get good at it. Then expand.
**Keep a prompt library.** When you write a prompt that works well, save it. Over time you'll build a personal library of prompts for your most common tasks.
**Review outputs before you use them.** Always. Not because AI is usually wrong — it's often right — but because the times it's confidently wrong will trip you up if you've stopped checking.
**Notice what it's bad at for your work.** Every domain has areas where AI underperforms. In legal: specific jurisdictional rules. In finance: current market data. In HR: nuanced interpersonal situations. Know your area's AI failure modes.
**The goal is augmentation, not replacement.** The most effective people use AI to handle the routine parts of complex tasks — the drafting, the summarising, the structuring — so they can focus their judgment on the parts that require it.
The biggest predictor of good AI output is not which model you use — it's how specifically you describe what you want. Most people's AI outputs are poor because their prompts are too vague, not because the AI isn't good enough.