Anatomy of a Good Prompt

A good prompt is not a clever trick; it is a clear brief. The same instinct that makes you good at handing work to an analyst makes you good at prompting Claude: say who they are, what you need, how you want it back, and what to watch for. That four-part shape (Role, Task, Format, Scrutiny) is the backbone of almost every strong prompt, and once you have it, the quality of your output stops being a matter of luck.

The four parts

Each part does a distinct job. Skip one and you feel the gap in the answer.

Role

Tell Claude who it is helping and from what perspective. This sets the tone, the assumptions, and the level of detail. “You are assisting a private equity associate” produces a different, sharper answer than no framing at all, because it tells Claude what matters and what can be taken as read.

“You are assisting a private equity associate screening a potential acquisition.”

Task

State plainly what you want done. Be specific about the thing: summarize, compare, draft, sense-check, list. Vague verbs produce vague output. “Have a look at this” is not a task; “summarize this teaser for an initial screen” is.

“Summarize the attached teaser so I can decide whether it is worth a first call.”

Format

Tell Claude how to hand the work back. Headings, bullets, a table, a one-paragraph answer, a maximum length. Format is what makes output usable without reshaping it yourself, and it is the part people most often forget.

“Use these four headings, bullet points under each, and keep it to one page.”

Scrutiny

Ask Claude to be honest about the limits of what it produced. Tell it to flag gaps, note anything the document is vague about, and mark anything unusual. This single habit is what turns Claude from a confident writer into a useful analyst, because it surfaces what you would otherwise have to catch yourself.

“Flag anything the document leaves unclear, and call out anything that looks unusual.”

A full prompt, assembled

Here is the four-part shape put together as one prompt you can paste and adapt. Read it and you can see each part doing its job.

You are assisting a private equity associate. I’ve attached a teaser for a potential acquisition (Project Atlas, a mid-market field-service software business).
Summarize it for an initial screen using these headings:

Business overview: what they do and how they make money
Financial highlights: revenue, EBITDA, growth, margins
Investment thesis: why this could be attractive
Key risks and open questions
Use bullet points under each heading and keep it to one page. Flag anything the teaser is vague about, and call out anything that looks unusual or that I should press management on.

This is what that prompt looks like in the message box, ready to send with the teaser attached:

Building one from scratch

When you face a new task, do not reach for a template you half-remember. Build the prompt in order, one part at a time.

1. Set the role

   Begin with who Claude is helping and the lens to use. One line is enough: “You are assisting a private equity associate.”

2. State the task

   Say exactly what you want done, with a specific verb: summarize, compare, draft, sense-check. If you would struggle to grade the result, the task is still too vague.

3. Specify the format

   Describe the shape of the answer: headings, bullets, a table, a word or page limit. This is what lets you use the output without rebuilding it.

4. Add the scrutiny instruction

   Ask Claude to flag gaps, note anything unclear, and mark anything unusual. Make this a standing habit on every analytical prompt.

5. Attach, send, then verify

   Add the relevant file, send, and check Claude’s flagged points (and any figures) against the source before the output goes near a deliverable.

Patterns that stack on top

The four parts cover most of what you will ask for. The moves below are the next layer: a small set of shapes that handle the situations that keep coming back in deal work. None of them is clever. They are reliable, and knowing them saves you reinventing the same prompt every week. Treat them as shapes to adapt rather than scripts to obey; the wording is a starting point, the structure is the value.

Iterate, do not restart

The most common mistake is treating each prompt as a fresh start. Claude holds the thread of a conversation, so the fastest route to a good answer is usually a decent first draft followed by precise corrections. You would not ask an analyst to redo a whole memo because one section is off; you would tell them what to change. Do the same here.

Good start. Now make three changes: tighten the risks section to the five that actually matter for an initial screen, drop the generic market commentary, and add a one-line “so what” under each risk explaining why it would affect the thesis. Keep everything else as it is.

Ask for the reasoning, not just the answer

For anything analytical, the conclusion is less useful than the path to it. Asking Claude to lay out its reasoning does two things: it gives you something to check, and it tends to produce a more careful answer, because the model has to justify itself rather than just assert.

Before you give your conclusion, walk through your reasoning step by step: what the document says, what you can infer from it, and where the gaps are. Then give your conclusion as a separate final paragraph, and note how confident you are in it and why.

This pairs naturally with the verification routine in Verify Everything. Visible reasoning is far quicker to check than a bare verdict.

Structured comparison

You constantly weigh things against each other: two targets, two structures, the case for and against a thesis. Left to its own devices, Claude may give you flowing prose when what you want is a grid. Ask for the structure explicitly, and give it the dimensions to compare on.

Compare these two acquisition candidates as a table. Use these rows: business model, revenue scale, growth rate, margin profile, customer concentration, and the single biggest risk. Add a final row with your one-line read on which is the more attractive screen and why. If a document is silent on a dimension, write “not stated” rather than guessing.

The “not stated” instruction matters. Without it, a comparison table is exactly the kind of place a confident-looking but invented figure can slip in.

The devil’s advocate

Claude is, by default, agreeable. Left unprompted it tends to support the direction you seem to be leaning. That is the opposite of what you want when you are pressure-testing a thesis. Ask it explicitly to argue against you.

I am inclined to pursue Project Atlas. Take the opposite view. Give me the strongest, most specific case against this investment: the risks I am most likely to be underweighting, the questions that would most worry an investment committee, and the single thing that, if true, would kill the deal. Be direct, not diplomatic.

Constrain the output

When you need output in a specific shape (a fixed word count, a particular set of headings, a tone, a template), say so up front and be concrete. Vague constraints produce vague compliance. “Keep it short” is weaker than “no more than 150 words”. “Make it formal” is weaker than “written for an investment committee paper: measured, specific, no marketing language”.

Draft the executive summary for an investment committee paper on Project Atlas. Hard constraints: no more than 200 words, three short paragraphs (thesis, value-creation angle, the main risk), measured and specific tone, no superlatives or marketing language. If you cannot fit something important in the word count, list it separately below the summary rather than cutting it silently.

Name the shape you cannot see

Sometimes the failure is not the content but the shape it comes back in. This is the Format part above, used deliberately: decide what you want back before you write the prompt, because if you cannot picture the output, Claude cannot either. Name the headings, the row labels, the length, the order. That specificity is what removes the reformatting step afterwards.

Then add the guard that most people forget: tell Claude what to do when a piece of information is missing. Write “not stated”, flag it, or list it separately. Without that instruction, a clean-looking output can quietly hide a hole.

Combining patterns

These stack. A strong real-world prompt might set the role, state the task, ask for visible reasoning, demand a fixed table format, and finish with a devil’s-advocate pass, all in one message. You do not need every pattern every time. Reach for the one that fits the job, layer a second if it helps, and keep the scrutiny instruction running underneath all of them.

What to take from this

Strong prompting is not a special skill reserved for the technically minded; it is good delegation written down. Set the role, state the task, specify the format, and demand scrutiny, and you will get reliable, usable work back the first time far more often. The patterns on top of that are habits more than techniques: refine instead of restart, ask for the reasoning, force structured comparisons, invite the counter-argument, and constrain the output precisely.

Next: the same job, next month

The four parts get you a good answer today. What they do not do is stop you rebuilding the same prompt from memory the next time the same job comes round, a little worse each time, with the one clause that made it work last quarter quietly missing.

Closing that gap is the whole subject of the next page. A prompt you have written, corrected, and run three or four times has earned the right to become a template: the inputs pulled out into named fields, the constraints written down instead of remembered, and enough structure that a colleague who was not there can run it and get your result rather than their own.

From Prompt to Template takes the four-part shape you have just learned and shows what it grows into: a fuller structure with an intake step that interviews you for the facts it needs, plus the templates that structure produced on live engagements. Read it next. Nothing in it replaces what is on this page, it all sits on top.

For everything else, Writing Effective Instructions goes deeper on phrasing and standing instructions, Why Context Is Everything explains why what you supply matters as much as how you word it, and the Prompt Library has short ready-to-paste prompts for common private equity tasks.