A Practical Guide for Your Team

How to Write Prompts
That Actually Work

Most people talk to AI like it's a search engine. Type a few words, hope for magic, get mush, blame the tool.

Here is the shift: a prompt is not a magic spell. It's a brief. You're handing work to the smartest, fastest teammate you've ever had — one who will do exactly what you ask, and nothing you forgot to mention.

This playbook is the exact way I write prompts when I build AI systems for companies. No jargon. No theory you can't use tomorrow morning. Just the patterns that separate a prompt that ships from a prompt that wastes your afternoon.

By Yuval Keshtcher AI Consultant & Implementer Read time: ~12 min
01 — MINDSET

The One Mental Shift

Before any technique, internalize this. It's worth more than every tip below combined.

Claude is not a search engine. It's a teammate you brief. The better the brief, the better the result.

A search engine takes keywords and hands back links. A teammate takes context and produces work. When you type three vague words, you're forcing the AI to guess everything you didn't say — who you are, what you're doing, what "good" looks like. It will guess. Sometimes wrong.

The whole game is simple: reduce the guessing. Every sentence of context you add is one less thing the AI has to invent. You already know how to do this — it's the same skill as briefing a new hire on their first day. You'd never say "write the thing" and walk away. You'd tell them who it's for, what it needs to do, and what done looks like.

And here's the freeing part: you don't need to be technical. The most valuable prompting skills — clarity, context, knowing what "good" looks like — are writing and thinking skills, not coding skills. If you can brief a colleague, you can prompt.

02 — THE FORMULA

The 3-Sentence Formula

When you're moving fast, you don't need a template. You need a reflex. This is it — the pattern that works 80% of the time:

The Formula

1. Describe who you are.
2. Describe what you want.
3. Describe how you want it formatted.

Three sentences, max. That's the difference between a reply you throw away and one you actually use. Watch what it does to a real request:

Weakwrite me a product update email
Strong# who you are I'm a product manager at a B2B software company writing to non-technical customers. # what you want Write a short email announcing that our reports now export to Excel. Warm, plain language, no hype. # how you want it Under 120 words. One clear subject line. End with a single sentence telling them where to click.

Same task. The second one gets used as-is. The first one gets rewritten by you, which means the AI saved you nothing.

03 — STRUCTURE

Anatomy of a Great Prompt

When the task is bigger, the 3-sentence reflex expands into five building blocks. You won't always use all five — but knowing they exist means you'll never wonder "why is this output so off?" It's almost always a missing block.

Role
"You are a senior financial analyst reviewing a quarterly report."Assigning a role is one of the strongest levers you have. It sets tone, depth, and vocabulary in a single line.
Context
"This is for our board. They're smart but not in the weeds on our metrics."Who's it for? What's the situation? What background does the AI need that lives only in your head?
Task
"Summarize the three most important trends and flag one risk."The actual ask. One clear verb. If you have several asks, number them.
Format
"Bullet points. Each under two lines. Lead with the number."How should the answer look? Length, structure, tone. This is where most "close but not quite" outputs get fixed.
Constraints
"Only use figures from the attached data. If something's unclear, say so — don't guess."The guardrails. What to avoid, what not to invent, the hard rules.

Read them top to bottom and you have a complete brief. Notice the order isn't random — role and context go first, because they frame everything that follows. Put the most important instructions at the very top, and re-state any critical rule at the very end. The AI pays the most attention to the beginning and the end of what you give it.

04 — THE CRAFT

The 7 Principles

These are the moves I come back to on every single build. Each one is a small habit with an outsized payoff.

PRINCIPLE 1

Give it a role

Telling the AI who to be is the cheapest way to level up an answer. "You are an experienced onboarding specialist" instantly sets the tone, the depth, and the vocabulary. One sentence does the work of a paragraph of instructions.

Instead ofexplain our refund policy
TryYou are a patient customer-support lead. Explain our refund policy to a frustrated first-time customer.
PRINCIPLE 2

Be specific — vague in, vague out

The AI can't read your mind, so it fills gaps with averages. "Make it better" means nothing. "Make it shorter, warmer, and cut the jargon" means everything. Specificity isn't extra work — it's the work.

Vaguewrite some social posts about our launch
SpecificWrite 3 LinkedIn posts about our launch. Each under 60 words, one concrete benefit per post, no emojis, ending in a question.
PRINCIPLE 3

Show an example, don't just describe

One good example beats three paragraphs of instructions. If you want a specific format or tone, paste a sample of it. The AI is brilliant at pattern-matching — give it a pattern to match.

Show, don't tellCategorize each support ticket. Match this exact format: Ticket: "App keeps crashing on login" → Category: Bug | Priority: High | Team: Engineering Now do the same for the 20 tickets below.
PRINCIPLE 4

Ask it to think step by step

For anything with reasoning — analysis, decisions, math, comparisons — tell it to work through the steps out loud before giving the answer. Accuracy jumps. "Think step by step" is a genuine cheat code for complex tasks.

TryAnalyze this deal step by step: 1) list the stakeholders, 2) their goals, 3) where they conflict, 4) then recommend a path. Show your reasoning for each.
PRINCIPLE 5

Speak plain human

Write the prompt the way you'd actually talk to a sharp colleague. You do not need robotic keywords, ALL CAPS commands, or fake "act as an expert" incantations stacked ten deep. Natural, clear language works best. If it reads like a normal request, you're doing it right.

Over-engineeredACT AS a world-class 10x elite copywriting GENIUS and CREATE viral-optimized MAXIMUM-impact copy...
HumanYou're a good copywriter. Write a homepage headline for a tool that turns meetings into to-do lists. Make it clear before clever.
PRINCIPLE 6

Give permission to say "I don't know"

The biggest fear people have — "what if it makes something up?" — has a one-line fix. Explicitly tell it that being unsure is allowed. This alone kills a huge share of confident-but-wrong answers.

Add this lineIf the data doesn't support a clear answer, say so instead of guessing. Never invent figures or sources.
PRINCIPLE 7

One job per prompt

When you cram five unrelated tasks into one message, quality drops on all five. The AI, like a person, does its best work when it's focused. If you catch yourself writing "and also… and while you're at it…", split it into separate prompts. Tight scope = sharper output, less drift, fewer mistakes.

05 — QUICK REFERENCE

Do's & Don'ts

Pin this to the wall. It's the whole playbook compressed into two columns.

✓ Do
  • Brief it like a teammate: who / what / how
  • Give it a clear role or persona
  • Say what "good" looks like — length, tone, format
  • Paste an example of what you want
  • Add context: the audience, the goal, the situation
  • Ask it to think step by step on hard tasks
  • Give permission to be unsure
  • Keep one job per prompt
  • Iterate — treat the first reply as a draft, then refine
✕ Don't
  • Type three vague words and expect magic
  • Say "make it better" (better how?)
  • Stack fake "elite genius expert" keywords
  • Cram five tasks into one message
  • Assume it knows context that's only in your head
  • Blindly trust it — always check the output
  • Add more and more rules when it drifts (split instead)
  • Give up after one bad reply
  • Paste confidential data you wouldn't email externally
06 — STEAL THESE

Copy-Paste Recipes

Real prompts for real work. Swap the bracketed bits for your details and go. Each one already bakes in the principles above.

Sales & Outreach

Turn a messy call into a follow-up

When: you just got off a discovery call and need to send something today.
PromptYou're my sales assistant. Below are my rough notes from a call with a prospect. Write a short follow-up email: warm, specific to what they said, one clear next step. Under 130 words. If a detail is missing, leave a [bracket] for me to fill — don't invent it. Notes: [paste your messy notes]
Marketing

One idea, five channels

When: you have one announcement and need it everywhere.
PromptYou're a content strategist. Take the announcement below and adapt it into: (1) a LinkedIn post, (2) a short tweet, (3) a one-line email subject. Keep my voice — direct, no hype. Match the length norms of each platform. Announcement: [paste the news]
Operations

Summarize a long thread into decisions

When: a 40-message email chain landed on you and you need the "so what."
PromptRead the thread below. Give me: (1) the decisions that were made, (2) the open questions still unanswered, (3) who owns what. Bullet points, no fluff. If something is ambiguous, flag it rather than guessing. Thread: [paste the thread]
Analysis

Pressure-test your own thinking

When: you've made a decision and want to catch the blind spot before someone else does.
PromptYou're a sharp, skeptical advisor. Here's a decision I'm about to make: [describe it]. Think step by step. Give me the three strongest arguments against it, the one thing I'm probably underestimating, and what you'd need to see to change your mind. Be direct, not diplomatic.
Anyone

Explain it like I don't have the background

When: someone sent you something technical and you need to actually understand it.
PromptExplain the concept below to me like I'm a smart person who has never worked in this field. Use one everyday analogy. Then tell me the one thing I'd need to know to talk about it confidently in a meeting. [paste the thing]
07 — TROUBLESHOOTING

When It Goes Wrong

It will, sometimes. The output misses, or the AI can't find what you meant. Here's the reflex that fixes 90% of it — and it's almost never "write a cleverer prompt."

When the answer is off, the fix is usually add context — not add cleverness.
  1. Add the missing context. Tell it what it couldn't have known — "this is for a non-technical audience," "look at the second attachment," "we're a healthcare company, so tone matters." Nine times out of ten, the gap was information, not phrasing.
  2. Say "explain it like I'm slow." If you're the one who's lost, ask it to re-explain the last step in plain language. No shame, no wasted time — it's the fastest way to unstick yourself.
  3. Show it broken vs. working. Stuck on something visual or structural? Paste an example of what it should look like and say "here's the target, here's what I got — help me close the gap." A comparison beats a description every time.
  4. One change at a time. Don't rewrite the whole prompt in frustration. Change one thing, run it, see what moved. Batch-changing means you never learn what actually worked.
  5. Split it if it's drifting. If the AI keeps wandering off-task, the answer is rarely "add more rules." It's usually "you asked for too much in one go." Break it into two focused prompts.
08 — QUALITY CONTROL

Trust, But Verify

This is the one habit that separates people who use AI safely from people who get burned by it.

Judge the output against reality — not against the AI's claim that it worked. The AI will always sound confident. Confidence is not correctness. Before you send, ship, or act on an answer, do the ten-second check: are the facts right? Do the numbers match the source? Would I stake my name on this?

The reflex to build

For anything that matters, ask yourself two questions before trusting it: "What does a correct answer look like?" and "Have I actually checked this one against that?" If you can't define what "correct" looks like, you can't tell whether the AI got there.

This isn't fear — it's fluency. Use AI for the first draft, the heavy lifting, the blank-page problem. Keep your judgment for the final call. That combination — the AI's speed plus your accountability — is exactly where the value lives.


09 — TL;DR

The Cheat Sheet

If you remember nothing else from this page, remember this table.

The moveWhat to actually do
ReframeBrief it like a teammate, not a search engine
The formulaWho you are · what you want · how you want it
RoleStart with "You are a…" — it sets everything
SpecificVague in, vague out. Name the length, tone, format
ExampleShow a sample instead of describing it
ReasoningSay "think step by step" for hard tasks
Honesty"If unsure, say so — don't invent"
FocusOne job per prompt. Split when it drifts
When stuckAdd context, not cleverness
AlwaysVerify the output against reality before trusting it

None of this requires code. It requires clarity. You already have that skill — now you know how to point it at the most capable teammate you'll ever brief.

The goal was never a clever prompt.
It's a result you can measure.
Yuval Keshtcher
Yuval Keshtcher
Building AI agents. Teaching people to build.

Grounded in patterns from real client builds, plus published prompting guidance from Anthropic's Claude documentation and prompt engineering best practices.