DEPTH · CLAUDE · 03 OF 05

Prompting Claude well

Three skills that change everything

15 min read·3 of 5 in the Claude track

There's a particular feeling when prompting starts working. The same question that gave you a generic, slightly-off answer suddenly returns something useful. Same tool. Different prompt. The gap between those two answers is the entire skill.

Most people never close that gap. They blame the tool, get bored, and quit. The people who do close it stop thinking about prompting as a chore and start using it like breathing.

Three skills get you there. None of them are clever tricks. All of them transfer to ChatGPT, Gemini, and whatever comes next.

The three skills below work in any chat AI. We teach in Claude because the failure modes are cleanest. Notes at the end of each section flag any tool-specific differences.

PART 01

Skill 1 — The anatomy of a useful prompt

Most beginner prompts are too short and too vague. 'Write an email.' 'Plan a trip.' 'Explain X.' The AI does its best with what you gave it, which is usually a generic answer that doesn't fit your situation.

Useful prompts have four parts. You don't need all four for every prompt — but you need to know which one's missing when an answer comes back wrong.

1. Context — what you're working with

What's the situation? Who are you? What's the audience? What's already been tried? The AI has no idea about any of this unless you say so. Three sentences of context usually beats a thousand words of detail.

Bad: 'Help me write an apology.'

Better: 'I run a small bakery. A regular customer is upset because we got their wedding cake order wrong on a date that mattered. They've been a customer for two years. I want to make this right.'

2. Goal — what you want at the end

What's the outcome? An email, a list, a decision, a draft, an explanation? Say it explicitly. The AI optimizes for whatever shape you name, so name the right shape.

Bad: 'Help with this situation.'

Better: 'I want a one-paragraph apology email that acknowledges the mistake, offers a specific gesture (free cake credit for a future occasion), and asks how we can make things right.'

3. Constraints — what would make a bad answer

What should it avoid? What length? What tone? Constraints are the most underused part of prompting. Saying 'don't make it longer than three sentences' is more useful than asking for 'a short email.'

Bad: 'Make it good.'

Better: 'Warm but not groveling. Three sentences maximum. No corporate-speak. Don't make excuses; just own it and offer the gesture.'

4. Example — what good looks like

Optional but powerful. If you have an example of the tone you want, paste it. If you have a previous email that worked well, show it. AI is much better at matching examples than at imagining what you mean by 'professional but warm.'

Bad: 'Sound like me.'

Better: 'Here's how I usually talk to customers: [paste 2-3 sentences]. Match that tone.'

Put it all together and you get a prompt that's specific enough to do real work. Not every prompt needs all four parts — but every weak prompt is missing at least one.

PART 02

Skill 2 — The iteration loop

The biggest misconception about prompting is that you write one perfect prompt and the AI returns one perfect answer. That's not how skilled people use these tools.

Skilled people prompt in loops. The first response is a starting point, not a final draft. You react to it, push it in the direction you want, and end up somewhere better than you could have specified up front. It's a conversation, not a vending machine.

The loop has five moves:

Describe what you want (the prompt).
AI responds. You read it.
Notice what's wrong, missing, or off. Be specific about which part.
Tell the AI exactly what to change. The more specific the feedback, the better the next round.
Repeat until the answer is good enough to ship.

The skill is in step three and four. Most beginners give vague feedback ('this isn't quite right,' 'make it better') and get vague improvements. The people who get great results give precise, surgical feedback.

Effective feedback

'The second paragraph is too long. Cut it to two sentences.'
'The tone is too apologetic. Make it confident without being arrogant.'
'You added a fact about [X] that I didn\'t mention. Remove it.'
'This works, but reorder the paragraphs so the offer comes first.'
'Good — now do three variations of the opening line.'

Ineffective feedback

'This isn\'t right.' (Which part?)
'Make it better.' (Better how?)
'I don\'t like it.' (Specifically?)
'Try again.' (With what change?)

If you find yourself giving the same vague feedback twice and getting the same wrong answer, the issue isn't the AI. It's that your feedback isn't specific enough for the AI to act on. Step back, articulate exactly what's wrong, then tell it that.

PART 03

Skill 3 — Magic phrases that consistently work

Certain phrases consistently produce sharper, more useful responses. They're not tricks. They work because they ask the AI to do something it does well but doesn't do by default. Keep them in your back pocket.

'Explain your reasoning.' Forces Claude to show the steps that led to its answer. Often surfaces where the logic is shaky.
'What could go wrong with this?' Surfaces risks and edge cases the default answer skipped. Best for plans, decisions, and code.
'What are the tradeoffs?' Gets a balanced view instead of a one-sided recommendation. Useful when Claude is being enthusiastic.
'Is there a simpler way?' Cuts through over-engineered or jargon-heavy responses. Often returns the actually-useful version.
'What would I miss with this approach?' Forces the response to acknowledge blind spots, not just sell its own answer.
'Push back on this.' Gets you out of yes-man mode. Combine with a specific claim or plan.
'Give me three versions of [X].' When you don't know which direction is right, comparing variations beats picking from one.
'Explain it like I'm new to this.' Defangs jargon. Great for technical answers or unfamiliar domains.

None of these are magic spells. They're requests for shapes of thinking Claude can produce on demand but won't produce by default. Use them like vocabulary — pull the right one when you need it.

TAKEAWAY

The 80/20 of prompting

If you remember only this: most prompting skill comes from three habits. Give Claude enough context to do useful work. Iterate with specific feedback instead of vague disappointment. Reach for a magic phrase when you want a different shape of answer.

Everything else is decoration. Practice these three and you'll be in the top 10% of users by your second week.

Try it now

Three exercises, one per skill. Pick anything from your life to fill the brackets — a real situation, not a hypothetical. Real friction makes the practice stick.

Practice 1 — Build a full-anatomy prompt

Use all four parts: context, goal, constraints, example.

Context: I [WHO YOU ARE] and I need [WHAT YOU'RE TRYING TO DO]. The situation: [2-3 SENTENCES OF DETAIL]. The audience: [WHO IT'S FOR].

Goal: I want [SPECIFIC OUTCOME — what shape, how long, what should it contain].

Constraints: Don't [WHAT TO AVOID]. Keep it [TONE / LENGTH / STYLE].

Example: Here's something in the tone I want — [PASTE 2-3 SENTENCES].

Now give me a first draft.

Practice 2 — Iterate with surgical feedback

(Use this after Claude responds to anything.)

That's a starting point. Three specific things to change:
1. [ONE SPECIFIC THING]
2. [ANOTHER SPECIFIC THING]
3. [A THIRD SPECIFIC THING]

Don't rewrite the whole thing — just fix those three.

Practice 3 — Use three magic phrases in one conversation

(After Claude has given you any answer.)

Round 1: Explain your reasoning for the recommendation above.

Round 2: What could go wrong with this approach? Be specific about the failure modes.

Round 3: Is there a simpler way to achieve the same outcome?

Notice how each one changes the shape of the answer.