The 4-Part Goal Prompt, With a Real Before and After

Spell out every step and Claude follows your plan, even a bad one. This short prompt gives Claude the goal instead, lets it research and plan, has it interview you, and shows it what good looks like. Plus a real run of both ways, side by side.

A free guide by Zach Crawford

Most people write giant prompts. "You're a world-class copywriter. Step 1, do this. Step 2, do that." Nine steps, seven rules.

When you spell out every step, Claude follows your plan. Even when it's a bad one.

So the best you get back is your own idea, written up nicer.

Give it the goal instead, and let it make the plan. Four short parts.

The 4-part goal prompt

Open Claude Code in the folder for the job. Fill in the three brackets first: the goal, the example, and what you love about it. Then paste the whole thing.

Part 1. The goal
[What the finished thing looks like, in one or two lines. Not the steps.]

Part 2. Research first
Before you write anything, find the best way to do this job. If you can search the web, do it, and tell me in a few lines what you found and where. Then make the plan yourself and show it to me. Use research only for how to do the job. Every fact about me, my offer and my clients comes from me.

Part 3. Interview me
Then interview me. Ask one question at a time and wait for my answer. Ask only what your plan needs and you can't find in the files in this folder. Stop when you could do the job without guessing. If I don't know something, leave a [blank] instead of making it up.

Part 4. Show it good
Here is one real example I love:
[paste it]
What I love about it: [one line]
Use it as the standard for quality and feel. Don't copy its words, its facts or its length.

When you're done, check every line against what I told you and the files in this folder, story details included. If I didn't say it, cut it or mark it [check]. Then list every fact you used and where it came from.

Here's what each part does.

Part 1. Start with the goal

Say what the finished thing looks like, not the steps to get there. One or two lines.

"A sales email for my new program that gets the right people to join and the wrong people to skip it." That's a goal. "Open with an exciting announcement, then list the features, then the price" is a plan. Leave the plan out.

Put everything you'd count as a win in that line. If you want people to hit reply, say so here.

Part 2. Let it research

Ask it to find the best way to do the job before it starts. Then it makes the plan, not you.

You see the plan before it writes a word. Read it. If something's missing, say so now. A plan is a lot faster to fix than a draft.

Claude Code also has plan mode. Press Shift+Tab until it says plan mode is on. Claude researches and writes a plan. It changes nothing until you say go.

Part 3. Let it interview you

One question at a time, until it understands what you want.

This is the part everyone skips. It's where Claude finds out the things you'd never think to type into a prompt: who should skip your offer, what your people are scared of, which client story is true and OK to share.

Part 4. Show it good

Paste one real example you love. Now it knows the bar.

Use something real: your best email, the post that got the most replies, a page you wish you'd written. Then say in one line what you love about it. That line tells Claude what to match.

A real before and after

Same job, both ways, in a clean Claude Code with none of my files. Same model both times. The coach is made up: Glen, a running coach selling Half Ready, a 12-week group program for runners over 40 training for their first half marathon. 15 spots, $597.

Before: the giant prompt. A typical one. It opens "You are a world-class email copywriter with 20 years of experience," then gives 9 steps in order: three subject lines "with curiosity and urgency," open with "an exciting announcement," introduce the coach, list the features, make the price "feel like a great deal" next to a one-on-one coach, add a client story, "create urgency," a button, and a P.S. about the deadline. Then 7 rules, like "enthusiastic, motivating tone." All of Glen's facts were in it.

Claude did every step. It ran twice and came out almost the same both times. It opened like this:

🎉 It's finally here: Half Ready is open!

What else came back:

• A made-up price claim: "One-on-one running coaches often charge that much per month." Glen never gave it a price for that.
• Promises he can't keep: "feeling strong, pain-free" and "Your turn could be next."
• Pressure: "When they're full, they're full, so grab yours today."
• Nothing that tells the wrong people to skip it.

That's the prompt's plan, done well. It was the wrong plan for Glen.

After: the 4-part prompt. Part 1 was one line: "A sales email to my list for my new 12-week coaching program, Half Ready. It should get the right people to join and the wrong people to skip it." Part 4 was his best past email, a short one he loves.

What Claude did:

1. Searched the web. It said up front the results were mostly blogs, "so treat them as practitioner opinion, not hard evidence," and used them only for the shape of the email.
2. Showed its plan. The plan had a part the giant prompt never asked for: who should skip it.
3. Asked 7 questions, one at a time: who it's for, who it's not for, what's in it, price and dates, which true story to open with, whether the client OK'd her name, and how he signs off.
4. Wrote the email, checked every line against his answers, and listed every fact with where it came from. It left out two lines it couldn't back up, one of them a line from his example email.

Here's the email it wrote:

Subject: She quit two plans. Then she finished smiling.

Linda quit two training plans because her knee wouldn't cooperate.

Last year she trained with me. She finished the Columbus half in 2:41 with no pain.

Her words: "I finished smiling. I didn't know that was possible for me."

That's exactly why I built Half Ready.

It's a 12-week coaching program for runners 45 to 55 who've done a 5K or 10K and want to finish a half marathon. They're worried their knees won't hold up, or that they'll get hurt like they have before. Maybe you've tried other plans and either quit or limped through.

Here's what you get:

• A personalized training plan that I adjust every two weeks
• A live group call every Tuesday night (recorded)
• A private group chat
• Video form reviews from me
• A 15-minute strength routine for knees and hips

You'll run four times a week: three shorter runs and a long run on the weekend.

Skip it if:

• You want to break 2 hours. You need a different plan.
• You're hurt right now. See a doctor first.
• You can't commit to 4 runs a week.

This only works if you've got the time and you're healthy enough to start.

It's $597 upfront, or three payments of $219. We start February 1. Doors close January 22, and there are 15 spots. No application. Click [LINK] and pay to join.

Run easy, Glen

How they compared. Both emails went to a fresh Claude with Glen's fact sheet. It didn't know which prompt wrote which. It ran twice, with the order swapped. Both times it said to send the 4-part email, and said nothing in it was made up. Scores out of 10:

What the 4-part email still missed. His refund policy, and an invite to hit reply if you're not sure. Claude never asked about either, and Glen never said, so they're not in it. That's why you read the plan in Part 2, and why everything you'd count as a win goes in Part 1.

When a long, detailed prompt is still the right call

Long prompts aren't always wrong. Write the steps out when:

• You already know the right steps and the order matters. A checklist, a process you've run many times, a form that has to come out the same way every time. Anthropic's prompting guide says to use numbered steps when the order or completeness of the steps matters.
• The format is fixed. A product listing with set fields, an email that has to fit a template. Tell it exactly.
• It's a repeat job you've already proven. Then don't retype it. Save the steps in a file and point Claude at it.
• It's a small job. "Fix the typo in line 3" doesn't need research or an interview.

Use the 4 parts when you know what you want but not the best way to get there. That's most of the writing, planning and offer work in a coaching or course business.

The context lives in files, not in a giant prompt

Every AI agent I run reads my business folder first.

That's why the prompt can stay short. Your offer, your voice, your best examples and your rules sit in files in the folder, and Claude reads them. Claude Code reads a file called CLAUDE.md in that folder at the start of every session.

So when the interview pulls out something new, like who your offer is for or who should skip it, ask Claude to save it to a file in that folder. Next time it has fewer questions to ask.

Where this comes from

• Anthropic's Best practices for Claude Code says to say up front what the result should look like, to have Claude interview you when you haven't worked it all out yet, and to hand over the real thing instead of describing it.
• Anthropic's Prompting best practices says a general instruction often gets better reasoning than a hand-written step-by-step plan, and that examples are one of the most reliable ways to steer Claude.
• Plan mode in Claude Code has Claude research and write a plan before it changes anything.

The limits, honestly

• It takes longer than one prompt. A search and a handful of questions come before the first draft. Worth it for a sales email, a page or an offer. Not for a quick reply.
• The interview only gets what it asks about. Read the plan in Part 2 and add what's missing.
• Web research finds blog posts, not proof. That's why the prompt says research is only for how to do the job. Facts about you come from you.
• One example is a start. Anthropic's prompting guide suggests 3 to 5 examples that differ from each other. If the work starts to look like a copy of your example, add a second one that's different.
• You're still the last check. Read the fact list at the end before anything goes out.

Want the next guide? Follow @iamzachcrawford