AI Writing Prompts That Produce Usable Drafts (And Why Most Do Not)

Most AI writing prompts fail for one reason. They tell the model who to be, but give it nothing to push against.

An AI writing prompt is the instruction you give a tool like Claude or ChatGPT to produce a piece of writing. For bloggers and creators, it decides whether you get a draft you edit in 20 minutes or one you rewrite from scratch.

The difference is rarely the tool.

It sits in five blocks that most prompts leave out. I will show you all five, using one weak prompt that I rebuild in stages, with the output at every step.

Why I Stopped Blaming the Tool

In my experience, bad AI drafts come from thin prompts, not weak models. When I fixed my prompts, the same model gave me drafts I could publish after one edit pass.

Here is how I learned that, and it started long before AI.

My first job was in a contact center, selling mobile phone plans over cold calls. I made zero sales in my first 40 days. I was close to losing the job.

My team leader did not tell me to "be a better salesperson." That would have been useless. He gave me two specific things to change in my calls.

Then I practiced with written role-plays for 15 days. Real objections and real replies.

Something concrete to react against.

Within 10 months I was the top salesperson on the floor, with the lowest returns.

That is exactly the gap in most AI writing prompts.

"You are an expert content writer" is the same as "be a better salesperson." It sounds like direction. It gives the model nothing to work with.

I have been writing for a living for over 20 years. I run 14 niche blogs and I built two software products, Threadeazy and VoiceLab. Every article on this blog now starts as a Claude draft, built from a long, structured brief.

The brief is the whole game. Let me show you why.

What Makes AI Writing Prompts Produce Average Drafts?

AI writing prompts produce average drafts when they only define a role and a task. The model fills every gap with the most common choice it has seen. The most common choice is average by definition.

Think about what the model knows when you type a short prompt. It knows the topic. It does not know your reader, your voice, your length, or what you refuse to say.

So it guesses. Every guess lands on the middle of the road.

Here is the weak prompt I will rebuild in this post. I used a version of this myself in early 2024.

You are an expert content writer. Write a blog post about email marketing for beginners.

That is the role-plus-task prompt. It is the prompt shape I see most often.

What a weak AI writing prompt gave me

The opening talked about the "digital landscape" and called email "a cornerstone of modern marketing." It listed five benefits, each one a bolded word followed by one vague sentence.

Nothing in it was wrong.

Nothing in it was usable either. It read like every other post on page three of Google, because that is roughly what it was trained on.

The fix is not a better role. The fix is more information in specific places.

The 5-Block Prompt Template for AI Writing Prompts

The 5-block template splits a prompt into Situation, Job, Fences, Sample, and Checkpoint. Each block closes one gap the model would otherwise fill with an average guess.

I did not invent these blocks in one sitting. They came from about two years of reading bad drafts and asking one question: what did the model not know here?

Here is each block, with the job it does.

1. Situation. Who reads this, where it will be published, and what the reader already knows.
2. Job. The one outcome this piece must produce, stated as a result, not a topic.
3. Fences. What the model must not do: banned phrases, length limits, format rules.
4. Sample. A short piece of real writing the model can react against, usually yours.
5. Checkpoint. A step the model runs before writing, like listing assumptions or asking questions.

Now watch the weak prompt change as each block goes in.

Stage 1: Add the Situation block

The Situation block tells the model who is reading and what they already know. This single block removes most of the beginner-level padding.

Here is what I added:

The reader is a 30 to 40 year old professional in India with a full-time job. They already know what an email list is. They have tried a free Mailchimp account and stopped after two weeks. This post goes on a personal blog, not a company site.

The output changed right away. The definitions disappeared. The draft started talking about why people abandon their list after the first two weeks.

That is a real topic. The first draft never touched it.

Stage 2: Add the Job block

The Job block states the outcome, not the subject. "Write about email marketing" is a subject. "Get the reader to send their first email this week" is an outcome.

My addition:

The job of this post: get the reader to send their first email to their list within 7 days. Everything in the post should serve that.

This one line cut two sections from the draft. The history of email marketing vanished. So did the section on advanced segmentation.

The model now had a way to decide what belonged. Before, it had no way to leave anything out.

Stage 3: Add the Fences block

The Fences block lists what the model must not do. Most prompts skip it, and it is the block that removes the "AI sound" fastest.

My fences for this post:

Do not use these words: landscape, cornerstone, unlock, elevate, crucial. No sentence over 20 words. No paragraph over 3 sentences. No bulleted list of benefits. Do not open with a definition.

The draft now opened with a moment. A reader staring at a list of 43 subscribers, scared to email them.

Fences work because the model defaults to the patterns it sees most. You have to name those patterns to block them.

I keep a running list of mine. If you want to build your own, my post on how to humanize AI content has the line-level checklist I use.

Stage 4: Add the Sample block

The Sample block gives the model a short piece of real writing to match. This is the block that moves a draft from "clean" to "sounds like you."

This is the role-play lesson from my sales job, applied to prompts.

I pasted three paragraphs from one of my own older posts, then wrote:

Here is a sample of how I write. Match the sentence length, the directness, and the use of short one-line paragraphs. Do not copy the content, only the style.

The difference was obvious in the first paragraph. Short lines, a question to the reader, and a specific number instead of a general claim.

Look, the model is very good at imitation. It is poor at inventing a voice from an adjective like "conversational." Give it the real thing.

Samples also fix most of the speed problem. I covered that side in how to write faster with AI without sounding like everyone else.

If you use Claude, you can store your samples once instead of pasting them every time. I explain that setup in how to use Claude Projects to keep your content sounding like you.

Stage 5: Add the Checkpoint block

The Checkpoint block makes the model stop and show its thinking before it writes. It catches wrong assumptions before they become 2,000 words of wrong draft.

My checkpoint:

Before you write, list the 5 assumptions you are making about the reader and the post. Then write a 6-line outline. Wait for my approval before drafting.

The first time I ran this, assumption number three was wrong. The model assumed the reader wanted to sell a product by email. My reader just wanted to start.

I corrected one line. The draft that followed needed one edit pass.

That is the whole difference between prompting and briefing.

The Full Prompt, Before and After

The weak prompt was 14 words long. The rebuilt prompt runs to about 280 words, and it saves me close to an hour of editing per article.

Here is the rebuilt version in one piece, so you can copy the shape.

Situation: The reader is a 30 to 40 year old professional in India with a full-time job. They know what an email list is. They tried a free Mailchimp account and stopped after two weeks. This goes on a personal blog.Job: Get the reader to send their first email to their list within 7 days.Fences: Do not use: landscape, cornerstone, unlock, elevate, crucial. No sentence over 20 words. No paragraph over 3 sentences. No bulleted benefit lists. Do not open with a definition.Sample: [Paste 3 paragraphs of your own writing.] Match the style, not the content.Checkpoint: Before writing, list 5 assumptions about the reader. Then write a 6-line outline and wait for my approval.

Notice what is missing. There is no "You are an expert."

I dropped the role line entirely. I tested it on and off across about 30 drafts. Once the other four blocks were in, the role line made no difference I could see.

Does that mean role lines are useless? Not really. They help when a task needs a lens, like "review this as a skeptical editor." For drafting, the Situation block does that job better.

Best AI Prompts for Writing Compared: Swipe Lists vs Built Prompts

Swipe-list prompts are fast to use but produce generic output, because they cannot know your reader. Built prompts take 10 minutes to set up once, and then produce drafts that need far less editing.

Here is how the two approaches compare in my own use.

Swipe prompts still have a place. I use them for headline ideas and outlines every week.

But for a full draft, the built prompt wins every time I compare them.

This works if you write for the same audience repeatedly. It does not work well if every piece targets a different reader, because you rebuild the Situation block each time.

Claude Prompts for Content: What Changes Between Tools

The 5-block structure works in both Claude and ChatGPT. The differences are in how each tool handles long instructions and where you can store them.

Claude follows long, structured instructions closely. Anthropic's own prompt engineering guide recommends clear sections and examples, which is exactly what the Sample block gives it.

In my experience, ChatGPT drifts more on the Fences block. It will obey a banned-word list for the first 800 words, then slip.

The fix is simple. Put the Fences block twice: once near the top, once at the very end of the prompt.

Here is the thing about both tools. Neither one remembers your prompt between chats unless you store it somewhere.

• In Claude, store the Situation, Fences, and Sample blocks in a Project.
• In ChatGPT, store them in custom instructions or a saved document you paste in.
• In either tool, keep the Job and Checkpoint blocks fresh for each piece.

The Job changes per article. The rest mostly does not.

Common Mistakes With AI Writing Prompts (And How to Fix Them)

The most common mistake is writing a long prompt that is all Situation and no Fences. The model understands the reader perfectly and still writes like a brochure.

I made every one of these mistakes. Here are the four I see most in prompts readers send me.

Mistake 1: Describing your voice with adjectives

"Write in a friendly, conversational, witty tone" gives the model three adjectives and no evidence. It will produce its own average idea of friendly.

Fix: Delete the adjectives. Paste a sample instead.

Mistake 2: Asking for the full draft in one step

When you ask for 2,500 words straight away, you only find the wrong assumption at the end. By then you have a long draft built on it.

Fix: Always run the Checkpoint block first. Approve the outline, then draft.

Mistake 3: Stuffing the prompt with every rule you know

I once wrote a prompt with 41 rules in it. The model followed about 30 and ignored the rest, and I could not predict which.

Fix: Keep Fences to the 8 to 10 patterns that actually show up in your drafts. Add a rule only when you see the problem twice.

Mistake 4: Never updating the prompt

Your first built prompt will be wrong in small ways. That is normal.

Fix: After every draft, ask one question. What did I fix by hand that the prompt could have prevented? Add that line.

My main prompt has gone through more than 40 versions. The latest one is better than the first by a mile, and it only got there one line at a time.

Prompt Engineering for Writers Is Really Briefing

Prompt engineering for writers is the skill of briefing a writer who cannot ask you questions. If you have ever briefed a freelancer, you already know most of it.

I spent 23 years in corporate jobs, and a lot of that time I managed people. The worst briefs I ever gave were one line long. The best ones answered questions before anyone asked them.

A prompt is a brief for a writer with no memory, no context, and endless patience.

That last part matters. The model will happily redraft 10 times. So the only real cost of a thin prompt is your own editing time.

At 3 articles a week, a thin prompt costs me about 3 extra hours. Over a year, that is roughly 150 hours.

That is almost a full month of evenings.

Where Ready-Made Prompts Fit In

Ready-made prompts are the fastest way to see what a well-built prompt looks like before you write your own. They work best as starting points you edit, not as final answers.

I resisted building prompt templates for a long time. My view was that everyone should write their own.

Then I watched readers try to build their first 5-block prompt from nothing. Most of them stalled on the Fences and Checkpoint blocks, because they had never seen good examples of either.

So I packaged the prompts I use for different content jobs. The Creator's Super Prompt Library has 100 fully written prompts, organized by task, plus 56 bonus prompts for editing and voice-matching. Each one already has its fences written in, so you edit the Situation and Sample to fit your blog and go.

Claude skills go one step further and turn a prompt into a reusable command. The Content Creator's Claude Skill Stack is the set of skills I built for exactly that, with no coding needed.

Start with prompts. Move to skills when you notice you paste the same prompt every week.

FAQ

What are AI writing prompts?

AI writing prompts are the instructions you give tools like Claude or ChatGPT to produce written content. A weak prompt names a role and a topic in about 15 words. A strong prompt adds the reader, the goal, the rules, a style sample, and a check step, usually in 200 to 300 words. The strong version typically cuts editing time on a 2,000-word draft from over an hour to under 25 minutes.

What is the best AI prompt for writing a blog post?

The best AI prompt for a blog post follows a structure, not a magic phrase. Use five blocks: Situation, Job, Fences, Sample, and Checkpoint. The Sample block, where you paste 3 paragraphs of your own writing, makes the biggest difference to how the draft sounds. The Checkpoint block, where the model lists its assumptions first, prevents the most wasted drafts.

Why does AI writing sound so generic?

AI writing sounds generic because the model fills every missing detail with the most common choice it has seen. If you do not specify the reader, voice, or banned patterns, you get the statistical average of the internet. Naming 8 to 10 banned words and pasting a real writing sample removes most of the generic feel. In my tests, those two blocks alone changed the draft more than any role line did.

Should I start my prompt with "You are an expert"?

You can, but it adds very little once the rest of the prompt is solid. I tested role lines on and off across about 30 drafts and saw no clear difference once the other blocks were in. Role lines help more for review tasks, like asking the model to critique a draft as a skeptical editor. For drafting, describing the reader does more work than describing the writer.

How long should an AI writing prompt be?

A good AI writing prompt for a full article is usually 200 to 300 words. Shorter prompts leave gaps the model fills with average guesses. Much longer prompts, like my old 41-rule version, cause the model to drop rules unpredictably. Aim for 8 to 10 fences and add new ones only after a problem appears twice.

Do the same prompts work in Claude and ChatGPT?

The same 5-block structure works in both tools. Claude tends to follow long instructions more closely from start to finish. ChatGPT sometimes drifts on banned words after 800 or so words, so repeat your Fences block at the end of the prompt. Both tools forget your prompt between chats unless you store it in a Project or custom instructions.

Are free prompt lists worth using?

Free prompt lists are worth using for ideas, outlines, and headlines. They are weak for full drafts because they cannot know your reader or your voice. I still use swipe-style prompts every week for brainstorming. For anything I plan to publish, I use a built prompt with my own Situation and Sample blocks.

How do I make AI write in my own voice?

Paste 3 paragraphs of your own writing into the prompt and tell the model to match the style, not the content. Adjectives like "friendly" or "witty" do not work well, because the model supplies its own average meaning. Pick paragraphs that show your sentence length and how you open a section. If you use Claude, store those samples in a Project so every chat starts with them.

How many times should I revise an AI prompt?

Revise your prompt after every draft for the first month. Ask one question each time: what did I fix by hand that the prompt could have prevented? My main writing prompt has gone through more than 40 versions this way. Most improvements are single lines, not rewrites.

What is the difference between a prompt and a Claude skill?

A prompt is an instruction you paste into a chat each time. A Claude skill is a saved set of instructions that Claude can load on its own when a task matches. Prompts are the right place to start while you figure out what works. Once you paste the same prompt more than once a week, turning it into a skill saves the setup time.

Conclusion

Those 40 silent days in the contact center taught me something I still use every morning.

Nobody gets better from being told to "be better." You get better from specifics you can react to.

AI writing prompts work the same way. "You are an expert content writer" is a pep talk. Situation, Job, Fences, Sample, and Checkpoint are the role-play scripts.

Take one prompt you already use. Add just the Fences block today, and a Sample block tomorrow. Then compare the drafts side by side.

Which of the five blocks is missing from the prompt you used most recently?