ChatGPT Writing Tips: Prompting, Editing and Quality Checks

Most disappointing ChatGPT output traces back to the prompt, and most of what's left traces back to skipping the edit.

The single highest-value change a ChatGPT user can make is to put constraints in the prompt: an audience, a voice, and at least one specific input the model could not have generated itself. A prompt that supplies a real example gets specific output back; a prompt that just names a topic gets the average of everything written about it.

Why a default prompt gives you a mediocre draft

Type "Write a 1000-word blog post about email marketing" and what comes back is technically complete and substantively empty:

• Advice the reader already knows
• The same shape every time: intro, five tips, conclusion
• No specific examples and no point of view
• A flat, uniform sentence rhythm

The output isn't wrong. It's just the statistical centre of the topic, which is exactly what you would expect from a system predicting the most likely next word. Fixing it takes work in two places, before the draft and after it.

Part 1: prompting for specific output

Give the model a persona and an audience

Instead of "Write a blog post about email marketing," try: "You are a senior email marketing strategist with ten years at SaaS companies. Write for startup founders who know the basics but want better open rates. Use a direct, no-fluff voice. Include at least three specific tactics with concrete numbers where possible."

The persona narrows the vocabulary and the audience constraint narrows the claims. Both push the output away from the average.

Ask for the angle everyone else misses

"What is the most counterintuitive or non-obvious insight about this topic that most guides miss? Write the post around that insight."

This works because it explicitly penalises the safest prediction, which is the one the model reaches for by default.

Supply your own material

"Here is a specific example I want you to use: [your example]. Write a post that treats it as the central case study and draws lessons for [audience]."

Once real specifics are in the prompt, the model has something to reason from and can't retreat into generalities. This is the technique with the largest effect on output quality, and it is also the slowest, because you have to bring something.

Refine across turns

Never take the first output as final. Useful follow-ups:

• "Make the opening more specific and cut the scene-setting."
• "Shorten every paragraph by 30%."
• "Replace transitions like 'furthermore' and 'additionally' with plain connectives."
• "Add a concrete data point to each section, and tell me which ones you are unsure about."
• "Rewrite the conclusion so it adds something the body didn't say."

Ask it to strip its own tells

"Rewrite this section to use contractions, vary sentence lengths, and remove phrases that feel formal or corporate."

This helps at the surface. It does not reliably fix the structural uniformity, which is why the edit pass below still exists and why a dedicated rewriter handles the mechanical part more consistently.

Part 2: the edit pass

Raw output needs a human pass. Seven changes cover nearly all of it.

Step	What to delete	What to add
1	The generic opening paragraph, which almost always starts with a clause about the modern landscape	Start on your second or third sentence, which usually has the actual point in it
2	Claims stated in general terms	One thing per section you know from experience: a client case, a mistake, a result
3	Uniform paragraph length	Deliberate variation: one single-sentence paragraph, one of six
4	Corporate-speak: "leverage," "seamlessly," "robust," "scalable," "holistic approach," "cutting-edge," "game-changer," "paradigm shift"	Plain English equivalents
5	Agentless passives: "was found," "is believed," "has been shown"	The agent: "researchers found," "most experts believe"
6	"do not," "it is," "you will," "we are" where the register allows contraction	"don't," "it's," "you'll," "we're"
7	Anything that sounds wrong read aloud	A rewrite of that sentence, done by ear

Step 4's list is a specimen of the vocabulary to remove, not a style to imitate. Step 7 is the one people skip and the one that catches the most, because a sentence that survives your eye rarely survives your voice. For the longer version of this work, including what to do with a whole draft rather than a paragraph, see how to humanize AI content. If the problem is length rather than rhythm, making writing more concise covers the cuts worth making first.

Part 3: checks worth running

Two tools help after the edit, and neither of them produces a verdict you should publish against.

The text improver flags wordy constructions, agentless passives and noun-heavy phrases so you can see where the draft is still padded. The pattern estimate on this site reports how varied your sentence lengths are and how many stock phrases remain; it describes prose and cannot establish who wrote anything, so treat it as an editing aid rather than a gate. There is no score to aim for and no band that means a draft is finished.

What does tell you a draft is finished: you can name the specific detail in every section, every number has a source you have checked, and reading it aloud doesn't make you wince.

Workflows by content type

Blog articles

Outline and first draft in the model, then manual editing for examples, clichés and structure, then a rewriter pass on the sections that are still stiff, then a proofread. The manual editing takes the longest and is worth the most.

Social posts

Generate ten variations on a theme, pick one, and rewrite the opening hook yourself. The hook carries the post, and it is the part the model does worst.

Email newsletters

Let the model draft the body. Write the subject line yourself. Add a specific recent observation at the top, because that is the part subscribers open for.

Product descriptions

Give the model your specs and three real customer reviews before asking for anything. The output improves in proportion to the quality of what you put in.

Common mistakes

Using AI for the wrong parts

Models are good at structure, coverage and getting past a blank page. They are bad at personality, contrarian positions, recency, your audience, and your voice. Split the work accordingly.

Publishing unread

Unread AI output is how wrong statistics, non-existent studies and invented quotes reach print. Read every word before it ships.

Treating a confident answer as a checked one

The model will state a statistic that doesn't exist, in the same tone it uses for one that does. Verify every specific claim against a real source, and delete the ones you can't verify rather than softening them.

Reusing one template

If every post you publish has exactly five tips, your format is doing none of the work. Vary it: case studies, interviews, frameworks, contrarian takes, tutorials.

---

Frequently asked questions

Does Google rank AI content lower?

Google's stated aim is to reward helpful content regardless of how it was produced, while its spam policies target content made at scale to manipulate rankings. Our guide to AI content and SEO works through what the policies actually say.

Should I disclose AI use?

In journalism and academic writing, disclosure is often required or expected, and many publications now have written policies. When you are unsure, disclose.

How do I get my brand voice out of it?

Paste two or three examples of your own best writing into the prompt and ask for the same style and voice. Concrete samples work better than adjectives describing your voice.

Which model should I use?

Use the most capable model your plan gives you for anything you intend to publish, and a faster one for throwaway drafting. Specific model names date within months, so treat any ranking you read as a snapshot rather than advice.

What to do first

Take your next prompt and add three things to it: who the reader is, what voice you want, and one specific example of your own. Then budget as much time for the edit pass as you spent generating the draft. The trade-off to expect is that AI-assisted writing moves the work rather than removing it, from producing sentences to judging them.