Published: July 22, 2026 | Last updated: July 22, 2026

Why Your Prompts Suck (and How to Write Better AI Prompts)

If you’ve ever typed a request into ChatGPT and gotten back a flat, generic answer, the problem probably wasn’t the model. Learning how to write better AI prompts is the single highest-leverage skill you can build right now, because the gap between a mediocre AI user and a sharp one isn’t access to a better tool. It’s four missing pieces of information the model never had.

ChatGPT alone serves roughly 900 million weekly users, according to TechCrunch’s February 2026 reporting on OpenAI’s funding announcement. Most of those people are typing one-line requests and expecting the model to read their mind. It can’t, and it was never going to.

Break The Ordinary has covered the broader picture in how to work with AI effectively, the operational layer in what changes when solo founders run agentic AI, and the applied layer in turning plain English into working spreadsheet formulas. Once this framework becomes automatic, setting up Claude Projects is the natural next step so you stop rebuilding the same context every time.

Table of Contents

• Why Does ChatGPT Give Generic Answers to Your Prompts?
• What Is the 4-Part Fix for Writing Better AI Prompts?
• How Do You Diagnose a Prompt That Already Failed?
• What Do Before-and-After AI Prompts Actually Look Like?
• What Mistakes Should You Avoid When Writing AI Prompts?
• How Does Prompting Differ Between ChatGPT, Claude, and Gemini?
• Frequently Asked Questions About How to Write Better AI Prompts

The fastest way to write better AI prompts is to give the model four things every time: context, a single specific task, the output format you need, and clear constraints on tone and length. Skip any one of the four and it fills the gap with its safest, most generic guess.

Quick Takeaways

• Bad AI output is almost always a prompt problem, not a model problem.
• The 4-part fix: Context, Task, Format, Constraints (CTFC).
• Heavy role-prompting (“act as an expert”) is now outdated for current models.
• One well-chosen example beats three extra paragraphs of instructions.
• Break big requests into smaller, sequential prompts instead of one giant ask.
• Diagnose old bad outputs by checking which of the 4 parts you skipped.

Why Does ChatGPT Give Generic Answers to Your Prompts?

ChatGPT gives generic answers when your prompt is under-specified, because the model can’t infer unstated context and defaults to the safest, most average response it can generate. This happens the same way in Claude and Gemini.

The fix isn’t a smarter model. It’s a more complete prompt.

The Root Cause Is Under-Specification, Not the Model

When you write “write me a LinkedIn post about pricing,” you’ve handed the model a topic, not a task. It has to guess your audience, your angle, your length, and your tone all at once. Faced with four unknowns, it picks the statistically safest answer for all of them – which reads as generic because it’s built to offend no one and fit anyone.

This pattern is independently confirmed across the official guidance from Anthropic, OpenAI, and Google – three competing labs that rarely agree on anything except this. Prompt engineering for beginners often starts with tricks and hacks. Learning how to write better AI prompts should start with this root cause instead.

What Is the 4-Part Fix for Writing Better AI Prompts?

The 4-part fix is a prompt engineering framework called CTFC: Context, Task, Format, and Constraints. Give a model all four and you get a precise, usable answer almost every time; leave one out and you’re back to guessing.

Google’s “Gemini for Google Workspace” prompting guide publishes a named four-part structure: Persona, Task, Context, and Format (PTCF). Anthropic and OpenAI don’t ship a single acronym, but their official best-practice guidance maps onto the same four elements under different names. That convergence across three competing labs is stronger evidence than any single blog post could be.

BTO’s version folds “Persona” into Context, because Anthropic’s own November 2025 guidance now lists heavy role-prompting – telling a model to “act as a world-class copywriter” – as outdated overhead for current models. Tone gets folded into Constraints instead of standing alone. Four parts, not five, is easier to remember and matches what the labs actually converge on functionally.

Source: Break The Ordinary – based on official prompting guidance from Google’s Gemini for Google Workspace PTCF guide, Anthropic’s prompt engineering best practices, and OpenAI’s prompt engineering guidance

Context – Who’s Asking and Why

Context means the model knows who this is for, what the goal is, and why the output matters. Anthropic’s own documentation recommends explaining “why something matters,” not just what you want. If a stranger read only your context sentence, they should know who this is for and why it matters.

Google’s PTCF framework folds this into its Persona and Context elements. In practice, one sentence usually covers it: “I’m a solo SaaS founder writing to other founders who struggle to raise prices without losing customers.” That single sentence eliminates most of the guessing a vague prompt forces onto the model.

Task – One Specific, Explicit Action

Task means naming a single instruction with a direct action verb – write, analyze, summarize, compare – instead of describing a topic. Anthropic recommends being explicit and skipping preambles: state the action first. If you can’t describe the task in one sentence with a verb at the front, you haven’t defined it yet.

Format – How the Output Should Be Structured

Format is the part of AI prompt structure most people skip – it tells the model the shape of the answer, not just the content. Google’s PTCF framework names Format as a standalone element for exactly this reason. Bullets, prose, a table, a Slack message, an email – each demands a different structure, and the model can’t guess which one you need.

Constraints – Boundaries, and One Example If It Matters

Constraints cover word limits, tone, and what to avoid – plus, when style is hard to describe, one well-chosen example. Anthropic’s documentation calls examples one of the most reliable techniques for controlling output, more reliable than paragraphs of written instruction in many cases. Even one example can turn a generic answer into exactly the voice you wanted.

The 60-Second CTFC Checklist

Before you send any prompt, run it through these four questions in order.

1. Context: does the model know who this is for and why it matters?
2. Task: is there one clear action verb, not just a topic?
3. Format: have you said what shape the answer should take?
4. Constraints: have you named the limits, or shown one example?

If you can answer yes to all four, send it. If not, that’s the exact part to fix – not the whole prompt, just that one piece.

How Do You Diagnose a Prompt That Already Failed?

Pull up your last disappointing AI response and check which of the four CTFC parts you skipped – that gap is usually the entire fix. Most guides only teach forward construction. This reframes prompting as a troubleshooting skill, not just a template to fill out.

Did the model misunderstand the audience or the goal? That’s a Context gap.

Did it do the wrong thing entirely, or try to do too many things at once? That’s a Task gap.

Was the structure wrong – too long, too short, or in the wrong shape? That’s a Format gap.

Did it ignore a boundary, or default to a bland tone with no example to follow? That’s a Constraints gap.

Most of the time, exactly one of these four is missing. That’s how to write better AI prompts without starting over: fix the one piece, resend it, and watch the output snap into place.

What Do Before-and-After AI Prompts Actually Look Like?

Seeing the CTFC framework applied to a real prompt makes it concrete faster than any explanation. Below are two examples: a vague one-line request, and the same request rebuilt with all four parts.

Example 1 – Writing a LinkedIn Post About Pricing

Before:

Write me a LinkedIn post about pricing.

After (CTFC applied):

I'm a solo SaaS founder writing to other founders who struggle to raise prices without losing customers. Write a 150-word LinkedIn post arguing that raising prices too late is more dangerous than raising them too early. Format it as 4 short paragraphs, no bullet points, ending in one question to drive comments. Keep the tone direct and slightly contrarian, and avoid generic advice like "know your worth." Here's an example of the voice I want: [paste one past post].

The first version gives the model a topic. The second gives it an audience, a specific argument, a format, a tone, and an example to match – which is exactly why the output changes.

Example 2 – Summarizing a Client Call

Before:

Summarize this meeting transcript.

After (CTFC applied):

This is a transcript of a client kickoff call. I need to send a recap to my team who wasn't on the call. Summarize the 3 decisions made and the 2 open questions still unresolved. Format as two short sections with headers "Decisions" and "Open Questions," each with a max of 3 bullets. Keep it under 120 words total, and don't include small talk or scheduling logistics.

This is a smaller task than the LinkedIn post, but the structure is identical: who needs this, what to extract, how to shape it, and where the boundaries are. That same structure is what makes tools like turning plain English into spreadsheet formulas work reliably instead of returning something close but unusable.

Example 3 – Writing a Product Description

Before:

Write a product description for my coffee mug.

After (CTFC applied):

This is a 12oz ceramic mug for our Shopify store, sold mostly as a gift for coffee lovers. Write a 60-word product description that highlights the double-wall insulation and the matte glaze finish. Format it as one short paragraph, no bullet points, ending with a single sentence about gifting. Keep the tone warm and simple, avoid words like "premium" or "artisan," and match this voice: [paste one existing product description].

Same four parts, a completely different task: an audience and a goal, one clear instruction, a fixed shape, and a boundary on both length and vocabulary. That’s the difference between a 60-word description that reads like it belongs on your site and generic ad copy that could be pasted onto anyone’s.

What Mistakes Should You Avoid When Writing AI Prompts?

The biggest mistake is assuming the AI is just bad at the task, when in the overwhelming majority of cases the prompt was under-specified. These five mistakes are what keep people from writing better AI prompts, more than any model limitation.

Believing More Detail Is Always Better

Guidance from Google DeepMind’s Phil Schmid on Gemini 3 warns that long prompts with unnecessary filler reduce output quality, not improve it. Precision beats volume. Say what matters once, clearly, and stop.

Role-Playing the AI Into Being an “Expert”

Telling a model to “act as a world-class copywriter” was standard advice in 2023. Anthropic now lists heavy role-prompting as outdated overhead for current models. Fold that same information into Context instead – “writing for a technical audience who hates jargon” does more work than a costume.

Front-Loading One Giant Prompt Instead of Chaining

Anthropic, OpenAI, and Google’s documentation all independently recommend breaking a complex task into smaller, sequential prompts instead of cramming everything into one shot. This is essentially what agentic AI workflows do at scale for solo founders – chain several focused prompts instead of one overloaded one. Three competing labs agreeing on the same point is not a coincidence.

Giving Up After One Bad Response

Every major lab treats iteration as the default workflow, not a failure state. OpenAI’s own help documentation frames prompting as draft, review, refine, repeat. Quitting after the first attempt means stopping one step before the actual fix.

Treating Examples as Optional Polish

Few-shot examples are repeatedly cited as one of the single highest-leverage techniques for controlling tone and format. Skipping them and relying only on written instructions leaves the model guessing at style. One example, pasted directly into the prompt, often outperforms three extra sentences of description.

How Does Prompting Differ Between ChatGPT, Claude, and Gemini?

The CTFC framework works the same way in all three tools, but each lab weights the four parts slightly differently in its own guidance. How to prompt ChatGPT effectively isn’t a fundamentally different skill from prompting Claude or Gemini well, just a different emphasis.

None of the three require a different mental model. All of them reward the same four inputs – they just differ in how much filler they can tolerate before quality drops.

Frequently Asked Questions About How to Write Better AI Prompts

What is the fastest way to write better AI prompts?

The fastest way is to run every prompt through CTFC before sending it: context, task, format, constraints. Most disappointing outputs are missing just one of the four, so checking all four takes seconds and saves a rewrite.

Why does ChatGPT give generic answers even when I explain what I want?

ChatGPT gives generic answers when your explanation covers the topic but skips one of the four CTFC parts, usually format or constraints. The model fills that gap with its safest, broadest guess rather than the specific one you had in mind.

Do I need to use role prompting like “act as an expert” for good results?

No. Anthropic’s current guidance lists heavy role-prompting as outdated overhead for its models, and the same audience information works better folded into Context. Say who the output is for instead of asking the model to play a character.

What’s the difference between prompt engineering and just talking to AI normally?

Prompt engineering for beginners is really just talking to AI with the four missing pieces added back in – audience, task, format, and boundaries. It’s less a separate technical skill than being as specific with a model as you’d be with a new employee.

Does the CTFC framework work the same way in Claude and Gemini as it does in ChatGPT?

Yes. All three labs’ own documentation converges on the same four elements, even though only Google’s Gemini for Google Workspace guide ships a named acronym (PTCF). The emphasis shifts slightly by tool, but the underlying logic of how to write a good AI prompt doesn’t change.

How do I know if my prompt is too long?

If your prompt repeats itself, hedges with filler phrases, or buries the actual instruction in the third paragraph, it’s too long. Google DeepMind’s Phil Schmid, writing on Gemini 3, specifically warns that unnecessary length reduces output quality rather than improving it.

How I Know This

I built the multi-agent AI pipeline that writes and publishes every article on Break The Ordinary, which means I write dozens of production prompts a week, not a handful for fun. When a prompt fails, it doesn’t just produce a bad paragraph – it breaks a step in a live system that has to run again the next day.

That pressure taught me the CTFC pattern faster than any course could have. I didn’t come to this with a coding background. I came to it by testing what actually made an AI agent’s output usable on the first try, then repeating whatever worked.

Every mistake in this article is one I made first, usually more than once, before I fixed it by adding the missing piece back into the prompt.

Prompting Well Is a System, Not a Talent

Learning how to write better AI prompts isn’t about being naturally clever with words. It’s a system you can run every time: context, task, format, constraints, checked in that order. That’s the same principle behind everything Break The Ordinary teaches – real independence comes from building repeatable systems, not from hoping you get lucky.

The people getting real leverage from their AI subscriptions right now aren’t the ones with the fanciest prompts. They’re the ones who stopped guessing and started checking the same four boxes every time.

If this is the first time you’ve thought about your AI workflow as a system instead of a chat window, how to work with AI effectively is the natural next read – it covers the collaboration layer this article’s prompt mechanics sit underneath.

About the Author

Randal | Break The Ordinary

I’m Randal, the founder of Break The Ordinary – a multi-niche media brand covering business, tech, health, and finance for people who want to build wealth, freedom, and a life worth living. I built and run the multi-agent AI pipeline that produces every article on this site, so I test how to write better AI prompts under real production stakes every week, not as a theoretical exercise. I share what actually works, what doesn’t, and what most people get wrong – my approach is direct, research-backed, and built on real experience, not theory.

×