How to Use ChatGPT to Come Up with a Blog Post Outline

When you’ve been blogging as long as I have, you fall into one of two categories. Either you’re constantly looking for ways to speed up your processes and get more out of every minute of your time, or you’re always seeking new knowledge and techniques and internalizing them to speed up your own process.

I tend to bounce from one side to the other. I seek out and learn new things, but I also love exploring new tools and incorporating them into my process – or stripping them out if they prove to be ineffective in the long term.

Obviously, one of the biggest new tools to hit the stage in recent years is generative AI and LLMs. ChatGPT being the biggest name among them, it gets a lot of attention, though other LLMs and more specialized tools also exist.

There are a million ways to use generative AI or your blogging, but a lot of them have risks. I would never recommend using ChatGPT to generate blog posts out of whole cloth, for example, because they tend to be very generic, boring, and often wrong about facts due to the way LLMs work. It ends up taking more time and effort to take LLM output and turn it into something valuable than it would to just create a blog post the normal way. Unless you don’t care about any of that, but I find that ignoring content quality is not a viable long-term strategy.

What I find a lot more useful is using an LLM to generate an outline or set of ideas that you can then use your human brain, skills, and knowledge to turn into a good blog post. It helps a lot with writer’s block, with finding the connective tissue to cover an issue from all of the angles you want to cover, and can even point out things you might not have thought about.

The question is, how do you do it? Unfortunately, it’s harder than most people think. You can’t just go to ChatGPT and say, “Write me an outline for a blog post about X,” and expect something worthwhile.

In my mind, using ChatGPT (or a similar broad LLM system) to generate blog post outlines is a two-phase process.

Phase 1: Competitive Research

The first phase actually has nothing to do with ChatGPT at all, but you can use other AI-powered tools if you want.

To start, you need a basic idea of a topic. You can just ask an AI for topic ideas, but you’re going to get basically the same results as if you typed “X topic ideas” into Google and looked at the top list, so that’s pretty basic and uninteresting.

One option is, of course, Topicfinder. I made Topicfinder for this explicit purpose, to scrape the topics and keywords from competition and authorities in a niche, to make it much easier to find ideas. If you want, you can just start a free trial and skip to the next phase, but I would recommend going a little deeper into your research.

For one thing, you want to know more than just the broad topic; you want to have some understanding of what your competition is for that topic, what angles they’re taking when covering it, and even what sentiment they’re using. All of this can help you either leverage consensus with Google or just find a way to engage with the topic at hand.

I have a list of blog post research tools I like over here, which you can also check out.

A lot of the modern AI-powered topic research tools do more than just throw an LLM at a problem. They pull in data like keyword value, competition estimates, traffic estimates, PPC costs for the keyword, and more. They can even do things like pull all of the subheadings from a post and even summarize the content, effectively taking live content and converting it into an outline so you can skim it quickly.

In particular, Perplexity is one I’ve found to be more useful than raw GPT output. It does a lot of your legwork by scanning all of the top Google results and providing not just information summarized from them but citations for the information it pulls.

I don’t find a lot of these AI tools to be very useful when I’m already familiar with a topic. But, if a topic is very fast-moving or fast-changing, or when it’s a subject I’m not as knowledgeable about yet (like for a new client I’ve just picked up), using AI to give me a birds’-eye view of the subject can be very helpful for initial research and awareness building.

Note that this is also why I like things like Perplexity; those citations are very important. Never trust AI output without looking into it yourself, especially if a “fact” doesn’t pass the sniff test.

As you build up all of this information, I recommend keeping a list of key points or facts you want to mention in a document. The AI you use for an outline later will either need that information fed into it, or you’ll need to cross-reference with it later to make sure you’re hitting all of the points you want to make.

Once you have a general idea of the topic, the details you want to cover, and the angle you want to take, you can start phase two.

Phase 2: Prompt Engineering

Phase two is where you turn to ChatGPT directly (or one of the guided LLM tools like Jasper) to generate your outline for you.

Prompt engineering is surprisingly tricky because it’s not as simple as just asking for an outline or even asking for an outline that hits your key points.

• You want something that doesn’t go off the rails.
• You want something that sticks to a reasonable, logical flow.
• You want something that leaves out fluff or points you don’t want to make.
• You want something that doesn’t run counter to your overall point of view.
• You want something that can be easily expanded into a full blog post.

You can get ChatGPT to output something that suits all of these needs, but you generally have to be very specific with your instructions. It can involve very lengthy and complex prompts, and even multiple back-and-forth iterations of your outline to get it refined the way you want it.

I have some tips on doing all of this, of course.

Be specific about the number of subheadings you want, and their length. One of the issues with just asking for an outline is that ChatGPT doesn’t know what an outline is. It knows what items in its training set have the flag “outline” on them, and what things have been referenced using the word outline, but the definition of a blog post outline is a lot narrower.

If left to its own devices, ChatGPT will happily go off the rails with its outlines. I’ve seen it repeat the same subheading several times (sometimes with slight variations, sometimes not). I’ve seen it put contradictory headings one after another. I’ve seen it write a subheading that’s more like 50-100 words long. It also loves to use common headings you see in a lot of blog posts, like “Recent Case Studies,” “Industry Insights,” or “References,” to take up space.

This is all based on how LLMs work. LLMs function by generating output that resembles existing human content, with some variation to that output, so it’s not just copy-pasting something else already out there. Since huge amounts of content out there have subheadings like Recent Case Studies or References, those have heavy weight in the LLM as things most of the content in that format has, so they get included.

Watch for overlap between points on the outline and specify to remove them if necessary. This one is also pretty common. LLMs don’t have any sort of flag or metric for what a discrete piece of information is. That means they don’t “know” when they’ve already covered a piece of information in their outline. For example, if you ask it to create an outline for a piece of software, it might make “how much does the software cost” and “software name pricing” as separate headlines.

Similarly, if you’re asking it to create an outline about something where one term is used in different ways, it might refuse to properly cover them both, in an overreaction to avoiding overlap. An outline reviewing an email manager might use the word “subscription” for both one of the main features and for the pricing structure, but GPT might not cover them both.

Basically, you often have to watch the output, compare notes with your desired output, and take a hammer to it if necessary.

Reinforce the core topic and main subtopics you want to talk about in your outline. LLMs also don’t have logic or a “thought process” the way you might think. The fact that they often output things in a similar order is reflective of the fact that most content they’re trained on has that ordering.

As with everything else I’ve mentioned, this is something the LLM can get wrong, and you need to specify a sentence or three in your prompt to fix it.

Don’t forget to specify “assumed” information. If you’ve ever tried to use a system like Jasper, you might notice that it asks you a bunch of sequential pieces of data and configuration options before it starts to generate output for you. Each of those is basically a GUI for a templated sentence or set of sentences that go into the back-end prompt.

That means things like your point of view, your sentiment, your perspective, and your tone of voice are all things you need to specify if you want output that reflects your style. You need less of this when you’re asking for an outline, but you do still need some of it.

Make sure your key points from your competitive research are covered. This is why I mentioned writing down all of the main points you want to cover; you can include them in your prompt to make sure they’re all covered.

Don’t forget to fact-check the output if necessary. This should be pretty obvious at a glance, but if the outline GPT gives you has a subheading clearly referencing something that is contradictory to your research, you’ll need to fix it.

Try out different GPT models and LLMs to see if the output works better for you and your topic. Whether you tab between GPT 4, GPT4o, or Claude, or even make use of a multi-agentic tool, you’ll find different results with different LLMs and even with different topics. Some will work better than others, so don’t be afraid to play around.

Once you’ve hammered out a prompt that works, copy it, streamline the back-and-forth into initial instructions, and genericize it into a template you can use for future prompts.

If you looked at all of the above and thought, “this doesn’t seem any faster than my usual outlining process,” that’s why. The initial prompt engineering will always be lengthy and time-consuming, but your goal is to get something that works with the quirks of the model and your topic that is then repeatable. That’s where the time savings come in. You’ll always have to play with it and tinker until it gives you what you want, but when 90% of the groundwork is done, that remaining 10% doesn’t take much time at all.

Leave a Comment