Building a Writing Assistant That Actually Helps

I started wondering if there was a way to offload some of the mechanical parts of writing. Not the thinking. Just the getting it down.

So I built something.

The idea was simple. Take the ChatGPT API and wrap it in a small interface that could handle the kinds of writing I do most often. Blog posts. Ad copy. Social media captions. Things with purpose. Things that need to sound like me but also need to actually work.

I am not a developer by trade. But I know enough to be dangerous with an API key and a text editor. The setup took an afternoon. The refining took weeks.

What emerged was not a content factory. It was more like a writing partner that never gets tired and never judges the first draft.

One of the first things I noticed when using raw ChatGPT was the voice problem. The default output has a certain texture. Useful but generic. It reads like someone who knows a lot but has never actually made anything.

I wanted the tool to understand difference between a LinkedIn post and a product description. Between a casual newsletter and a landing page that needs to convert.

So I built a tone selector. It started with three options: casual, professional, and warm. Over time, that grew into a small menu of voices I actually use. The casual one sounds like how I write here. The professional one strips out the casual asides. The warm one leans into empathy without getting sentimental.

It is not perfect. But it gets me seventy percent of the way there. I take over from that point.

I do a fair amount of content that needs to rank. Nothing aggressive. Just writing that has a chance of being found.

The challenge with keyword optimization is that it often ruins the flow. You end up with sentences that are technically correct but feel forced. Like someone is checking a box.

The assistant now has a keyword field. You drop in a few terms. It weaves them into the draft in a way that does not disrupt the natural rhythm. Sometimes I have to nudge it. Most times, it works on the first pass.

I learned that the key is giving it context. Just feeding a list of keywords leads to clumsy results. But if I also paste in the surrounding paragraphs, the output stays cohesive.

I am not a fan of rigid templates. They tend to produce work that feels like it came from a template.

But I found that loose structural guides work well. A blog post outline with sections for problem, solution, and example. An ad copy structure with attention, interest, desire, action. Social media formats that follow a pattern without being formulaic.

The templates are stored as prompts behind the scenes. When I select one, the assistant fills in the structure with the content I provide. It saves me from starting with a blank page. That is often the hardest part.

I open the tool. I pick what I am writing. I choose a tone. I paste in any notes or research I have. I add keywords if they matter.

Then I hit generate. The text appears in a few seconds. I read it. I cut about thirty percent. I rearrange some sentences. I add a few lines that only I could write. Personal details. Specific observations. Things the tool does not know because I have not told it.

That second pass is where the writing happens. The first pass just gives me something to work with.

I have stopped staring at cursors. That alone was worth the effort.

The assistant is not a replacement. It is a starting point. The best results come when I treat it like a junior writer who needs clear instructions and a final edit.

I also learned that the API costs are minimal. A month of heavy use runs me less than a cup of coffee each week. The time saved is significant.

There were moments where I tried to automate too much. When I let the tool write full drafts without intervention, the results were usable but flat. The magic is in the combination. My thinking plus its output equals something that moves faster than either alone.

I use a simple web interface that connects to OpenAI’s API. Nothing hosted publicly. Just a local tool that sits in my browser. There are plenty of no code options that do similar things if you prefer not to build.

What matters more than the setup is the prompt library. I spent time crafting prompts that yield consistent results. Each one includes instructions about voice, structure, and what to avoid. That small investment made everything else work.

The tool still struggles with nuance. It does not understand brand voice the way I do. It cannot tell when a joke will land or when to hold back. It writes well but does not know my audience the way I know them.

I also have to watch for repetition. Sometimes it latches onto a phrase and uses it three times in two paragraphs. I catch that in editing.

And it does not have opinions. I have to supply those. The opinions are what make writing worth reading.

I built this for myself but I have started letting a few friends use it. The feedback has been consistent. They say it removes the friction of starting. That is exactly what I wanted.

I am not selling it. I am not building a business around it. It is just a tool that lives in my workflow now. I use it almost every day. Not for everything. Just for the tasks that benefit from a faster first draft.

There is a certain calm in knowing the mechanical part is handled. I can focus on the part I actually enjoy. The shaping. The refining. The small choices that turn a draft into something worth sending.

I do not know if this approach would work for everyone. But for me, it has been quietly useful. No hype. No big promises. Just a little less friction and a little more time for the work that matters.

That is the kind of tool I can get behind.

A few people have asked how I structured the prompts for different tones. I will probably write about that next. The prompt engineering part turned out to be more interesting than I expected.

The keyword piece took the longest to dial in. I tried different approaches. Pushing keywords into headings. Spreading them across the first paragraph. Eventually I settled on a method that looks for natural insertion points rather than forcing placement.

It works best when I provide three to five keywords. More than that and the text starts to feel crowded. Less than that and the optimization is barely noticeable.

I have added and removed templates over time. The ones that stayed are the formats I use weekly. The ones I removed were either too specific or too rigid. A good template gives you a container. It does not tell you what to put inside.

What surprised me was how much the templates changed my own thinking. Even when I am not using the tool, I find myself reaching for those structures. They became part of how I organize ideas.

The numbers are rough estimates. I did not track everything perfectly. But the time saved is real. I used to block out two hours for a draft. Now I block out one. The other hour goes to editing and adding the details that make the writing mine.

I think that is the part I value most. Not the speed. The space it creates for the kind of work I actually want to do.