How to use AI in your workflow as a tech writer

As tech writers we have a lot of tools at our disposal, and a lot of tools that we have to learn how to use. AI is one more tool that, used well, can be extremely powerful in your day-to-day tasks. So, assuming your company doesn't have a CEO that thinks they can replace people with AI, what does it mean exactly to use AI to help you? Do you have to be an engineer to learn this new skill? What can AI do for you?

Where to start

AI is a great tool to have in your roster. But it can also be intimidating. There are so many new concepts to learn: LLM, intelligence level, harness (yes, really), agents, skills… The list goes on. So what do you need to start?

Let's start with the basics: there are a few companies offering language models that differ in their capacities and intelligence. The best-known ones are Gemini (Google), Claude (Anthropic), and ChatGPT (OpenAI). Let's focus on these for now since there are too many to mention here.

There are two main ways of using these models: through their chat interface and as an agent. The difference between these lies in their capacities. If you chat with these LLMs, they answer your questions, but if you use them as an agent, they transform into something that can perform tasks for you. The difference lies in the harness - what the LLM has to interface with the computer, which provides tools to the AI model, like read, write, search the internet, etc.

Agents are where it's at

An agent is what tech writers should be using. Because agents are basically AI models with tools at their disposal, LLMs stop being brains without bodies to become little helpers that can perform a huge array of tasks for you.

If you start simple, you can think of an agent as a souped-up find-and-replace tool that, instead of blindly making substitutions in your documentation, can look at the context and only replace what you really want. From here, the sky is the limit. You can use an agent to:

• Enforce style guides
• Help you with awkward sentences
• Research PRDs and other technical docs that usually are scattered around and difficult to find
• Research support tickets to understand what customers are having trouble with, and where the documentation might need to be improved to help

Automating repetitive tasks

After a while using an agent, it's probable that you'll find yourself doing the same task over and over again. Instead of doing it, or typing the same prompt to your agent every time, you can automate repetitive tasks through skills.

Skills are markdown files that tell the agent how to do something, and what the goal is. For example, say that you have a ticket for a new feature. If your ticket has no details about what's needed, you'll know that one of the first things you need to do is investigate what the ticket is about, where you can find more information about the requirements, who the stakeholders are, etc.

Instead of doing this manually, you can create a skill to automate it all, and give you a summary with the information you require.

Now, I know what you're thinking: this seems complicated. Won't I need to learn how to code to do this? Not at all! What you need is to tell the AI what you need, and be as specific as possible. Since AI is an expert in a lot of fields, let it help you create a skill.

Don't read this as you need to think of every detail beforehand; think of it more as a conversation with a buddy that will help you with what you need. So, continuing with our ticket example, you could prompt it with something like:

I need you to create a skill that will research Jira tickets. This research needs to have:

• Who the stakeholders are
• What's needed
• Where are the PRDs or any other internal documents that can help understand the ticket
• Any public documents that already exist

Create a structured report with this information. The report should be factual. Do not make up information that does not exist.

This is only an example, but you get the idea. There are two other concepts you should be aware of, though, and they are really important.

MCP servers

Before you can let your agent loose on your company's internal databases, you need a way for it to connect and read that information. This is done through Model Context Protocol (MCP) servers that act as an interface between your agent and your company at large. Without it, your AI agent cannot just go and find information for you, or read your tickets. Ask your IT department about this if you don't have this information already.

Get a second pair of eyes

The other concept you need to be aware of is that AI agents are not very good at being self-critical. That is, when they gather information, they will be pretty happy telling you that the information they got you is the most accurate ever. So to make sure that what you got is really factual, you need a second pair of eyes, a skeptical agent, so to speak. So when you're creating your skills, ask the agent to spawn a subprocess with a fresh agent whose only job is to go and fact check everything that the report says, for example. Because this fresh agent spawns in a new process and has no context about what the previous agent did, it can be really useful to find factual errors, or incomplete information, or information that the first agent just made up. But it needs to be a different agent spawned in a new process, or it won't work.

The importance of the AI model

Different AI models have different capacities and levels of intelligence. For example, Claude Sonnet is less intelligent than Claude Opus. And different models within the family have different capacities as well. In my experience, Claude Opus 5 is more difficult to work with than the previous Opus 4.8.

Just because you're working with documentation, don't assume a lesser model is able to do the task. Documentation can be really complex depending on the tools you have, workflows, and a lot of other variables. And some AI models will struggle with the complexity. Test different AI models in your workflows and see the results you get before deciding on what's best to use. But I'd advise you to use more capable models whenever you can, especially if you're trying to make sure that the research you did is factual or that there are no errors in your documentation.

Wrapping up

These are just a few examples of what you can do with agents and the way you use them will really depend on your workflows. I would say to pay attention to the mechanical tasks you're constantly repeating and start there. Then you can graduate to research and other more complex workflows when you feel confident about the results you're getting.