AI and Documentation

Using LLMs to Assist with Documentation 

TL;DR 

AI has always shortened the amount of time it takes to produce documentation, allowing developers to vastly increase the amount of time they can spend working on integration solutions and producing deliverables that we provide to our customers. As the LLMs improve not only in their intelligence level, but also with their integration capabilities, the tedium of documentation is further reduced, which subsequently increases developer productivity. 

LLM Selection 

In this day and age of the AI boom, there are a whole host of companies each of whom have multiple LLM models to choose from, each seeming to have an area they specialize in to set them apart from their competitors.  

Gemini for instance, is automatically integrated into our Google accounts, assisting with emails and chats, suggesting revisions, autocompleting common phrases, and summarizing Google search results in a succinct way.  

ChatGPT, one of the first of its kind, is touted as a great general purpose AI tool that provides detailed information on just about every topic imaginable, with the ability to carry on a natural conversation and have your responses tailored to your exact word choice or use case. 

Copilot was groundbreaking when we gained the ability to integrate it directly into our IDEs, offering advantages such as context based code completion and suggestion, the ability to write and/or modify code for you directly in your development environment in real time, rather than having to cherry pick the few useful lines of information from a larger response another LLM may have given you. 
 
All of these advantages together have made successful strides to increase developer productivity, but the one that has caused the biggest wave in my opinion, is Claude by Anthropic. I initially was quite tied to ChatGPT for a long time as my all-purpose go to for technical questions and I was introduced to Claude as a replacement for that sole purpose. 

Within mere weeks of making the switch, I not only noticed Claude’s advantages in intelligence and reasoning, but also with its further integration capabilities, namely its ability to be integrated into my terminal, allowing it access to all the same info and context that I as a developer have when working on a project. As such, its ability to read, write, review, and create detailed documentation from nothing, far outclasses any of the other LLMs I’ve used in the past. 

Claude and Documentation 

No developer I know enjoys having to write documentation. It must be neat, accurate, thorough, and all encompassing, and ensuring our documentation abides by these standards takes a significant amount of time and care. But it’s something that we at Behaim have made our customers grow to expect from us as a standard part of our services, so it’s a job that must be done as diligently as the integration that it accompanies. 

In the past I have used LLMs to assist with documentation, but it still involved quite a bit of thought on the authors’ part, as well as quite a bit of copy paste, and particularly reformatting in a way that is legible and presentable. This is a task now trivialized by Claude (and more specifically Claude Code) and is something that, in my opinion, sets it far apart from other LLMs. 

Using Claude Code to directly integrate into my terminal, and thus giving it the ability to read the full context of my file system (or project root of whatever I may be working on), it can quickly scan and understand why something was implemented a certain way, and to describe how it works. Whether the project has the simplicity of a few script files or is comprised of many microservices with dependencies on other files in other directories, Claude Code is able to quickly extrapolate and summarize all this information and put it into an organized document in a matter of minutes, compared to what might take a developer hours or even days. 

 
We typically provide our customers with different types of documentation throughout the course of the project, be it an as is analysis at the start, describing the current state of the client environment before we are involved, or a “readme” or a “how-to” in the event that we migrate a client to a new platform or tech stack that requires a knowledge transfer. Whatever the case is, be it readme, a guide, technical documentation, or anything in between, Claude Code has been able to cut down hours to minutes, collectively saving me and other developers valuable time that can now be spent more productively on more technical tasks. 
 

Documentation in Practice 

The ability of Claude to provide top tier documentation is particularly topical for me as the past 3 weeks have been spent analyzing and documenting a new client environment. The start of any project will always be documentation heavy, and this was my first time starting a new project with Claude Code in my toolkit. 

Reformatting this table took less than 30 seconds with the help of Claude Code, you can see my question to it at the top specifying which file I want it to look at, where in the file, and what I want it to do 

All the time I’ve saved by using Claude Code to write and format my documentation has allowed me the flexibility to multitask on both the documentation, while being able to get a jump start on the actual implementation in tandem, which otherwise would have taken weeks separate from the documentation, putting me far ahead on where I would be without Claude. 

Overall, documentation is just one area of many in a programmer’s day to day that Claude has been able to streamline that I can speak to from personal experience. In just a few short weeks it’s already saved me numerous time and headaches and I am not the only developer who feels this way, and I am sure my use of it will only get more frequent.