My experiences with AI and Claude Code in 2025

I figured, for the history of the universe, that I would write down some of my experiences with AI. I have this feeling that people think that talking about AI usage is a little taboo. However, these tools have taken our field and the world by storm. People should write more about their experiences with them. I put the year 2025 in the title to make it more timely for end-of-year wrap-up posts, but I’ll add some pre-2025 background also

My (mostly pre-2025) impressions with AI

My first experience with AI was being incredibly skeptical. I really disliked OpenAI…I still do not use any of their products, and I really dislike seeing prose that I can tell was generated with it.

However, early on I started seeing projects everywhere on GitHub that require an “OPENAI_API_KEY”. I did not use these, and did not really feel like I needed to, though there was clearly some “blood in the water” making everyone go crazy

My early feelings about AI also involved being upset that AI couldn’t “cite its sources”. I had this idea in my head that AI was a galaxy brain…or should be a galaxy brain… that could draw from an immense pool of knowledge. However, one thing that shifted my thinking on this was reading “4.2 Gigabytes, or: How to Draw Anything” https://debugti.me/posts/how-to-draw/

This made me realize AI are ‘compressed representations’ of a lot of knowledge…weird!

My early experiences with using AI chatbots

Despite my OpenAI stance, once Google came out with Bard, I started asking it some questions. It was pretty interesting.

Some early experiences I had with Bard were trying to get it to convert code from one language (python) to another (JS). It was very good at doing this, which was wild to me. These language translations would take a lot of time to do by hand… and it would spit it out in seconds!

Early experience with text editor integration: Avante.nvim

My first experience with ‘buying’ into the AI was purchasing 20 bucks of API tokens from Anthropic, and using it with Avante.nvim (tagline: “Use your Neovim like using Cursor AI IDE!”). I liked the minimalism of Avante.nvim, and got stuck with the keybindings, and didn’t really want to try any others like CodeCompanion. You could launch a chat sidebar with <leader>aa or get it to edit a block with visual select <leader>ae

I had some pretty unique successes with Avante.nvim and Claude. It fixed some bugs parsing a very complex file format (CRAM). I can’t imagine it had much ‘training’ data on this, but it came up with a very successful solution to a weird bug, and I posted a full chatlog on my early PRs using this (example Dec 2024, https://github.com/GMOD/cram-js/pull/150 please note that all things considered this is now a very basic early example of my AI usage)

Around February 2025, Claude introduced 3.7 which added tool calling, which are basically ‘agentic’ workflows. These were integrated with Avante.nvim but they were not super reliable. I was pretty unhappy with the tool calling because it seemed like it did not work well. I reported several bugs related to this on the Avante.nvim github (e.g. “bug: Avante Gets stuck on tool calling and never progresses past that point”, “bug: Sonnet just runs in loops, reading the same file over and over again”, etc.). It also costs more in API tokens

In any case, one rainy day in May 2025, when Avante.nvim was really not working and I felt very stuck, I think I realized that maybe Avante.nvim was causing me more grief than help…see also Footnote 2

Switching to Gemini CLI and Claude Code

Not sure what exact date, but eventually, I decided to try Claude Code. I watched a video on YouTube almost right after it came out in February 2025, https://www.youtube.com/watch?v=AJpK3YTTKZ4, and I immediately saw that Claude Code could be very powerful, but I didn’t actually jump in and try it. It still seemed maybe “not for me” as I had my Avante.nvim integration.

However, I think once I saw Gemini CLI get released, I was like, ok I should try that out. The Gemini CLI has a free tier. Gemini CLI did ok, but it seemed not quite as powerful as Claude. I had also already used Claude via the Avante.nvim integration, so I switched to Claude Code

I started making Claude Code do very challenging tasks, and it was quite successful on many of them. One of the most mindboggling things it has done to date for me, and there are many, is that it completely reached into a very complex codebase called Bandage, a C++ codebase, and managed to extract the graph layout functions and compile it using emscripten to WASM and use it in Javascript. This produced https://github.com/cmdcolin/bandagejs

BandageJS is a demo that I still need to work on, but I was pretty mindblown by this

Switching from API billing to subscription

I am not a big fan of subscriptions, but I also started burning through a lot of API tokens. I’m talking like…it burned through 50+ bucks to do that bandageJS project even though it only did it in a couple hours. That cost is high, but at the same time, it would literally make breakthroughs that I would (likely) never have done.

In any case, to try to bring ‘down the cost’ from API billing I decided to change to a subscription at some point.

I bought the basic 20 dollar tier of Claude, and ran out of it immediately. Then I upgraded to Claude Max and now am at 100 bucks a month. That’s where I am “happily” sitting now.

I can sometimes ‘max out’ the Claude Max tier, but it is probably the right level for me. This is a weird thing to ‘confess’ but I buy the subscription myself, and do not get it re-imbursed through my work. Might need to have that talk at some point with my boss.

How I use Claude Code

Running multiple Claude Code sessions at once

I frequently have 2-3 Claude Code sessions going in different terminals, and I assign them distinct tasks

I have seen some people use ‘git worktrees’ to enable parallel Claude Code sessions, but I generally just use multiple separate clones of the repo

This is a little chaotic on the brain, as you have to mentally juggle multiple distinct tasks, but it lets you literally solve like…2-3x the tasks

Asking Claude Code to simplify, refactor, and identify bugs repeatedly

After almost every question, I will ask it different questions like

• “Are there any simplifications or refactorings that you would make”
• “Are there any other changes you would suggest”
• “Are there any tests you would add”
• “Are there any bugs you can identify”
• “Are there any correctness issues you can identify”

Maybe there is more optimal wording but in general it can identify some reasonable simplifications that are good-to-have

I sometimes also add things like “You can create helper functions” to clue to it that this is an option

Similar idea from the “Rule of Five” mentioned here. Essentially just repeatedly telling the AI to check itself https://steve-yegge.medium.com/six-new-tips-for-better-coding-with-agents-d4e9c86e42a9

Telling Claude Code to find optimizations

I work with ‘big data’, and being fast is important. A lot of the optimizations that Claude Code recommends do not move the needle at all, but if you set up a benchmark, as I wrote about previously https://cmdcolin.github.io/posts/2025-11-24-simplebench/, you can start to measurably achieve speed improvements

Telling Claude Code to solve merge conflicts

Solving merge conflicts by hand is tricky. I have historically used the Meld graphical interface, but even so it is finicky, laborious, and I can get into a mess. Instead, I can ask Claude to “solve merge conflicts” and it’ll do it quite well from my experience so far

Telling Claude Code that it can add debug logging

This is a very important one

Often, Claude will think very hard about how to solve particular issues. Who knows what’s going on in that mechabrain.

However, like a human programmer, adding debug logging can reveal a lot of useful info

it is often good, and even necessary, to let Claude know that it can add debug logging. Then you can either manually report the debug logging back to it, or it can automatically see the logging by making it run the tests or whatnot

Telling Claude Code to vendor entire NPM packages

Our app uses a variety of NPM packages, and some of them are quite old by now.

Sometimes this doesn’t really matter, but when you start to ‘ossify’ and you can only use version 1.0.x of some package because of reason Y, it is really annoying. One way to plow through this is you can ask Claude Code to vendor the entire package, and subsequently make bugfixes or improvements, update it to use typescript, remove unused functionality, and more

I read this essay sometime in 2025 and it was influential to me regarding vendoring: https://htmx.org/essays/vendoring/

I am a big fan of using libraries, but the open source library ecosystem is threatened by giant hacks like Shai-Halud (https://www.cisa.gov/news-events/alerts/2025/09/23/widespread-supply-chain-compromise-impacting-npm-ecosystem)

Vendoring can maybe help, but you also become responsible for maintaining this ‘dependency’ yourself. See also https://nolanlawson.com/2025/11/16/the-fate-of-small-open-source/

Telling Claude Code to solve github issues

You can literally paste a URL of a github issue into Claude Code, and it will do its best to go and solve it. Sometimes it actually works. Amazing

Don’t be too specific in your prompts if you don’t actually know what you want

Some people say to be very specific when talking to AI tools, but I have found very adequate results from NOT being specific. I just state the problem, even in vague terms, and see what the AI comes up with

This can often be better or simpler than what I might have prompted it with.

This is in line with the ‘be careful with what you wish for’ doctrine: don’t tell it exactly what to do if you don’t actually know what’s best, you can just see what it suggests

It is hard to tell if this has an impact but I am often pretty wishy washy with Claude, I’ll say things like “if needed, do X” or whatnot. I imagine that the power of suggestion of even typing that will strongly steer the direction it takes, but I still like to just give it the option to decide against the suggestion

Concluding thoughts

This is only a brief snapshot into my Claude Code usage. AI (applied to everything these days) is scary, exciting, dystopian…it’s a lot

I generally use Claude Code to work on a large (100k+ lines of code) project that I’ve worked on for many years, and I have a good mental model of how it works.

It may be more troublesome to work on new things that you have no experience with, as you may not learn or develop the right mental frameworks for understanding the code

AI coding is a superpower you can use, but it is also a superpower your competition can use to leapfrog you and your work… However, so much AI usage is derivative drivel that rips off what others create.

In any event…Good luck out there, and happy holidays

Footnote 1

The rabid anti-AI sentiment on Reddit (generally criticizing it for being inadequate, which increasingly is not true), combined with increasing AI-generated text and rage/engagement bait, made me quit Reddit and social media in general in mid 2025.

Now, I just watch random YouTube videos if I need a dopamine hit. I feel bad that I won’t be able to help out random strangers that are asking Reddit questions, but I have more time and don’t have my brain halfway into a random internet argument.

Footnote 2

This was also a little bit of a harbinger of things to come. When the AI is not working…or you run out of tokens…you don’t have the wind behind your sails anymore

Footnote 3

This post was inspired by https://nolanlawson.com/2025/12/22/how-i-use-ai-agents-to-write-code/

His post includes a note about customizing your CLAUDE.md

My ~/.claude/CLAUDE.md is fairly minimal. I instruct it not to run tests and stuff because it can take a long time on the large-ish (100k LOC) codebase I work on. Exact copy of my CLAUDE.md:

Please do not run tests, lint, format, or build manually. Just ask me to run
them and report results back. Use yarn v1 instead of npm

For typescript, avoid adding explicit return types and avoid using 'as' to cast
types where possible

We use noUncheckedIndexedAccess in most of our tsconfig, so be aware of this

Be minimal about commenting code, if the code is self explanatory it needs no
comment

Prefer node --experimental-strip-types over tsx or other solutions

I tend to not like "early return" style code

Do not use push(...list), it can cause stack overflow

If we are in the process of debugging an issue, please do not remove debug
logging unless instructed

Footnote 4

I link to Yegge here because his post on CHOP-oriented dev (https://sourcegraph.com/blog/the-death-of-the-junior-developer) was one of the first posts to make me realize that CHOP aka vibecoding is rapidly changing our field.

Footnote 5

If you are a AI skeptic, or someone that complains that they always spit out garbage, I can only say, you have to use at least Opus 4.5

The other models will frequently come up with bad ideas. Sometimes I will downgrade in Claude Code (due to thinking my task is simple or running out of credits) and select e.g. Sonnet or Haiku, but if I forget to change it back, it’ll continue making really bad output and I’ll be like wtf is this? This is terrible!

If you want a good experience, you have to use the most powerful model that you can. I imagine this is not environmentally friendly, but you will waste more of your time chasing down ridiculous bugs if you do anything non-trivial with weaker models