Claude Review
So back in 2024, I took a good look at AI assisted code generation using Chat GPT. There was some promise there but as everyone has seen over the last two years, things have progressed rapidly.
People are automating their jobs with agents, entire industries are adjusting. Software engineering as a career has changed from one where you spend most of your time coding to most of your time prompting and reviewing code.
The realy question that I’m asked a lot by my peers who still have to yet get on board, is “does it actually work? Does it write good code? Isn’t it all just … slop?”
The answer is more nuanced than that, and I think its worth digging into the details.
The Project
To kick this project off, I gave the following prompt to Claude.
Today, we will develop a complete authentication API service in Go.
This API service will use github.com/labstack/echo/v4 for the HTTP framework, and we will need to support account registration, authentication, authorization, and user profile management (both by the authenticated user, and by admin users) using a REST API. The REST API should be documented in an OpenAPI 3 document, and the API should serve the documentation via the swagger-ui from /docs.
For storage, I’d like to use the github.com/jackc/pgx/v5 database library with a postgresql database.
Users should be stored with their email address as the primary key, the password should use a strong one-way hashing algorithm like argon2. Users should be able to be a member of a Role, and there should be at least two roles initially defined, one for administrative users and one for normal users.
For context building, I installed the following plugins:
- code-review
- code-simplifier
- frontend-design
- superpowers
There are more plugins available which are useful but I find these are generally the most useful to run with on a daily basis.
After starting with the prompt, Claude created a plan, which I reviewed.
It asked me two questions, it originally specified options for the profile metadata which would have added columns for the metadata fields. I suggested to use dynamic schema and it added it to the design.
The other thing it asked for was the application architecture. Because the application was simple, it recommended a fairly flat architecture, but it also offered the Ports & Adapters pattern as an option, which is what I chose.
It then generated an implementation plan which goes into detail exactly how the service would be implemented. I reviewed, did not see anything objectionable so let it implement.
Thoughts
The service it generated was fully functional as described out of the box. I did ask it to add some Makefile targets for local testing, but other than that - it just worked.
I’ve been doing much more advanced things with Claude in my work, so this didn’t surprise me. But what I wanted to convey is just how powerful this is becoming. If you know what you want, driving Claude to a desired output is incredibly easy.
There’s a huge risk here of course; if you do not prompt carefully like me, Claude will happily go and build something that is not a good fit for your organization.
If you’re a software engineer, and AI tooling is not a part of your toolbox, you’re leaving a lot of the table. If you’re stubbornly resisting adopting any AI tooling because you feel it’s just generating “slop” I would encourage you to examine the codebase. I would feel pretty challenged to write something like this in a few days, let alone a few hours. And I’m pretty sure it would have more bugs.
The only things I could really see are a lack of rate limiting, and limits to post bodies etc. Nothing truly horrendous but things that are pretty common to see in real world implementations. And you could see how easy it would be to iterate on the service to close gaps. “Find security flaws. Fix them”.
Wherever you stand on “where AI is going” we should all acknowldege the power of the tools. Its here, its real, and people are using them, furiously, to build faster than you.
It doesn’t mean what they build is better - but given well constructed prompts it can get pretty damn good.