AI Augmented Coding

by Brent Rager
December 20, 2025

I know there is a swath of folks in our industry that felt like AI augmented coding is infringing on our agency as developers. Causing our practice and our knowledge to "rot". I get that and those feelings are valid, but I don't think we have much of a choice.

Expectations for delivery, efficiency, and quality will continue to increase from c suite, product managers, etc.

There will be continued downward pressure on headcount as business analysts note that efficiency in software delivery should increase in the AI era.

In the last year I spent about 6 months straight building a one click self generating AI support agent with fully customizable embeddable website chat widget. I built everything solo from the ground up back to front. Frankly, it was the best thing I ever built. It didn't ultimately launch due to a rug pull, but the learning was invaluable. Everything from the customer experience, auth, config, delivery, payments, ci/cd, support experience, go to market, landing pages, sales conversion tracking. Everything.

Some of the most important learning I pulled from that experience was learning where Claude Code sped me up, and where it slowed me down. And there were definitely examples of both!

Bottom line is, as experienced developers who have been through the toil it takes to learn the intricacies of this profession, AI augmented coding is a superpower.

Where did it slow me down? Frankly for me the biggest problem with AI augmented coding was that it to some extent ruined my ability to estimate this project. There were several points along the line when I thought I was 95% done because parts "felt" close to done without much testing, but they weren't really. The last mile problem is not fixed by AI augmented coding. We just get to the last mile faster.

An example of a particular part that didn't go well: I built a really awesome TypeScript to Python fully typed SocketIO RPC library with code generated client/server using Zod schemas auto translated to Pydantic schemas (through JSON Schema). The code generation aspects were awesome to build with Claude, but getting the low level intricacies of the comms to work just perfectly took longer than it felt it should and was largely a manual process. It was awesome though and I plan to open source this soon (see my open source collection)!

In general, closing out features and flows still took the labor it took before. Crossing all the ts and dotting the lower case js. That's where vibe coding won't take you all the way, and it's still a good place to become an expert of your code that you and your AI assistant built. You own it and must support it after all.

There were more places, though, where it truly felt like a 10x multiplier. Those are, in no particular order:

• React landing pages and low-interactivity pages (even with Payload CMS backing). Making fantastic marketing landing pages is easy as can be.
• React components. Especially using component libraries like ShadCN to compose more complex components. The better the building blocks, the better the outcome.
• React flows with great DX tools like react-hook-form (with zod resolvers), TanStack libraries like react-query and tables, nuqs, etc. The better the DX of the tools, the better Claude is at working with it.
• Data Pipelines and Data Engineering flows. Platform integration and data transformations. It's never been simpler to have the perfect data integrity to drive your business intelligence. Knowing the tools and when they work for the use case (pandas, polars, duckdb, pyspark, etc.) makes you and your AI assistant a badass data engineer.
• To add, platform integrations are so stupid simple. API integration no longer takes too long to set up. Say the word and flash the api doc url and you're in business. Integrations take hours not weeks to get right. Do you use Jira issue numbers in your git branch names as a standard? Just whip up a script that takes a URL with the issue number in it, have it read the Jira API, and have it make a branch name based on the Jira issue and summary. Ez-pz.
• Did I mention DX? Used to be, we all had a list a mile long of manual process automations we'd get done "when we have time". Well now we have time. If you can automate it, you have time to get it done now. Did you set up a nice Payload CMS email templating system and need to transfer those auth email templates to Supabase for prod usage? Set up a script that uses Supabase's management API. Tired of your local Supabase dev env having generic auth emails? Extend that script to automatically copy them into your dev environment too. Minutes.
• Not done with DX. The better your tooling, the better your Claude will do. Lean into the era of Rust-based tools to speed up your type checking and linting. Biome, ruff in python. Excited to see ty come into beta to replace basedpyright. Check out things like e18e.dev and use the recommended replacements. The better and faster your monorepo tooling, the better the signals and faster your AI can make high quality code changes.
• IaC. The days of using the AWS or GCP console to whip up PoC is over. Use great IaC coding libraries like Pulumi to manage your infrastructure from your codebase from the get-go.
• Testing. Unit tests, integration tests, e2e tests. Gone are the excuses. Write the tests and ship with confidence.
• CI/CD. Github actions. ArgoCD. You name it. Build your workflows, acceptance tests, and deployments with the level of detail and automation we could only dream of before.

Bottom line. There's a lot to like. It's a paradigm shift. Not everyone is going to like it. It's another tool to put in your tool belt and frankly, this tool is gonna make it a lot easier to use all the other tools and make your tool belt big enough to meet the modern demand in software.

You're not getting replaced. You're just gonna be the best. Lean in.

Happy coding!