Using an AI Coding Agent With an Existing Codebase
AI coding is often presented as a way to build a new app from scratch.
But one of the more useful applications is working on code that already exists.
Most real projects aren't empty folders. They already have components, dependencies, configuration, APIs, tests, and plenty of code that you didn't write yourself.
That's where an AI coding agent can be useful.
Instead of asking an AI to generate a piece of code in isolation, you can give the agent a task and let it first understand how the project is structured.
A typical workflow might look like this:
Inspect the codebase → find the relevant files → understand the existing implementation → make the changes → run commands or tests → fix anything that breaks
This is different from simply asking a chatbot for a code snippet.
We've built Clixad around this kind of workflow.
Clixad is a cloud-based AI coding agent that runs in the terminal. It can read an existing project, modify files, execute commands, test changes, debug errors and continue iterating on the task.
For example, instead of asking:
"How do I add authentication to a React app?"
you can give an agent an existing project and ask it to add authentication to that project.
The agent can then inspect the relevant files and work with the code that's already there instead of giving you a generic implementation that you have to adapt manually.
This is particularly useful for:
adding features to an existing SaaS
fixing bugs across multiple files
refactoring old code
updating dependencies
writing or fixing tests
understanding an unfamiliar codebase
continuing a project that was started with AI
The cloud aspect also matters.
With Clixad, the AI models run remotely, so you don't need to run a coding model locally or provide your own API key. There is also no mandatory monthly subscription.
Users get free credits, with additional credits available through advertiser-funded offers, forms and surveys.
We've also made the five cheapest models free within their daily limits, which gives users a way to work on smaller tasks without spending credits while they're within those limits.
For new projects, AI can help generate the first version.
For existing projects, the more interesting question is whether an agent can understand what is already there and safely work within it.
That's where we think agentic coding becomes much more useful than simple code generation.
The step that quietly breaks this on a real repo is the first one. "Inspect the codebase" is the most expensive thing the agent does, and it spends the whole budget before writing a single line.
Concretely: it greps, reads six files to work out the pattern, reads the tests, reads the config. That's most of the window gone. Then it starts editing, compaction fires, and the understanding it just paid for gets summarised down to two sentences. From there it's editing a codebase it no longer actually has in front of it. The tell is that it starts reintroducing a pattern the repo already rejected, or re-reads a file it read twenty minutes earlier and draws a different conclusion from it.
Two things that helped me more than a better prompt:
Make the inspect step produce a durable artifact, not context. Have the agent write down what it learned in a file in the repo: the conventions, the three files that actually matter, what it must not touch. Then have it work from that file. The file survives compaction. Context doesn't. It also means the second task on that repo starts warm.
Scope by blast radius, not by how big the task sounds. "Add a field to this model" sounds tiny and touches nine files. "Rewrite this component" sounds big and touches one. The first one is where agents go wrong on existing code.
One thing worth checking before tuning any of this: the context window you actually get is set by whoever serves the model, not by the model card. Most hosted endpoints serve around 32K regardless of the advertised number, and it fails silently, no error, just truncation. On an empty folder you never notice. On an existing codebase it's the whole game.
Disclosure: I build Grunz, a chat and coding agent on open weights, so I've hit all of this the hard way.
Interesting. How are you measuring whether it is working?
Interesting. How are you measuring whether it is working?
Interesting take. Would you still recommend this approach to someone starting today?
Appreciate the honesty here, most people only share the wins.
Helpful post. How did you get your first bit of traction?
Clear and practical, thanks. Did anything surprise you along the way?
Curious how long it took before you saw the first real results?
Nice work shipping it. What has been the biggest challenge since launch?
Interesting take. Would you still recommend this approach to someone starting today?
Helpful post. How did you get your first bit of traction?
Appreciate the honesty here, most people only share the wins.
Clear and practical, thanks. Did anything surprise you along the way?
Curious how long it took before you saw the first real results?
Nice work shipping it. What has been the biggest challenge since launch?
Interesting. How are you measuring whether it is working?
This resonates a lot — how long did it take before you saw any real signal on it?
Interesting take. Would you still recommend this approach to someone starting today?
Appreciate the honesty here, most people only share the wins.
Clear and practical, thanks. Did anything surprise you along the way?
Helpful post. How did you get your first bit of traction?
Curious how long it took before you saw the first real results?
Nice work shipping it. What has been the biggest challenge since launch?
Curious how long it took before you saw the first real results?
Nice work shipping it. What has been the biggest challenge since launch?