How I Use AI to Build a 55-Crate Rust Project (Honestly)

I build Temps alone. The codebase is 55 Rust crates. That's a lot of surface area for one person. So yes, I use AI assistance heavily. Claude (Anthropic's AI assistant), mostly.

I want to be specific about how I use it, because the honest answer is more nuanced than "AI writes my code" or "I write everything myself."

TL;DR: AI makes me roughly 30% faster overall. The gains are concentrated in boilerplate, tests, and documentation. Architecture decisions, race conditions, and subtle lifetime issues are the same speed as before. The biggest workflow shift: I stopped writing first drafts of boilerplate myself, which freed my attention for the hard problems.

What AI Is Actually Good At

Boilerplate. Rust has a lot of it. Implementing a trait for a new type means writing the same three methods with the same signature patterns I've written fifty times. I'll describe what I need and let Claude draft the implementation. I read every line before it goes in, but the drafting is fast.

Tests. Writing unit tests is often more tedious than writing the code being tested. AI drafts test cases quickly. I review them for completeness, add the edge cases it missed (there are always edge cases it missed), and move on.

Documentation. Rust has rustdoc, and public APIs need doc comments. Describing what a function does in a comment is the kind of task where AI is faster than I am and the output is usually as good or better.

Debugging known error classes. If I have a borrow checker error I've seen before, I'll paste it in and let Claude explain what's happening. This is faster than re-deriving the explanation from first principles every time.

Translation between representations. I have a Postgres schema and I want the corresponding Sea-ORM entity structs. That's mechanical translation. AI handles it in seconds. I verify the output and check the types.

What AI Is Not Good At

Architecture decisions. I've tried asking Claude to help me decide how to structure a new subsystem and the answers are plausible but shallow. They don't account for the specific constraints of the codebase's design, the things I know from having built it over two years that aren't captured in any single context window.

Debugging race conditions. When something is timing-dependent, when a test fails 1 in 20 runs, when a deployment works on the third retry but not the first, AI suggestions are often confident and wrong. These bugs require understanding the system's actual runtime behavior, not its code structure.

Knowing when to say no. AI will implement almost any feature I describe if I frame the request confidently. It rarely pushes back with "this is the wrong approach" or "this will create a problem in three months." That skepticism is something I have to supply myself.

Subtle Rust lifetime issues. Straightforward borrow checker errors are fine. But the ones that involve async lifetimes, shared mutable state across task boundaries, or complex interaction between the proxy and the deployment engine — those require careful human reasoning. AI gets them wrong often enough that I've learned to treat its suggestions as a starting point, not an answer.

How I've Changed How I Work

The biggest shift is that I've stopped writing first drafts of boilerplate myself. I used to start every new module by typing from scratch. Now I describe what I want in a comment and let Claude draft the structure, then I revise from there.

This sounds minor, but it changed my relationship with tedious tasks. The parts of coding I used to procrastinate because they were boring are now the fastest parts. The cognitive overhead of starting a new file went to nearly zero.

What it means is that my attention is available for the hard parts. Designing the state machine for a new deployment feature. Thinking through the failure modes. Deciding what to leave out. That's where I spend my time now.

The Parts I Want to Get Better

Prompt engineering for Rust codebases is still somewhat rough. I've gotten better at providing context (I include relevant type signatures, trait bounds, and architectural constraints in my prompts) but there's still a lot of back-and-forth on anything non-trivial.

I'd like better integration with the actual codebase — asking questions about my specific code rather than generic Rust patterns. That's getting better. But it's not there yet.

Honest Assessment

AI made me probably 30% faster overall. Not 10x, not 2x. 30%. The gains are concentrated in the boring parts, which means the hard parts (the architecture, the edge cases, the debugging) are still slow. But they're the same speed they would have been without AI. The tedium tax came down.

For a solo developer building something this large, 30% faster is the difference between shipping and not shipping.

If you're curious about the technical decisions that shaped the codebase, see why we chose Rust for a deployment platform. For the practical result of all this work, here's how git-push deployments work under the hood.