Thoughts from using Claude Code for two and bit weeks at work. The expectation was to use CC to write all of our code for us.

The reality : probably about 50% of my code was written by AI . I wasn’t quite ready to give the robot full reigns over writing the feature end-to-end. why was this?

I was also building my mental model of how everything fitted together. My task for the last two weeks as extend a completed GQL resolver to support more sophistated graphs, and the existing logic is complex. I had’nt fully grokked the data flow end-to-end so although I could have prompted CC to build the feature, I would have struggled to understand the changes, I probably would have taken to validate its correctness . I would have had to unpick it anyways.

What I did instead. I used CC as very high-tech rubber ducker to talk me though the full flow of the resolvers, where data transformations were happening, why we were transforming the data here, where the data splits of . I asked it to produce me diagrams ascii art to check and recheck my understanding. This saved me hours easily. I have really good understanding our complex codebase now.

AI had successfully helped pin-point where changes needed to go, this was half the battle.

Next was tackling new buisness logic. Time for spec driven development? Not quite. but sort of.

Maybe im too close to the detail, maybe i’m a control freak? I knew how I was going to implement this. So should i have written all the steps down and in english explained the data structures precisiely? (maybe…)

The thinking , the time it would have take me to explain the data structure in english would have been quicker to write up the the zod / type interface in my IDE . this became the central talking point with CC and all my prompts was around this data structure.

These new graphs the calculations had to be tight . I knew debugging any data dicrepancies was going to be painful. I was working with big JSON data structures. but the the inputs and outputs were clear. This was a perfect scenario for TDD.

Getting CC to write the tests first. Sounds dangersous? Yes if you dont know what you want to test in the first place. I knew in my head the the critical paths. let see if CC would get them too but it would have been time consuming and labourious to type up. lets see if CC would get them… and yes it did! yay go AI and some other edges cases too which i didnt consider. scary.

strong opinions matter more than ever. The tests was not how I organised it. re prompt it. ask it to make it more try. organise these test into this test suites, these ones into that test suite, refactors are quick now! CC once again save me a tonne of time!

These tests give me a sane way to review my JSON transformations are working as expected. The buiness logic is complicated but with AI i have created my test harness to reason with my data better.

Time to write the buisness logic. CC can do that that in pinch too. point it at the test we wrote together. let it cook. voila! all tests pass, well done CC you’re very smart. Wait this function is unreadable, its not how i would have written it, dont do a while loop lets do a for of instead becasuse its easier for a human to understand this! remember refactors are cheap now. your taste and preference to coding is what matters more than ever now

things are coming together, tests are comprehensive. buisness logic implemented, bringing it altogether and testing e2e time. things dont work as expected

integration not working. assumptions about the data were not fully correct.

This was invetiably going to happen.

The messy part— humans getting together, why is this not working? it took us all day . but we figured out.

Tweak the sql on the data model. adjust the config in the DB. tweak schema slighlty in GQL. was i supposed to prompt all this? It was quicker to do it manually have and have that fine control at this point in the project. but yes AI was once again useful for the labourious parts , those test we painstakingly built all had to be updated AI was able to update them in one swathe. cool!

this is what software engineering is like a in product time that is trying new untested things.

We get things wrong, we iterate. we dont have all the answers right out the gate. AI helps yes. But can we use it for 100% of writing code. I personally cant yet. Could I possibly in future? I’m not sure, maybe it could be 80-90% . but maybe its because i think the code quality is still important, it needs that slight manually reallignment sometimes.