”Hey Claude. Execute the handoff.” Sure is nice starting a session like that. I watched Fable cook the handoff into code, and was then let down by the results. I realized that far too much had been left open to interpretation. I could see that existing code patterns and conventions I’d designed with got ignored. Most critically, the auto-tuning to bodyScale didn’t even fully work—only VFX graph params were getting auto-adjusted, but many other critical scale-dependent scene properties such as defaultUnscaledSize and gridScale were not. It all felt extremely vibe-coded (which it was) ((which is not what I wanted)).
I told the executing agent I had issues with the handoff and it asked me to elaborate on the gripes I had, so I started a second session and asked that new agent to audit the recent uncommitted changes in light of my desire to better adhere to the existing codebase (“What would the dev of this project have done had he coded it himself? Would he have taken an alternate route?”). I then fed the results of that audit back to the original agent as my feedback to its question about my gripes. It then stashed the half-baked changes it made in a new branch and fixed the handoff doc accordingly. I had that second agent take a look at the changes to the handoff just to make sure it all looked clear and in line, and it suggested some additional fixes. Back and forth I went. It kind of felt like using CodeRabbit. Now, my handoff is ready to go for another attempt tomorrow.
Tags: ai claude workflow refactoring vfx scaling vibe-coding