It has already consumed so many of my tokens that it would be enough to rebuild a small town.
My personal recommendations:
- Enabling parallel Codex review in the settings significantly improves quality, but increases Claude token usage by about 30%, because they often argue about style rather than substance.
- I enabled the use of maximum-capability models. This significantly increased costs, but also improved quality. - From the very start of a project, make sure to give an instruction to cover the entire codebase with all types of tests (from units to e2e tests).
- Plans can be executed in one “batch” sequentially if they do not conflict with each other—just ask Claude to write a script.
- After each plan, make sure to run all tests and migrtions.
- After each batch of plans, do a deployment and manual visual testing. Otherwise, something that breaks once will cause further issues down the line.
- One-way Telegram integration is useless—knowing that something ran or failed does not really help.
- It adds Claude as a co-author in Git. This is not a problem for me, but it might be worth disabling for others.
- With each next plan, the cost keeps increasing because the solution becomes more complex.
- Splitting the system into microservices is the only way to handle the multiplied growth in complexity and cost.
- It can write not only code, but also texts and templates. I just haven’t tried images yet.
It’s impossible to stop!
