- A delay of more than one month without justification = immediate suspension of service.

- If the service was suspended due to non-payment, next collaboration only continues with full prepayment for the next month.

- Transfer of intellectual property and copyrights only after final payment.

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!