It’s getting more and more addictive! I’ve already launched 53 plans, and there are more than 40 ahead!
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 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!