Where the money goes
The same job costs a lot more tokens (AI usage) with /pstack:poteto-mode. The extra goes to three places:
- Helper AIs. Reading and writing code goes to several subagents, and each one costs separately.
- Review panels.
arena,architect, andinterrogatehave several models each do a pass. The default is two, so two costs. - Thinking longer. The default level is
xhigh. The longer it thinks, the more it costs.
That money buys checks and second opinions. The moves below spend less without losing them.
Skip /pstack:poteto-mode for small things
A typo, a color tweak, one more log line: if you can see at a glance whether it’s right, just ask.
- Only want to know how: use
/pstack:poteto-help. It answers and doesn’t start work. - Don’t understand the code yet: ask
/pstack:howfirst. It’s cheaper than redoing a wrong change.
Strong main model, cheap models for code
Run the main conversation, the one you talk to, on a strong model. Switch the code-writing roles to cheaper models.The main conversation has to understand you and review changes, so skimping there causes mistakes. Coding helpers get clearly scoped work, which cheaper models handle well.
Run /pstack:setup-pstack. At the last step, “confirm roles,” change the feature, refactoring, bug-fix, and perf-issue lines to a cheaper model.
feature, refactoringCheaper modelbug-fixCheaper modelperf-issueCheaper model
The helpers get clearly defined work, which cheaper models handle well
Drop a budget level
It offers four budgets:
| Budget | Level | Notes |
|---|---|---|
unlimited |
Up to max |
Most expensive |
large |
xhigh |
The default |
medium |
high |
A bit cheaper |
small |
medium |
Cheapest |
The budget only changes how long models think, not which models. Roles you changed yourself are kept. Start a new conversation after you pick, or it won’t apply.
Send one fewer reviewer
The arena runners, architect runners, and interrogate reviewers lines send one helper per model listed. The default is two:
arena runners: claude-opus-5-5-xhigh, grok-4.7-xhigh-fast
interrogate reviewers: claude-opus-5-5-xhigh, grok-4.7-xhigh-fastDelete one and you pay for one less, but you also lose one model’s view. You can also say it per task:
/pstack:arena this naming question isn't a big deal. 2 options is enough.auto and inherit-parent: follow the main conversation
Putting auto or inherit-parent on a role means the same thing: no model is set, so it uses the main conversation’s model. auto isn’t a model called auto.
When the main conversation is cheap, these roles are cheap too. In Claude Code, the coding roles use it by default.
Stuck?
After switching to small, results got noticeably worse.
Try medium first. Or keep large and only switch the coding roles to a cheaper model. A bad judgment call usually costs more than the tokens you save.
I edited a model name in the settings file by hand, and it says the model isn't available.
The name may be misspelled, or you may not have access. It falls back to the default model and says so. To change models, rerun /pstack:setup-pstack and pick from its list.
I changed the budget, but the cost didn't change.
New settings only apply to new conversations. Start a new one and check again. On tools like Codex, make sure ~/.agents/pstack-models.md was actually written.