Install it
In the Claude Code chat box, type these two lines one after the other:
/plugin marketplace add cursor/plugins
/plugin install pstack@cursor-pluginsThis installs the full set, so /pstack:poteto-mode has everything it needs. If you’re not sure, use this.
To pick skills yourself, type this in your terminal (not the chat box):
Check the skills you want and choose Claude Code as the tool. The skills go into .claude/skills/. For what to pick, see Install only a few skills.
Add pstack: before commands
If you installed from the marketplace, every command lives under the plugin name. Plain /poteto-mode isn’t recognized; you type:
The same goes for other commands, like /pstack:how. That’s Claude Code’s rule for plugins, so two plugins can’t clash over a name. If you used npx skills, there’s no prefix.
Commands on this site are written for the plugin by default, so copying works. If you installed with npx skills, pick “Installed with npx” here:
I use
Commands on this site are shown for this tool, so copying just works
Shown as: /pstack:poteto-mode
Pick models, then start a new session
After installing, run this once:
It asks about your reasoning budget (how long models think), then assigns a model to each model roles. Finally it writes one file: ~/.claude/rules/pstack-models.md.
Claude Code reads the .md files in that folder each time a new session starts. So after setup, quit and open a new session before you start working. The current session still has the old settings.
A model is written as “name + level,” like opus xhigh. The only names are opus, sonnet, haiku, and fable. Keep these lines in this one file. Don’t copy them into CLAUDE.md.
How to set it up well
Most people should pick Balanced: the session runs on opus, and sonnet writes the code.
Why? Two things decide which models get used:
- The session model: the one you pick with
/model, the AI you talk to directly. It picks the approach, hands out work, and reviews every change. - Each line in the settings file: which model each helper AI (subagent) runs on.
By default, pstack has grok write the code. Claude Code can’t use grok, so those lines become inherit-parent. That means “follow the session model.”
So if you change nothing and run the session on opus, opus does all the work. It costs the most, and both review seats are the same model. Balanced moves the coding to sonnet. That halves the price and puts two different models on reviews.
Prices are API list prices as of October 2026, per million tokens. On a subscription, pricier models use up your limits faster:
| Model | Input / output | Good for |
|---|---|---|
| fable | $10 / $50 | Strongest, when cost is no object |
| opus | $4 / $20 | Session, judgment, hardest changes |
| sonnet | $2 / $10 | Writing code, second review seat |
| haiku | $0.1 / $0.5 | Not recommended |
Why not haiku: checking the results uses the same model as the coding line. Make that the cheapest model, and the checks get weaker too.
Three plans
Run the session on opus, and change every inherit-parent in the settings file to sonnet xhigh.
The session has to understand you and review changes, so it gets opus. Coding helpers get clearly scoped work, which sonnet can handle. The hardest changes go through hardest tasks, still on opus.
# pstack model configuration. One line per role. Delete a line to fall back to the skill default.
# budget: large (xhigh)
feature, refactoring: sonnet xhigh
bug-fix: sonnet xhigh
perf-issue: sonnet xhigh
hillclimb: sonnet xhigh
judgment and prose: opus xhigh
hardest tasks: opus xhigh
how explorer: sonnet xhigh
how explainer: opus xhigh
why investigators: sonnet xhigh
why synthesizer: opus xhigh
reflect tooling: sonnet xhigh
reflect judgment, divergent, synthesizer: opus xhigh
arena runners: opus xhigh, sonnet xhigh
arena cross-judge pool: opus xhigh, sonnet xhigh
swarm workers: sonnet xhigh
architect runners: opus xhigh, sonnet xhigh
interrogate reviewers: opus xhigh, sonnet xhighRun the session on sonnet, keep the default settings file, and pick the medium budget.
You don’t change a line. inherit-parent follows the session to sonnet, and opus still makes the judgment calls. Medium lowers every level from xhigh to high. For complex work, switch back to opus with /model before you start.
# pstack model configuration. One line per role. Delete a line to fall back to the skill default.
# budget: medium (high)
feature, refactoring: inherit-parent
bug-fix: inherit-parent
perf-issue: inherit-parent
hillclimb: inherit-parent
judgment and prose: opus high
hardest tasks: opus high
how explorer: inherit-parent
how explainer: opus high
why investigators: inherit-parent
why synthesizer: opus high
reflect tooling: inherit-parent
reflect judgment, divergent, synthesizer: opus high
arena runners: opus high, inherit-parent
arena cross-judge pool: opus high, inherit-parent
swarm workers: inherit-parent
architect runners: opus high, inherit-parent
interrogate reviewers: opus high, inherit-parentRun the session on opus, give judgment to fable, the code to opus, and seat three models on each review.
fable goes where judgment matters most: explaining, summarizing, the hardest changes. A review lists three models, so it sends three helpers. fable’s list price is 2.5 times opus. If your account can’t use fable, setup won’t offer it.
# pstack model configuration. One line per role. Delete a line to fall back to the skill default.
# budget: large (xhigh)
feature, refactoring: opus xhigh
bug-fix: opus xhigh
perf-issue: opus xhigh
hillclimb: opus xhigh
judgment and prose: fable xhigh
hardest tasks: fable xhigh
how explorer: sonnet xhigh
how explainer: fable xhigh
why investigators: opus xhigh
why synthesizer: fable xhigh
reflect tooling: opus xhigh
reflect judgment, divergent, synthesizer: fable xhigh
arena runners: fable xhigh, opus xhigh, sonnet xhigh
arena cross-judge pool: fable xhigh, opus xhigh
swarm workers: opus xhigh
architect runners: fable xhigh, opus xhigh, sonnet xhigh
interrogate reviewers: fable xhigh, opus xhigh, sonnet xhighNone of the plans use the unlimited budget. pstack already raises the level one step for the hardest step on its own.
How to write it in
- Pick the session model with
/model. - Run
/pstack:setup-pstackand pick the budget for your plan. - When it asks whether to change any roles, paste it the file above and have it write that.
- Quit and start a new session.
You see “Skipped in Claude Code”
Cursor’s scheduled tasks, its Bugbot review bot, and its cloud agents don’t exist in Claude Code. Steps that use them get skipped. The reply gets one line, and it moves on:
Skipped in Claude Code: <which step>. <why it was skipped>.Opening a PR still works, using gh pr create instead. /loop exists here too. If it says it left a placeholder (like a missing link), check what’s missing and fill it in if needed.
Turn on auto-update
Claude Code doesn’t update third-party plugins by default. To turn on auto-update:
- Type
/pluginin the chat box. - Open the marketplaces tab.
- Select
cursor-plugins. - Choose enable auto-update.
To update by hand, run this in your terminal:
If an update changes the default models, your old settings file keeps the old ones. For the fix, see Update, inspect, uninstall.
Uninstall or pause
Type this in the chat box:
Then type /reload-plugins, or restart Claude Code. To pause it for a while, use disable in place of uninstall, and enable to turn it back on. If you ran setup, also delete ~/.claude/rules/pstack-models.md after uninstalling, or every new session keeps reading it.
Stuck?
I typed a command and nothing happened. It's not in the menu either.
If you installed from the marketplace, type /pstack:poteto-mode; without pstack: it isn’t recognized. Type /pstack: first, and the menu lists only pstack’s commands.
I ran setup, but the models I chose didn't take effect.
The settings file is only read when a new session starts. Quit and open a new one. If that doesn’t help, open ~/.claude/rules/pstack-models.md and check the line shows the model you chose.
A reply says /pstack:arena skipped its judge, or found no reviewer from another family.
A few steps want a model from “another family” to review, like the /pstack:arena judge. All Claude models count as one family. So it skips the judge, picks itself, and says so. The arena cross-judge pool line does almost nothing here. For another vendor’s eyes, take the review that matters to Codex or Cursor and run /pstack:interrogate there.
The skills use Cursor's names. Do I translate them?
No. The plugin includes a mapping guide (docs/harness/claude-code.md), and the AI translates on its own. For example, .cursor/rules/ becomes CLAUDE.md (or AGENTS.md if there isn’t one), and cloud agents become a local worktree.
The AI says the Skill tool can't call pstack's skills. Is it broken?
No. In Claude Code, most pstack skills can’t be called that way. The AI reads that skill’s SKILL.md directly and follows it. Keep typing commands as usual.
I let it run overnight and found it stuck on a permission prompt.
Before you let it run alone, allow the common commands: git, gh, your test and build commands. Claude Code’s built-in /fewer-permission-prompts finds the look-only commands you use often and writes them into the project’s .claude/settings.json. Whether to allow commands that change things is your call.