Setting Claude Code effort per prompt is what effortless automates: a Claude Code Mod that GitHub user HeyCubit published on October 6, 2026. A small judge model reads each prompt and picks the reasoning effort before it goes out, so quick questions run at Low and hard jobs at High without anyone touching /effort. One bar above the prompt also shows how full the context is and how long the prompt cache stays warm, with buttons to compact or hand off to a fresh chat.
Everything below is what the author reports in the project’s README and on its project page.
On this page
The problem it targets
Claude Code already lets you choose effort by hand. Anthropic’s model configuration docs list low, medium, high, xhigh and max for Opus 5.5, Sonnet 5.5 and Haiku 5.5, with medium as the default on those three models. You change it with /effort, the --effort flag, the CLAUDE_CODE_EFFORT_LEVEL environment variable or a per-model level in settings, and /effort auto clears your saved level for the active model.
Per the same docs, a level you confirm with Enter in /effort is saved as your default and applies in later sessions. The author’s pitch is that people forget to switch, leave everything on a high setting, and pay for reasoning a rename or a one-line question does not need. effortless makes the choice per prompt instead.
How the setup works
From the README. Install it like any plugin from a marketplace, inside Claude Code:
/plugin marketplace add HeyCubit/effortless
/plugin install effortless@effortless
Or from a terminal:
claude plugin marketplace add HeyCubit/effortless; claude plugin install effortless@effortless
Then run /reload-plugins or restart Claude Code. Anthropic’s plugin docs describe the same two steps: add the marketplace, then install a plugin from it. A short setup then appears above the prompt.
The judge. By default Haiku 5.5 reads each prompt and picks the effort, on your own Claude login and with no API key. Short follow-ups such as “go”, “ok” or “yes” keep the effort already picked and skip the judge. Optionally you can add Jev, a judge from TypeSafe that needs a TypeSafe key; the author says it answers in about 0.25 seconds against about 1 second for Haiku, and Haiku steps in when Jev is unsure or fails.
The bar. It shows the effort and model for the current prompt, how full the context is, a countdown to when the prompt cache goes cold, an Auto switch, and a Compact button that turns into Handoff when Haiku judges that a fresh chat would pay off. The bar also warns when a 5-hour or weekly limit passes 80%, with a Save mode that caps effort at Medium. Changing effort yourself turns Auto off at your next message.
Compaction. By default every compaction, /compact and the automatic one included, is summarized by Haiku 5.5; if Haiku fails, Claude Code compacts as usual.
Useful commands from the README:
| Command | What it does |
|---|---|
/effortless settings | opens the settings panel |
/effortless auto | turns automatic effort on or off |
/effortless handoff | hands off without the bar |
/effortless stats | what the prompts Auto steered cost this session, per effort |
/effortless bench | scores each judge on the labelled prompts in bench/judge-cases.json |
If the mod does not load in the terminal after a restart, the README says to add this to ~/.claude/settings.json:
{ "env": { "CLAUDE_CODE_ENABLE_FUNCTION_HOOKS": "1" } }
What the author measured
The README reports a run on October 10, 2026: four coding tasks in a fresh copy of the repository, each run three times with and without the mod, through Anthropic’s API. Every answer was checked, and all 12 runs were correct in each setup.
| Opus 5.5 | Change in cost per task with effortless on |
|---|---|
against fixed medium | 21% lower |
against fixed high | 29% lower |
against fixed xhigh | 42% lower |
On Sonnet the author reports no difference against medium, and savings of 10% against high and 35% against xhigh. The author attributes the saving to fewer requests per task at lower effort, and lists the limits: four tasks, three runs each, mostly easy work, and fresh chats through the API rather than long chats on a plan. On a separate set of 73 labelled prompts, Haiku picked an acceptable effort in 92 to 93% of cases and Jev in 99%, by the author’s own scoring.
What it does not do by default
- No model switching. An experimental setting can move a prompt the judge calls simple to a cheaper model, but it is off by default. The author found that each model keeps its own prompt cache, so a mid-chat switch first has to write the whole chat into the other model’s cache, which on an 80k-token chat cost about 5 times what staying on a warm Opus did.
- Paused on Fable. The README says a change of effort between requests rewrote most of the prompt cache on Fable 5.1 and older models, so Auto pauses there. Anthropic’s docs likewise note that changing effort while Claude is working can bring up a cache warning to confirm.
- Split view. In the desktop app’s split view, the bar only appears in the left pane.
Before you install
The README is direct about trust: like any Claude Code plugin, the mod runs code on your machine, and its code is in the repository’s hooks/ folder to read first. It starts claude (to update or uninstall itself), git (to check for new versions) and a file copy on update. A Jev key entered with /plugin configure goes to Claude Code’s secure storage; a key pasted into the setup guide is saved in plain text to ~/.config/jev/.env, so the author recommends /plugin configure on a shared machine. The project is MIT licensed and marked beta on its page.
For another community Mod that changes what each session loads, see harness-scope. For how plugin marketplaces work in general, see Claude skills marketplace.
