HeyCubit/effortless

effortless Picks the reasoning effort for every prompt. On real coding tasks, 21% cheaper than Opus on medium, with the same answers. Website · Install

textless While Claude works, one pixel-art picture stands in for the wall of steps. Pick a set, use your own art, or find one in the Marketplace. Website · Install

Install · What it shows · Judges

A Claude Code mod that picks the reasoning effort for every prompt. Quick questions get Low, hard jobs get High, and you never touch the Effort control. On real coding tasks it cut the cost of Opus by 21% against Anthropic's recommended medium, 29% against high and 42% against xhigh, with every answer still right (the benchmark). One bar above the prompt also shows how full the chat is, how long the prompt cache stays warm, and hands off or compacts in one click when a chat gets heavy.

Haiku 5.5 costs about 75% less than Haiku 4.5, and Anthropic names compaction and quick, well-defined work among what it is built for. effortless uses it in three places:

The cheaper model is switched on in Settings → Model; compaction is switched in Settings → Handoff → Compact with.

Two ways. Both take a minute and need no terminal.

1. Ask Claude (easiest). Paste this into Claude Code and press Enter:

Install the effortless plugin for me: run `claude plugin marketplace add HeyCubit/effortless` and then `claude plugin install effortless@effortless`. When both succeed, tell me to run /reload-plugins.

2. Commands. Type these two lines in Claude Code's chat box:

/plugin marketplace add HeyCubit/effortless

/plugin install effortless@effortless

Or in a terminal, in one line:

claude plugin marketplace add HeyCubit/effortless; claude plugin install effortless@effortless

Then run /reload-plugins (or restart Claude Code). A short setup opens above the prompt: keep Haiku alone (one click, no key, runs on your own Claude login) or add Jev (a TypeSafe key, about 4x faster effort calls), lean cheaper or smarter, and pick how handoffs are written. Run /effortless setup to go through it again, or change any of it in ⚙.

Updates come to you: when a new version is out, a card above the prompt offers Update or Later. Update loads the new version in the chat you pressed it in, with nothing to type.

Prefer it quiet? Settings → Customize has a Minimal look with no bar. It also sets the Appearance (match Claude Code, dark or light) and the Theme (violet, Claude orange or cherry blossom).

When the cache has gone cold on a big chat, the bar turns to ice with Compact and Handoff:

Handoff writes a summary of the chat and carries on in a clean one. Quick takes a few seconds; Full checks git and saves HANDOFF.md (or runs your own skill). Then clear and carry on, clear and wait, or keep the chat and copy.

Compact takes an optional note for what the summary should keep. Every compaction, also /compact and the automatic one, is written by Haiku 5.5 by default, which Anthropic recommends for compaction at a fraction of the chat model's price; if Haiku fails, Claude Code compacts as usual. Switch it in Settings → Handoff → Compact with.

The bar also warns when a chat is getting swamped (each message re-reads a lot of context), when a 5-hour or weekly limit passes 80% (with a Save mode that caps effort at Medium), and when your judge stops answering.

Haiku 5.5 runs on your own Claude login and needs no key. It always makes the handoff call: from 30% of context, every second message, it reads what the chat was for, the trail of topics, the last reply and how full the context is, and says a fresh chat would suit only for a clear reason. Then the Compact button turns into a lit Handoff with the reason. Want Handoff on the bar all the time? Settings → Handoff → Handoff button → Always.

Haiku also picks the effort, unless you add Jev, TypeSafe's faster judge (about 0.25 s against about 1 s). With a key, Jev answers the effort first and Haiku steps in whenever Jev is unsure. Run /plugin configure effortless@effortless in Claude Code, or use the setup guide or the Judge card in Settings.

A key set with /plugin configure is kept in Claude Code's secure storage. A key pasted in the setup guide or the Judge card is written in plain text to ~/.config/jev/.env (the file the jev skills read), never to a file of this repo. If Jev fails or takes longer than 3 seconds, Haiku judges that prompt and effortless tells you why once per session (out of credits, key rejected, no answer).

Short follow-ups such as "go", "ok" or "yes" keep the effort already picked and ask no judge.

• The default judge needs no key: Haiku runs on your own Claude login.
• A key is only needed to add Jev. Enter it with /plugin configure, never on the command line, so it stays out of your shell history: Claude Code then keeps it in its secure storage. A key pasted in the setup guide or the Judge card is saved in plain text to ~/.config/jev/.env instead, so prefer /plugin configure on a shared machine.
• A key is sent only to TypeSafe, and to nothing else. Use a key with a spending limit if your provider offers one.
• Like any Claude Code plugin, this mod runs code on your machine. Its code is in hooks/: read it before you install if you do not know the author. The programs it starts are claude itself (to update or uninstall the mod when you press those buttons), git (to see if a new version is out) and, on Update, a plain file copy of the new version into the folder your open chat runs from.

Four real coding tasks in a fresh copy of this repo, each run 3 times with effortless on and with it off, through Anthropic's API so the cost is real. Every answer was checked. Run 2026-10-10, 96 runs:

That is 21% less than medium, 29% less than high and 42% less than xhigh. On Sonnet it came out even against medium (Sonnet at medium is already lean) and saved 10% against high and 35% against xhigh.

The saving comes from fewer steps: every request rereads the whole chat, and at a lower effort the model makes fewer of them. Over 80k requests of real Claude Code use, about 76% of the cost was that reread, 16% cache writes and 8% output. Keeping chats short and compacting before the cache goes cold saves on the same 76%; that is what the countdown is for.

Limits: four tasks, three runs each, mostly easy work where effortless goes low. On hard work it picks more effort and saves less. Fresh chats through the API, not long chats on a plan. Run it again with bash bench/agentic/run.sh; all numbers and limits are in bench/RESULTS.md.

Moving a simple prompt to Haiku looks cheap on a single prompt: the cheaper answers were good enough in 27 of 30 and cost 60 to 77% less. But each model keeps its own prompt cache, so a prompt moved mid-chat has to write the whole chat into the other model's cache first. Measured on an 80k-token chat, that costs about 5 times what staying on a warm Opus does (10 times for Sonnet). So Cheaper model when it can is off by default and marked experimental. Details.

/effortless bench runs labelled prompts through each judge you have set up, using the same code a real prompt goes through, and scores them against a fixed effort. Each case lists the efforts a careful person would accept for that message on that model. The cases are in bench/judge-cases.json; run it yourself.

Two runs on version 1.5.1, 73 cases, Sonnet and Opus 5.5:

What this does and does not show:

• It measures whether the judge picks a sensible effort, not how much a session costs or how good the answers are.
• 73 cases is a small set, and the judges are not fully deterministic: runs differ by a few points.
• The judge prompts were tuned on 53 of the cases. The 20 held-out cases were never tuned against, so that column is the honest one. They were written after the first run showed which kinds of message miss, so they are not blind.
• Haiku's remaining misses are mostly "yes" or "thanks" after hard work, where it keeps the effort high.

Effort only changes on Opus 5.5 and Sonnet 5.5. On Fable 5.1 and older models a change of effort between requests rewrote most of the prompt cache, which costs more than it saves, so Auto pauses there and the footer says Paused. Haiku takes no effort setting.

In the desktop app's split view, Claude Code draws plugin bars only in the left pane. The right pane still draws replies, so when a chat went cold, is getting swamped or runs hot, a small card in the bar's colours hangs under the newest reply, naming the command that does what the bar's button would: /compact, /effortless handoff or /effortless save.

effortless draws in the terminal CLI too, with moving pixel art in its bands. Claude Code may not load a plugin's code there yet: if /effortless says the mod is not loaded after a restart, add this to ~/.claude/settings.json and restart:

{ "env": { "CLAUDE_CODE_ENABLE_FUNCTION_HOOKS": "1" } }

Tested on Claude Code 2.1.285 (stable), 2.1.286 and 2.1.288 at 80 and 120 columns. Below 90 columns the bands leave their art out.

While Claude works, Claude Code shows every step it takes. textless puts one picture there instead, and when Claude is done you get the answer, with a pixel word that says whether it is done, needs you or stopped.

• Pick what shows: a set (Clawd, Scenery, Paintings, Abstract, Working), one favourite scene, or your own art in ~/.claude/textless/scenes.
• The Marketplace: scenes other people made, checked before they go live. One command puts one on: /textless get <id>.
• One switch per chat: the ⏻ next to the model name, or /textless.

Website, scenes and Marketplace: heycubit.github.io/effortless/textless. The code is in textless/.

effortless is MIT licensed. Like any Claude Code plugin it runs code on your machine, so the whole mod is here to read: hooks/. It sends nothing anywhere except your prompt to the judge you picked.

claude plugin validate .
claude plugin test .
claude --plugin-dir .

tools/render-band draws the bar the way the desktop app does, without opening it; the images in this README come from it.

MIT