What is it?
What it is
A token‑efficiency skill for coding agents that teaches five habits to reduce cost and token usage without affecting quality.
Why it exists
It exists to eliminate wasted tokens and cost from redundant lookups, polling, and inefficient steps, as demonstrated by paired A/B measurements.
Who should use it
Developers and teams using AI coding agents (e.g., Claude Code, Codex CLI) who aim to reduce token costs and improve operational efficiency without sacrificing code quality.
Who should avoid it
Users not employing AI coding agents, or those with minimal task complexity where the potential savings do not justify the integration effort.
How it works
A quick walkthrough in plain English
How benjamin-plus-skill works
Step 1 of 3
You interact with it
Open benjamin-plus-skill, send a request, or connect it to your stack.
Features
Advantages
- Up to -18% cost and -22% tokens per task with quality unchanged across SkillsBench
- Measured solve rate unchanged (p=0.22) on Java SWE-bench (Codex CLI, gpt-5.6-luna, 675 paired replicas)
- Injection via hook or AGENTS.md saves -17.9% cost median on Claude Code and -4.4% on Java/Codex
- Quality statistically indistinguishable from baseline: 7 better / 5 worse / 68 ties (sign p=0.77)
- Verifier reward improved from 0.362 to 0.392
- Polling rule alone can cut nearly half of agent steps on some platforms
- Simple integration — a single cat command and one ~3 KB file
- Open methodology with paired A/B and per-trial adoption check (80/80 vs 0/80)
- MIT license enables broad reuse
- Maintained by JetBrains with iteration based on real benchmark evidence
Disadvantages
- Not powered as an equivalence test — large effects ruled out but small quality differences not detectable
- Savings depend on baseline bloat: agents with lean sessions may see much smaller gains
- Hard-task tails can erode aggregate savings on a few outlier tasks
- As a discoverable skill folder (vs injected) it saves nothing (-0.5%, n.s.); requires explicit hook/AGENTS.md wiring
- Some rules require platform-specific polling cadence assumptions that may not generalize
- Limited to the agent platforms tested (Claude Code, Codex CLI); untested harnesses may behave differently
- Shell-based delivery assumes user comfort with editing settings.json or appending files
- Feedback loop relies on users supplying paired before/after numbers and traces for regressions
- Only 291 GitHub stars, suggesting limited community validation beyond the maintainer's benchmarks
- No topics/tags metadata to aid discoverability in agent skill registries
Installation
native
git clone https://github.com/JetBrains/benjamin-plus-skill ~/.benjamin-plus
For Claude Code, add to ~/.claude/settings.json:
{ "hooks": { "SessionStart": [ { "matcher": "startup|resume|clear|compact", "hooks": [ { "type": "command", "command": "cat ~/.benjamin-plus/injected-instruction.md" } ] } ] } }
Or per-project: cat ~/.benjamin-plus/injected-instruction.md >> CLAUDE.md
For Codex CLI: cat ~/.benjamin-plus/injected-instruction.md >> ~/.codex/AGENTS.md (or >> AGENTS.md in a repo)
For any other agent: append ~/.benjamin-plus/injected-instruction.md to the system prompt.FAQ
What exactly does the benjamin-plus skill do?
It is a token-efficiency skill for coding agents that optimizes how they perform lookups and polling. It teaches five specific habits—such as 'Recon in one pass' and 'Keyhole reads'—to reduce cost and token consumption without sacrificing task quality.
How much can I expect to save in terms of cost and tokens?
Users can expect a median cost reduction of approximately -10% to -18% and a reduction of up to -22% in tokens per task, depending on the baseline bloat of your sessions.
Does using this skill affect the quality of the agent's output?
No. Based on measured benchmarks, the quality remains unchanged. In testing, the mean verifier reward actually showed a slight increase (0.362 to 0.392), and solve rates remained stable.
Should I install this as a discoverable skill folder or inject it?
You must inject it. Testing showed that making it a discoverable skill folder saves nothing (-0.5%), as agents waste steps trying to find the documentation. For best results, append the instruction file directly to your agent's system prompt or configuration.
How do I integrate this with Claude Code?
You can add it to your `~/.claude/settings.json` file under the 'hooks' section to trigger a command that cats the instruction file during session startup, or simply append the content of `injected-instruction.md` to your project's `CLAUDE.md` file.
How was this skill developed and validated?
The skill was developed through auto-research, involving an agent analyzing 1,200 traces to identify inefficiencies. It was then validated through a rigorous paired A/B test using 80 SkillsBench tasks to ensure savings were achieved without degrading performance.
Featured in Videos
YouTube tutorials and walkthroughs for benjamin-plus-skill
Alternatives
Similar projects ranked by category, topics, and text overlap.
