JetBrains/benjamin-plus-skill

What it solves

It reduces the operational cost and token consumption of coding agents without sacrificing the quality of the work they produce. It targets "bloat" in agent workflows—such as redundant file lookups, inefficient environment probing, and excessive polling—which increases costs because agents pay for both the initial action and the repeated reading of the conversation history.

How it works

The project provides a set of optimized behavioral rules (a "skill") that are injected into an agent's system prompt or session start. These rules teach the agent five specific habits:

1. Combined Reconnaissance: Gathering all necessary facts in one step rather than multiple separate lookups.
2. Keyhole Reads: Reading only the necessary lines of a file (e.g., 50 lines) instead of the entire file when full content isn't needed for transformation.
3. Single-Pass Environment Probing: Checking and installing all missing dependencies in one command.
4. Strict Verification: Using the task's own verification command as the definitive signal to stop.
5. Efficient Polling: Checking the status of long-running builds every 30 seconds instead of every second.

Who it’s for

Developers and teams using coding agents (such as Claude Code or Codex CLI) who want to lower their API costs and token usage while maintaining the same solve rate.

Highlights

• Measured Savings: Demonstrated up to 18% reduction in cost and 22% reduction in tokens per task.
• Evidence-Based: Developed through an automated research loop analyzing 1,200 agent traces to keep only rules that improved efficiency without hurting quality.
• Cross-Platform: Compatible with various agents via system prompt injection or specific tool hooks (e.g., Claude Code, Codex CLI).
• Quality Preserved: A/B testing showed no significant change in the solve rate or quality of results.