← sources

I built checkmate: a Claude Code plugin that makes the agent take the shortest path to the goal. Pass rate went 86.7% → 98.3% in my benchmark, output tokens halved

source: reddit (reddit:ClaudeCode)published 2026-10-24 08:19ingested 2026-10-24 11:00

retained text

Left alone, coding agents add preamble, option menus, extra files, and a "want me to also...?" at the end. checkmate is a goal-driven mode that fixes that. It has 9 rules: identify the win condition, take the least complex path that fully works, answer first, decide instead of menu, finish, verify against the goal, close with done / not done / blocked.

What I learned building it: a SKILL.md alone doesn't reliably activate. It only worked once a SessionStart hook injected the rules, so the plugin ships with one. Modes are lite, full, ultra and off. Brevity never overrides correctness, and if you ask for detail you get detail.

Benchmark (40 tasks × 3 runs, feature / bugfix / refactor / QA, automated pass/fail, medians among passed runs only):

Pass rate

• baseline: 86.7%

• checkmate: 98.3%

• ultra: 95.8%

Median output tokens

• baseline: 562

• checkmate: 274

• ultra: 203

Median steps

• baseline: 4

• checkmate: 2

• ultra: 2

Median wall time

• baseline: 6.5s

• checkmate: 3.8s

• ultra: 3.1s

Total tokens only drop about 20% (53.8k → 42.9k), since context dominates. The big win is output and steps.

Caveats: I ran this on one model (claude-opus-5.5), so treat it as a first result and not a general claim. Run it yourself with npm test && npm run build, and the harness is in /benchmarks.

Works with Claude Code, plus generated rules for Cursor, OpenCode and Codex. Zero dependencies, MIT. Inspired by ponytail, which targets code volume; this targets moves to the goal.

https://github.com/rover-001/checkmate

I'd like to hear where it's too terse or skips something it shouldn't.

document fetched 2026-10-25 00:35 · canonical https://reddit.com/r/ClaudeCode/comments/1xe34m5/i_built_checkmate_a_claude_code_plugin_that_makes