GrugDone

AI think less. Work still done.

A tiny Agent Skill designed to reduce unnecessary reasoning verbosity while keeping final answers clear and complete.

What it does

Long reasoning traces can restate the prompt, narrate obvious steps, or polish thoughts that never need to be prose. GrugDone asks agents to keep that trace compact while preserving exact terms, constraints, and corrections.

Concise-output prompts shorten what the user sees. GrugDone targets unnecessary reasoning verbosity instead.

How it works

The skill favors short fragments and direct movement from observation to decision to action. It changes the reasoning style requested from the agent; actual behavior still depends on the model and environment.

Some harnesses expose reasoning text that can be measured. GrugDone makes no claim about unobserved internal computation.

Benchmarks

Captured reasoning traces

Two clean pairs selected from the same 20-prompt GLM-5.2 high-reasoning run. The cells contain only reasoning text exposed by the benchmark harness. Counts are words because usable token counts were not returned.

Without GrugDone — 41 words	With GrugDone — 8 words
The user asks a SQL conceptual question. No tool use needed. Answer concisely. The issue: WHERE clause on the right table's column turns the LEFT JOIN into effectively an INNER JOIN, filtering out NULL rows. Fix: move predicate to ON clause.	Classic SQL LEFT JOIN + WHERE filter issue.
Without GrugDone — 89 words	With GrugDone — 50 words
This is a simple probability question. Let me just answer it directly. Total balls: 3 red + 2 blue = 5 balls. Two drawn without replacement. Both same colour = both red OR both blue. P(both red) = (3/5) × (2/4) = 6/20 = 3/10 P(both blue) = (2/5) × (1/4) = 2/20 = 1/10 P(same colour) = 3/10 + 1/10 = 4/10 = 2/5 Or via combinations: C(3,2) + C(2,2) over C(5,2) = (3 + 1)/10 = 4/10 = 2/5. This is a math question, no tools needed.	Simple probability problem. No tools needed. Total balls: 5 Red: 3, Blue: 2 Draw 2 without replacement. Same colour = both red OR both blue. P(both red) = (3/5)(2/4) = 6/20 = 3/10 P(both blue) = (2/5)(1/4) = 2/20 = 1/10 P(same) = 3/10 + 1/10 = 4/10 = 2/5

Overview

Measured text	With GrugDone	Without GrugDone	Observed difference
Exposed-reasoning words	1,741	2,564	32.1% fewer
Final-answer words	4,819	5,037	4.3% fewer
Combined words	6,560	7,601	13.7% fewer

This was one exploratory run, not evidence of token savings or preserved answer quality. See the report, paired data, and raw outputs.

Installation

Quick installation

Copy skills/grugdone into .agents/skills/grugdone in the target repository. Claude Code users can copy it to .claude/skills/grugdone instead.

This repository includes both ready-to-discover layouts.

Full compatibility details

Environment	Skill location	Invoke
Claude Code	.claude/skills/grugdone/SKILL.md	/grugdone
Codex	.agents/skills/grugdone/SKILL.md	$grugdone
Kimi Code CLI	.agents/skills/grugdone/SKILL.md	/skill:grugdone
OpenCode	.agents/skills/grugdone/SKILL.md	Mention GrugDone; the agent loads it through the skill tool
Cursor	.agents/skills/grugdone/SKILL.md	/grugdone
Antigravity IDE	.agents/skills/grugdone/SKILL.md	Mention GrugDone or let the agent activate it
Antigravity CLI	.agents/skills/grugdone.md	/grugdone
ZCode	Import the Codex/Claude copy in Settings → Skills, or copy it to ~/.zcode/skills/grugdone/SKILL.md	$grugdone

For a global installation, use the environment's documented user-level skill directory.

Z.ai documents GLM Coding Plan as model access for supported coding tools, not as a separate GLM CLI. ZCode is the documented Z.ai environment with native Skill support.

Inspiration

GrugDone started after I watched one of Theo's videos about how efficiently ChatGPT models reason. It made me wonder whether other models could benefit from the same approach: less unnecessary reasoning while keeping final answers useful and complete.

GrugDone is an experiment built around that question, not a claim that every model will achieve identical performance.

Examples

The captured traces above are the current examples. examples/README.md provides a template for future additions.

Limitations

GrugDone is experimental. Results may vary across models, settings, prompts, and agent environments. It cannot guarantee changes to unobserved computation, token use, latency, or answer quality.

Contributing

See CONTRIBUTING.md for guidance on benchmarks, regressions, and documentation improvements.

License

Licensed under the MIT License.