Terse plugin for Claude Code enforces 22 writing rules to curb verbose AI prose

The Terse plugin for Claude Code addresses the tendency of AI models to produce verbose, metaphor-heavy, and repetitive responses, often referred to as "Claude-speak." By enforcing 22 specific writing rules, the plugin mandates that the first sentence contains the answer, limits responses to 150 words by default, and prohibits recaps, slogans, and closing offers like "happy to dig further."

Performance and Cost Impact

While the primary goal is stylistic clarity, the reduction in verbosity yields significant efficiency gains. Benchmarks comparing vanilla Claude Code against the Terse plugin across 20 prompts show:

• Opus 5.5: Word count dropped from 9,705 to 4,468 (a ~54% reduction), resulting in a 32% lower cost.
• Fable 5.1: Word count decreased by 46%, with costs reduced by approximately 50%.

Technical Implementation

The plugin operates as a forced output style rather than a simple token saver. It includes a hook that re-sends a short digest with every prompt to ensure rules persist across long sessions (up to turn 30) and applies the same constraints to subagents. Unlike the built-in "Concise" configuration, which only trims preambles, Terse actively replaces metaphors with literal phrases and contrasts with direct mechanisms.

Development and Availability

Built over three weeks using Claude Code itself, the tool includes a custom scorer and an Opus-based judge to count style violations. The project is free, MIT-licensed, and available via the Claude plugin directory or by running /plugin marketplace add lowenbjer/claude-terse.