Detailed Analysis
A GitHub project called "justsaydone" has surfaced promising a dramatic reduction in Claude Code's output verbosity—up to 95% shorter replies—aimed squarely at cutting token spend for developers using Anthropic's coding-focused AI tool. The premise is straightforward: Claude Code, like many agentic coding assistants, often generates lengthy explanatory text alongside or instead of direct code changes, and this tool appears designed to suppress or minimize that chatter, prompting the model to simply confirm task completion rather than narrate its reasoning or summarize its work in prose. For developers running high-volume, iterative coding sessions, this kind of intervention can translate into meaningful cost savings, since Anthropic's API pricing is metered by token consumption on both the input and output sides.
The underlying issue this project addresses is a well-known tension in agentic AI tools: verbosity versus efficiency. Claude Code, like Cursor, GitHub Copilot, and other AI pair-programming tools, is often criticized by power users for over-explaining its actions—restating what a function does, summarizing changes it just made, or offering unsolicited context—when the user simply wants the diff or the file update. This behavior is partly a byproduct of how large language models are trained to be helpful and communicative by default, optimized for satisfying a broad range of use cases including one-off Q&A rather than tight, repetitive developer loops. When multiplied across thousands of tool calls in a long agentic session, these small verbal flourishes compound into real token costs and slower turnaround times, which matters considerably to teams running Claude Code at scale or through API billing.
This grassroots tooling response is emblematic of a broader trend in the AI coding assistant ecosystem: third-party developers building lightweight wrappers, prompts, or configuration layers on top of foundation model APIs to tune behavior that the underlying vendor hasn't yet optimized out of the box. Anthropic itself has iterated on Claude Code's default behavior and system prompts over time, and has introduced features like prompt caching and more configurable verbosity settings, but community-built solutions like this one suggest that developers still find gaps between default behavior and their actual workflow needs. The rise of such community tools also reflects the maturation of Claude Code's ecosystem—enough people are using it heavily enough, and hitting real cost friction, that homegrown efficiency hacks are becoming a genre of their own, similar to the plugin and extension economies that emerged around earlier developer tools like VS Code.
More broadly, this development touches on the economics of agentic AI adoption. As coding agents move from novelty to daily-driver status inside engineering organizations, token efficiency becomes a first-order concern alongside raw model capability. Anthropic's pricing model, tied directly to input and output token volume, creates a direct incentive for the community to optimize away unnecessary verbosity, and tools like "justsaydone" represent an early, informal signal of what enterprise buyers will likely demand more formally: fine-grained control over response length, tone, and format to manage spend as usage scales. It also hints at a future where "output discipline" becomes a standard configuration option or fine-tuning target for coding-specific models, rather than something developers have to hack around themselves.
Read original article →