Rodman reduces Claude tokens to 3.7M

Leonard Rodman said on May 24 that he had reduced Claude Code token usage in a documented run from 10.4 million to 3.7 million by using InsForge Skills and the InsForge CLI. Rodman, an open-source builder who posts as @RodmanAi, said in an X post that the same setup cut errors to zero from 10. The post framed the change as “context engineering, not prompt engineering,” and pointed readers to InsForge’s GitHub repositories for the implementation. ### What exactly did Rodman say changed? Rodman’s May 24 post said the before-and-after comparison was 10.4 million tokens versus 3.7 million, with errors falling from 10 to zero. The figures matter because Claude Code is Anthropic’s terminal-based coding agent, where long sessions can accumulate large context windows and repeated tool calls. Anthropic describes Claude Code as an agentic coding tool that lives in the terminal, understands a codebase and executes routine tasks through natural-language commands. (github.com) The claim has not been independently benchmarked in the material available here. What can be verified is that the workflow Rodman referenced maps to public InsForge components on GitHub: a main InsForge repository, a separate CLI repository and a skills repository that is explicitly designed for Claude Code and other coding agents. (github.com) ### What are InsForge Skills and the CLI doing in that workflow? InsForge says its platform gives coding agents database, auth, storage, compute, hosting and an AI gateway for building full-stack applications. Its public documentation says agents can interact with InsForge in two ways: through an MCP server, or through a CLI paired with Skills that agents invoke directly from the terminal. (github.com) The CLI repository says running setup commands can automatically install InsForge agent skills for supported coding agents including Claude Code. The skills repository separately describes itself as “Agent Skills for InsForge” and shows installation paths through a skills registry or through Claude Code’s `/install-skills` command. ### Why would that reduce token use? (github.com) Anthropic’s public skills repository describes Skills as reusable agent capabilities for Claude. In practice, that means a coding agent can call structured tools and predefined workflows instead of repeatedly re-deriving instructions from raw prompt context. That is consistent with Rodman’s description of the gain as coming from context engineering rather than prompt changes. (github.com) InsForge’s documentation makes the same distinction in implementation terms. Its docs say the CLI-plus-Skills path lets coding agents operate backend tasks directly from the terminal, which can reduce the need to restate project structure and operational steps inside the model’s conversational context. That is an inference from the published architecture, not an independently measured explanation of Rodman’s exact run. (github.com) ### How much of this is verifiable today? The public repositories are verifiable today. InsForge’s GitHub organization lists the main InsForge project, the CLI package and the insforge-skills package, and the repositories were updated recently. Anthropic’s Claude Code repository and public skills repository are also live on GitHub. The performance result itself remains Rodman’s own reported demonstration. (github.com) Anyone trying to reproduce it would need Rodman’s exact task, baseline setup, model configuration and error-counting method. The next concrete step is in the code: InsForge’s GitHub repositories publish the CLI and skills packages Rodman referenced, and Anthropic’s Claude Code and skills repositories document how those components are installed and used. (github.com 1) (github.com 2)