Three months of Claude Code sessions on medium-sized projects, and the first 25,000 to 60,000 tokens of every conversation were gone before a single line of code got written. All of it going toward Claude reading files, mapping the directory structure, figuring out what the codebase does. One developer called it "paying to orient the AI" - and built a tool to stop it.
Codesight is a zero-dependency command-line tool that pre-generates a compact context pack for AI coding assistants. Run it once in your project directory and it writes a .codesight/ folder containing a compressed project structure map, file summaries with key functions and exports, a dependency graph showing how modules connect, and a list of entry points so the AI knows where the application actually starts.
When you open a new Claude Code session, you paste the context pack into your first message. Claude starts already knowing the shape of the project - no exploration calls needed.
The Token Math
For small projects the benefit is modest. For medium to large codebases - roughly 50+ files - the developer reports cutting per-session orientation cost from 25,000-60,000 tokens down to under 2,000. At Claude Sonnet 4.6 pricing, that's $0.075 to $0.18 per session saved. Small individually, but if you're opening 20+ Claude Code sessions daily on a larger project, it adds up to real money over a month.
Codesight works with any AI coding assistant that accepts pasted context: Claude Code, Cursor, GitHub Copilot's chat interface, or plain API calls. The output is plain text, not formatted for any specific tool.
One real limitation: the context pack is a snapshot of your project at the moment you ran it. Major structural changes - new modules, refactored imports, shifted entry points - require a fresh run. The developer suggests adding it to a pre-commit hook or running it at the start of each coding session on active projects. For projects with stable architecture that you return to repeatedly, it's a set-and-forget improvement.
Codesight installs via npm with no configuration required for most projects. It's available on GitHub.