Member-only story
Cut Claude Code tokens: rules, skills, and prompts
Now token burning is real need to look in skills, prompt, context to reduce it and limit upfront token cost $.
You can run the whole software lifecycle through Claude Code and still burn the plan limit before lunch. The feature is rarely the expensive part. The expensive part is what the session already loaded: a long instruction file, unscoped rules, every skill description in a plugin, and a conversation you never cleared.
Anthropic’s costs guide states the mechanism: token cost scales with context size, because each request sends that context again. Prompt caching makes the repeated prefix cheaper. It does not make a bloated prefix free. On a subscription, /usage shows plan usage rather than an API invoice. On an API key or cloud provider, the same screen estimates dollars. Either way, start by reading that screen before you rewrite prompts.
This is the layout I would commit to a GitHub repo on day one, then the habits that keep an agent-only workflow inside it. Examples use an ASP.NET Core orders API with Entity Framework Core. The practices come from Claude Code’s own docs, Matt Pocock’s skills, and codewithmukesh/dotnet-claude-kit. I did not run a benchmark on your solution. Where a repo publishes its own token estimate, it is labeled as theirs.
