Claude Profiles

You can offload code work and even thinking to an agent. You cannot offload understanding what's loaded into its context.

A Claude Code session's behavior is the sum of a stack: the harness system prompt, API-side injections, the difference between claude.ai and claude-code, MCPs, skills, CLAUDE.md, and plugins. Plugins themselves can smuggle in skills, hooks, and agents. When something goes sideways, the cause is usually in a layer you didn't know was loaded.

This site argues for a small set of shell aliases that replace the system prompt instead of layering on top, and defer MCP and tool loading, so each context starts from a known baseline you control. A handful of aliases you wrote and understand beats thirty plugin and skill definitions scattered across project folders.

The argument has a number under it. The harness ships about 6,400 tokens of baseline system prompt; the reference setup replaces it with 4,500 tokens. With ENABLE_TOOL_SEARCH=true, tool schemas drop from "every MCP tool upfront" to nine, totalling another ~9,100 tokens. About 15,800 input tokens are spent before the user says a word. That floor is paid every turn, every session, every account. Measure your setup walks through the captures.

Who this is for

Engineers ramping up on Claude Code who are starting to install plugins, skills, and MCPs without a clear picture of what each one injects. If you're running on AWS Bedrock or any usage-priced API, that 15,800-token floor is real money on a real ledger.

How to read this

The context stack is the diagnosis: what's actually in your window before your first prompt. Measure your setup is the proof: capture the harness baseline, capture your real session, diff them. Profile structure is the prescription. Build your own walks through assembly. Reference setup shows the working configuration this site runs against.