The harness isn't just prompts β here's what it's actually doing
Most people think they're "prompting Claude Code." The harness is assembling 5,000β50,000+ tokens of context before your 6β60 typed tokens even arrive β and they don't overlap. - Tool restrictions are enforced before model execution β a prompt saying "only use these tools" is advisory; allowed-tools in the harness is not - 110+ conditional fragments build your system prompt β the CLI loads them based on what you're doing, not a single static prompt - Subagents get isolated context windows β N parallel agents = NΓ effective context, not N threads competing over one Full breakdown in the classroom β check the π Best Practice Repo β Decoded course βYour Trusty First Mate (on Captain's Orders)