The complete system prompt behind Claude Code, Anthropic's command-line AI coding agent, has been published as a public GitHub gist - and it's a fascinating look at how a production AI coding tool actually thinks.
The prompt runs over 4,000 words of structured instructions covering everything from file editing rules to security boundaries. A few details stand out. Claude Code runs on Opus 4.6 with a 1 million token context window (roughly 2,500 pages of text). It uses a file-based persistent memory system stored at ~/.claude/projects/, letting it carry context about you and your codebase across sessions. When conversations get long, the system automatically compresses older messages rather than losing them.
The coding philosophy baked into the prompt is opinionated in ways that experienced developers will appreciate. "Three similar lines of code is better than a premature abstraction." Don't add docstrings to code you didn't change. Don't build for hypothetical future requirements. Only validate at system boundaries. These aren't generic AI instructions - they read like a senior engineer's code review checklist.
The safety layer is equally specific. Before any destructive action like git reset --hard or deleting files, Claude Code must check with the user first. It refuses force-pushes to main without explicit permission. It won't skip git hooks. For security work, it requires clear authorization context like a pentest engagement or CTF competition.
None of this is secret, exactly. System prompts for AI tools have been surfacing regularly, and Anthropic hasn't tried to obfuscate this one. But seeing the full instructions in one place gives a clearer picture of what "agentic coding" actually means in practice right now: a tightly scoped set of rules trying to balance capability with the kind of caution you'd want from anything that can run shell commands on your machine.