Last week Claude Code source code was leaked and became public available. So I asked my agent to compare the reverse engineered tools that I asked Claude Code to do a few weeks ago with the actual implementation.
Apart from funny wording “Worth stealing from Claude” outputted from the LLM model, generally speaking the tools are pretty similar, with Claude Code having more safeguards that what I vibe coded. Some interesting discoveries:
- System prompt is dynamically built based on model of choice and available tools
- Core agent loop is nothing special, standard turn based agent conversation
- Nothing concerning found on the telemetry side, basic application tracking. They do track every command and tool call made in the agent
To me the only value to this is that now I can better adjust my tool prompts to match what their LLM expects.
Ah, and there’s a tamagochi-style buddy disabled by a feature flag in the code.