Claude Code's system prompts, archived and diffed by a bot

What it does This repo automatically detects new Claude Code releases, extracts the underlying system prompts and feature flags, and archives them as tagged, browsable artifacts. It organizes prompts by category in system-prompts/ and aggregates metadata—token counts, CLI surface changes, flag states—in meta/. The result is a searchable history of how Anthropic’s coding agent talks to itself.

The interesting bit The project is essentially a transparency engine built on top of cchistory’s extraction tooling, feeding the @ClaudeCodeLog X account with structured diff summaries. It treats Claude Code’s closed-source prompt engineering as a moving target worth tracking, not a black box to ignore.

Key highlights

• 45 tracked prompt files with indices sorted by token count, init date, and last edit
• Automated detection of new releases with annotated Git tags for historical comparison
• Derived metadata including estimated token totals, prompt statistics, and CLI surface changes
• Powers an X/Twitter bot that publishes threaded diff summaries with screenshots
• Community-maintained and unofficial; built on prior extraction work by cchistory and Piebald-AI

Caveats

• Token counts and line-of-code metrics are estimates, not exact figures
• Prompt files may contain near-duplicates or intentionally duplicated variants across releases
• File names and structures can shift between versions (renames, splits, merges), making longitudinal tracking noisy

Verdict Worth bookmarking if you study prompt engineering, agent behavior drift, or Anthropic’s product strategy. Skip it if you just want to use Claude Code without knowing what the system prompt told it five versions ago.