This git timeline makes Claude system prompts easy to track

Simon Willison has published a new project that turns Anthropic’s Claude system prompt history into a browsable git repo. With prompts split into granular files and timestamped commits, you can use git log, diff, and blame to trace changes over time.

•
Claude
claude cover

TL;DR

  • Project: Claude system prompts as a git timeline reformats Anthropic’s Claude system prompt history into a git repo
  • Purpose: Easier prompt-change review over time than a single monolithic markdown document
  • Repo structure: Prompts split into granular files, grouped by model, family, and revision, with timestamped commits
  • Git workflows enabled: git log for change tracing, diff for revision comparisons, blame for date-attributed lines
  • Published as a “beat” on 18 April 2026, 12:25 pm
  • Related links: PyCon US 2026 AI/security tracks; Qwen drawing vs Claude Opus 4.7; Meta Muse Spark in meta.ai chat

Simon Willison has published a new research project, Claude system prompts as a git timeline, that turns Anthropic’s published system prompt history for Claude into a format that can be examined like a software codebase. The goal is straightforward: make it easier to understand how Claude’s system prompts have changed over time, without having to manually sift through a single large markdown document.

In Willison’s description, the original monolithic markdown source is broken into smaller, more navigable pieces, with extracted prompts organized into granular files and paired with timestamped commits. The repository’s structure groups prompts by model, family, and revision, so changes can be reviewed at different levels of detail.

Why a git-based timeline matters

By putting the prompt history into git, the project makes common tooling do the heavy lifting. Willison highlights standard git workflows that become immediately useful once prompts are split into individual files:

  • git log to trace when prompts changed
  • diff to compare revisions
  • blame to attribute lines of text to specific dates

Taken together, that enables prompt evolution to be tracked and compared over time, with changes tied to specific points on the timeline—without manual parsing.

Publication details

Willison posted the item as a short “beat” on 18th April 2026 at 12:25 pm.

The page also points to three recent posts from Willison: one about PyCon US 2026 in Long Beach and new AI and security tracks (17th April 2026), another comparing a local Qwen model’s drawing results with Claude Opus 4.7 (16th April 2026), and a third about Meta’s Muse Spark and tools in meta.ai chat (8th April 2026). Separately, it includes a link to sponsor a monthly briefing for $10/month.

Source: https://simonwillison.net/2026/Apr/18/extract-system-prompts/

Continue the conversation on Slack

Did this article spark your interest? Join our community of experts and enthusiasts to dive deeper, ask questions, and share your ideas.

Join our community

Related Articles

claude cover

Claude adds caching dashboard to show cost and TTFT impact

Claude has just rolled out a caching-focused dashboard in the Claude Developer Console, aiming to give developers clearer visibility into how prompt caching affects spend and time-to-first-token. The company is also spotlighting new resources to help improve cache hit rates.

1 shared tag
Opus 4.7’s mixed reviews expose cracks in AI evaluation

Opus 4.7’s mixed reviews expose cracks in AI evaluation

Anthropic has just rolled out Opus 4.7, and reactions are sharply split. Simone Civetta argues the divide reflects uneven workflow impact, shifting benchmarks, and prompt breakage tied to stronger instruction following. He suggests higher “effort” modes to curb under-thinking.

1 shared tag
claude cover

Claude Opus 4.7 is more literal: 10 migration moves

Paweł Huryn breaks down what changed in Claude Opus 4.7 versus 4.6—and why some “regressions” are really prompt friction. He shares 10 practical migration moves, from intent-first context files to batching asks and updating adaptive thinking settings.

1 shared tag