The CLAUDE.md you'd write
if you had the time.
A free best-practices template for Claude Code, updated automatically each week. Thirteen practitioner sources monitored and scored. Every technical claim verified against the official documentation before publication.
Maintained by the State of CLAUDE.md weekly skill · sf1997-spec/state-of-claude-md
Claude Code moves fast.
CLAUDE.md files don't.
Every week, practitioners at companies including Anthropic, Marmelab, and PSPDFKit publish new guidance on extracting reliable output from Claude Code. That guidance is scattered across blog posts, GitHub commits, Hacker News threads, and conference talks — each in a different format, with different levels of rigour, rarely cross-referenced.
The standard response is to write a CLAUDE.md once and leave it. Within weeks it reflects conventions for a model version that has been superseded, references commands that have been renamed, and omits patterns the community has since validated through sustained daily use.
State of CLAUDE.md exists to close that gap. It monitors the sources, verifies the technical claims, scores the candidates, and publishes the result.
Under 200 lines. Four core rules.
The template is structured around four universal behavioural rules — Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution — that survive model updates because they address how Claude reasons rather than how it behaves in a specific version. Surrounding them are sections on context management, verification discipline, and safety constraints, each updated as new evidence emerges.
Loading the current edition…View the full template ↗ github.com/sf1997-spec/state-of-claude-md
Loading this week's edition…
Not curated by hand.
An automated evidence pipeline.
Each week the system fetches content from thirteen sources, filters by a relevance vocabulary, and extracts candidate items — specific claims about how to configure Claude Code effectively. Tier 1 sources are Anthropic's official documentation. Tier 2 sources are practitioner blogs and high-starred GitHub repositories. Tier 3 sources are community signals that require corroboration before reaching the template.
A verification subagent checks every technical claim against the current official documentation before scoring begins. Named commands, keyboard shortcuts, configuration field names, and file paths must match verbatim. A partial match fails. Items that pass verification are scored against a published rubric across four dimensions: source authority, cross-source validation, recency, and claim type. Only items with two or more independent sources reaching the minimum threshold are committed to the template.
Read the scoring rubric →Get each new edition by email.
A short note whenever the CLAUDE.md is updated — what changed, what was excluded, and what's under evaluation. No marketing. Unsubscribe any time.