Using GraphViz for Claude.md (blog.fsck.com)

An informal researcher experiment turned into a practical workflow idea: after hearing that agents might formalize processes with .dot (GraphViz), the author asked Claude to convert their existing CLAUDE.md process document into a GraphViz .dot graph. They iterated a dozen times with the model, fixed inconsistencies and one over-ambitious hallucination, produced a readable .dot representation, swapped it in for the original Markdown rules, and found Claude followed the .dot rules at least as well — and in some quick, informal comparisons, better — than the original text. The project also produced a .dot style guide to make it easier for Claude to generate or update its own process graphs; the author notes this mirrors Anthropic’s SKILL.md concept (Markdown files with YAML frontmatter), but uses .dot for explicit structure and visualization. Why this matters: Graph-based .dot specs turn freeform process prose into a machine- and human-readable formalism that reduces English ambiguity, makes workflows easy to visualize and verify, and can reveal rule conflicts quickly. That improves rule-following, debugging, and potentially enables agents to self-document and evolve capabilities in a verifiable format. Caveats: the tests were informal, and LLMs can still hallucinate (the author saw one sci‑fi divergence). Still, using .dot as an intermediate spec language — plus a style guide and tooling for conversion — is a low-friction way to make agent behavior more explicit, testable, and maintainable.