Using the WAT Framework: Writing Sanity MCP Workflows That Make Claude Consistent and Reliable
The text explains why open-ended AI instructions like “write a blog post about TypeScript” lead to inconsistent results. Because models are probabilistic, they vary structure, miss fields, and overlook edge cases, which is problematic for repetitive, structured tasks such as publishing to a CMS. To solve this, it introduces the WAT framework: Workflows, Agents, Tools. Workflows are plain-language markdown SOPs that encode domain knowledge and specify steps. Agents (Claude) handle reasoning and decisions. Tools are deterministic scripts or APIs, like the Sanity MCP, that execute actions. This separation narrows the decision space and keeps behavior consistent across sessions. A concrete example is the draft_blog_post workflow, which fetches authors and categories from Sanity, requires outline approval, and strictly defines document shape and constraints, including a 5000-byte body limit. Workflows evolve through a self-improvement loop: each failure adds new rules and edge cases. To get started, you document repeatable tasks, inputs, tools, steps, and edge cases, store them in .claude/wat/workflows/, and reuse them for faster, cheaper, and more reliable AI-assisted work.
