Using the WAT Framework: Writing Sanity MCP Workflows That Make Claude Consistent and Reliable

The Problem With Asking AI to "Just Figure It Out"

When you give Claude an open-ended task like "write a blog post about TypeScript," you get a different result every time. The structure varies. Fields get missed. Edge cases are ignored. That's not a bug — it's the nature of probabilistic reasoning. Without constraints, the model explores a wide solution space. For repetitive, structured tasks like publishing content to a CMS, that's a liability.

What the WAT Framework Is

WAT stands for Workflows, Agents, Tools. It's a three-layer architecture that separates concerns: Workflows are plain-language markdown SOPs that define what needs to happen and how. Agents — Claude — handle reasoning, sequencing, and decision-making. Tools are deterministic scripts or APIs (like the Sanity MCP) that do the actual execution. Each layer does what it does best. The workflow encodes your domain knowledge. Claude handles judgment calls. The tools handle the mechanics.

Why Workflows Make AI Better

A workflow narrows the decision space. Instead of reasoning from scratch each time, Claude follows a defined path: collect inputs, fetch context, draft an outline, get approval, write content, create the document. Every step is explicit. Edge cases are handled in advance. The result is consistent, repeatable behavior — more like a reliable colleague following a process than a creative generalist making it up as they go. If each step is 90% accurate and you have five steps, an improvised chain lands at 59% success. A workflow keeps you close to 90% at every step.

Building a Sanity MCP Workflow

The draft_blog_post workflow is a practical example. It starts by fetching authors and categories from Sanity via GROQ so references are always valid. It requires outline approval before writing — a quality gate that prevents wasted effort. The document shape is specified exactly: the Portable Text body structure, SEO fields, what not to set (autoSummary is generated on publish). Discovered constraints get documented too, like the 5000-byte body limit imposed by the audio narration pipeline.

The Self-Improvement Loop

Workflows are not static. When you hit a slug conflict, you add a rule: append -draft. When you discover the body size limit, you document it and add a trimming instruction. Each failure makes the workflow stronger. Over time it accumulates hard-won knowledge about your specific setup — knowledge that would otherwise live only in your head or get rediscovered every session.

Getting Started With Your Own Workflows

A good workflow starts with a repeatable task you find yourself explaining more than once. Write down the inputs, the steps, which tools to call, and the expected output. Add an edge cases table as you encounter them. Store it in .claude/wat/workflows/ and reference it by name when you start a task. The payoff compounds: every session that follows the workflow is faster, more consistent, and cheaper than one that improvises from scratch.