What gets measured
The ⊨ Claude Code plugin turns scoring into a feedback loop. Evaluate your draft, then let the agent iterate until traits converge.
Streamable HTTP — works with Claude Code, Cursor, Windsurf, or anything that speaks MCP.
# Claude Code
$ claude mcp add --transport http u22a8 https://u22a8.ai/mcp
$ cat post.md | curl -s -d @- \
-H "Accept: text/plain" \
https://u22a8.ai/m/u22a8.technical-writing
Status: ready — not yet trained.
Usable guidance ↔ Passive description
Whether the writing gives the reader something they can immediately do, apply, or run — concrete steps, runnable code, reusable templates — versus describing things without enabling action
Problem-anchored ↔ Unmotivated exposition
Whether each new concept or technique is introduced in response to a visible problem or limitation — versus presenting information without explaining what drives the need for it
Evidence-backed precision ↔ Vague handwaving
Whether the writing uses specific numbers, real examples, production graphs, named systems, and concrete scenarios — versus vague generalizations, marketing language, or abstract reassurances
Layered build-up ↔ Complexity dump
Whether the writing introduces components or concepts one at a time and composes them step by step — versus presenting the full system upfront and expecting the reader to absorb everything simultaneously
Coherent arc ↔ Disconnected sections
Whether the writing maintains a visible story or argument that connects each section back to a central question or metaphor — versus presenting information as a flat list of loosely related topics
Concrete-first ↔ Abstract-first
Whether the writing starts with a specific, tangible example (a real server, a single neuron, an arithmetic expression) and builds toward abstraction — versus leading with definitions, formulas, or theoretical frameworks