AHD / Artificial Human Design

A framework for running design past a human before it ships.

You have already seen the slop. The sameness. The confident wrongness of generated output that no one reviewed. AHD does not lecture about this problem; it gives you a toolchain to prevent it in your own workflow.

The system compiles a design brief into a structured evaluation, runs it against frontier vision models with explicit rubrics, and returns scored criticism you can act on. Not scores for vanity. Scores that block merge until they are addressed.

$ ahd eval-live swiss-editorial --brief briefs/landing.yml --models <specs> --n 10
# Compiles brief → runs 28-rule lint → dispatches to vision critic → returns scored report

OUTPUT

A directory of evaluated frames with per-rule scores, a deduplicated issue list ranked by severity, and a structured brief you can check in and diff. The vision critic runs with a frozen system prompt and token-constrained output, so results are reproducible across runs and models.

What ships

Brief compiler YAML/JSON spec to internal representation

28-rule linter Static analysis without model calls

Eval harness Parallel runner with timeout and retry

Vision-critic scaffold System prompt, rubric JSON schema, scorer

MCP server Model Context Protocol for IDE integration

Editor plugins VS Code and Neovim (source in repo, manual install)

Eight tokens Curated design vocabulary for prompt grounding

What is gated

Live frontier-model calls Requires your own API keys

Live vision critique Needs multimodal key + screenshot pipeline setup

Standalone npm packages for editor plugins Publishing pipeline pending

Additional tokens Process for community submission not yet open

Read the code — GitHub →

forgejo mirror available on request