Menu
Grade and steer UI work using Anthropic-style design dimensions—coherence, originality, craft, and usability—for generators and evaluators.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Run red-green-refactor in vertical slices, emphasizing behavior-first tests and harness quality gates.
Break a plan/PRD into independent vertical-slice Beads issues with explicit dependencies.
Synthesize current context into a PRD and file it as a Beads feature/epic issue.
Phase 2 of the Forge pipeline. Turn a research document into an implementation plan that is built around Beads issues and TDD. Reads plans/research/<slug>.md, designs the build, sets up red-green-refactor test strategy using the tdd skill, materializes a Beads epic/feature/tasks with dependencies and priorities, and defines demo/test checkpoints. Use when starting /forge-plan or the plan step of /forgemaster.
Fact-check an implementation plan against the real codebase and first-party documentation, flagging inaccuracies with what / why / evidence and a severity tier. Use when the user wants a plan reviewed or validated for accuracy, runs /review-plan, or wants a plan checked against real code before implementing. Accepts a Beads issue, a filesystem path, or a plan pasted into chat.
Phase 3 of the Forge pipeline. Execute the plan's Beads tasks with red-green-refactor TDD, committing per checkpoint and pausing at each natural stopping point so the user can run a live demo or test and see the work themselves. Use when starting /forge-implement or the implement step of /forgemaster.
| name | ui-originality-criteria |
| description | Grade and steer UI work using Anthropic-style design dimensions—coherence, originality, craft, and usability—for generators and evaluators. |
Use when building or reviewing user-visible UI (marketing pages, dashboards, app shells, flows). Encodes the four grading dimensions from Anthropic — Harness design for long-running application development so subjective quality stays gradable and anti-“AI slop”.
Pair with .claude/protocols/evaluation-rubric.md for harness-wide severity (BLOCKER / HIGH / …); this skill is the product-design lens.
Question: Does the layout feel like one intentional system—not a pile of unrelated components?
Strong: Color, type, spacing, imagery, and motion share a clear mood; hierarchy tells a single story.
FAIL signals: Mismatched radii/shadows; competing focal points; “default component” look with no unifying thread.
Question: Would a human designer see custom decisions, or stock templates / library defaults / clichéd AI patterns?
Strong: Distinct typography pairing, palette, or layout rhythm tied to the product’s domain.
FAIL signals: Unmodified stock cards; purple-on-white hero gradients “because SaaS”; generic three-column icon grids with lorem tone; interchangeable with ten other AI landing pages.
Question: Are fundamentals solid—type scale, spacing rhythm, contrast, alignment?
This is competence, not creativity. Most honest implementations pass unless something is broken.
FAIL signals: Illegible contrast; broken hierarchy; ragged grids; tap targets too small; obvious misalignment.
Question: Can a user understand what the UI does, find primary actions, and complete tasks without guessing?
FAIL signals: Hidden primary actions; misleading labels; dead ends; keyboard traps; unclear empty/error states.
When both generator and evaluator see these criteria, bias feedback toward design quality and originality first—Claude often scores craft and functionality well by default. Use craft and functionality as hygiene gates, not as excuses to ship bland work.
.claude/protocols/evaluation-rubric.mddocs/index.html (Agent Forge static dashboard — NASA retro / space-console theme)