interform

keep priority P2 effort S wave 2

What it does: A single-skill plugin (1 skill, 0 agents/commands/hooks/MCP). The whole substance is skills/distinctive-design/SKILL.md: a ~40-line prompt-injection that steers Claude away from generic “AI slop” aesthetics (named anti-patterns: Inter/Roboto/Arial fonts, purple-on-white gradients, card grids, symmetric layouts) and toward a committed, medium-aware aesthetic direction across web, TUI, native, and print. It raises the floor on generated-interface visual quality; it explicitly does not enforce anything at runtime.

Rationale: The plugin is correctly scoped and correctly packaged by 2026 standards. Its single thin SKILL.md with on-demand metadata loading is exactly the progressive-disclosure / “skills beat prompts” pattern the field converged on (SKILL.md as the portable capability unit). There is no MCP bloat, no eager tool loading, nothing to modernize on harness mechanics. The content is concrete and actionable (named fonts/clichés to avoid) rather than vague, which is what makes a design skill useful. It does one thing well and its own vision doc deliberately caps complexity (“remain a single-skill plugin”). That is a healthy mature plugin — hence keep, with targeted improvement rather than rework.

SOTA gap: Lacks a verification/critic loop. SOTA (2026) is emphatic that “verification is the new bottleneck” and that generic unstructured NL critique is largely ineffective — improvement comes from structured, tool-grounded feedback. interform is a pure generator: it produces interfaces but never checks its own output against its own stated anti-patterns. Its PHILOSOPHY.md shows “Brainstorms analyzed: 0 / Plans analyzed: 0 / Source confidence: inferred,” and its vision doc concedes it does not enforce rules via hooks/checks. The frontier idea it lacks is closing the generate→verify loop with grounded signals (e.g., delegate a post-generation design audit to intersight’s token/component extraction), turning the skill’s anti-slop opinions into checkable assertions rather than hopeful prose.

Redundancy: Complementary, not duplicative today: intersight does automated UI/UX analysis (W3C DTCG token extraction, component inventory, vision-based layout analysis from URLs) and interbrowse:teardown does UX teardowns — these are the analysis/critique side, while interform is the generation side. The risk is in interform’s own roadmap (IFRM-N3 “layout and accessibility checks,” IFRM-L1 “visual style registry”): if built inside interform these would directly overlap intersight. Keep interform generation-only and route verification through intersight rather than reimplementing token/layout analysis.

Actions:

Close the generate->verify loop: after generating an interface, optionally delegate a grounded audit to intersight (DTCG token + component extraction) so the skill’s named anti-patterns (Inter font, purple gradients, card grids) become checkable assertions instead of unverified prose — cites the ‘verification is the bottleneck’ + ‘tool-grounded critique beats NL critique’ SOTA.
Build the docs/examples/ before/after gallery the vision doc promises; concrete artifacts feeding back into the guideline (per the self-improvement / experience-distillation trend) materially strengthen a prompt-only skill.
Reconcile roadmap.json with reality: IFRM-N1 (validation hooks) and IFRM-N3 (accessibility/layout checks) imply tooling the plugin does not ship and would overlap intersight — either delete these or re-scope them as ‘delegate to intersight’ so the roadmap stops promising a duplicate analysis engine.
Add a thin accessibility floor to the SKILL.md (contrast ratios, focus states, reduced-motion) — currently it optimizes for boldness with no a11y guardrail, which is a credibility gap for ‘production-grade.’