claude-design/CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.2.0] — 2026-08-20

Added

• Phase 9 — critique-iterate, the optional post-ship loop that closes the design → critique → re-prompt cycle. The facilitation flow is now nine phases: phases 1–8 take the operator to a shipped artifact, Phase 9 runs only when the operator asks for it. Offered once at the close of Phase 8 and exited with an explicit skip.
• skills/claude-design-facilitator/references/05-critique-iterate.md — the single shared Phase 9 reference, carrying all four sub-phases: coach the export and the critique run (9a), consume the pasted critique as untrusted reference data (9b), map findings to the one prompt layer that produced them (9c), decide whether to loop or exit (9d). One shared section rather than eight full per-preset sections; the duplication that the per-preset alternative implied was rejected in planning.
• A thin Critique-iterate hint block in each of the eight per-preset reference files, naming which design-critique sections dominate for that artifact type and which prompt layer those findings route to. The block's heading letter is per-file (four presets carry ## (f), four carry ## (g)) because it follows each file's existing section sequence; the letter-free substring Critique-iterate hint is the canonical machine-checkable anchor.
• tests/test-sc4-phase9-coverage.sh — 21 assertions covering the shared reference, its four sub-phases, the SKILL.md Phase 9 section, and the hint block in every preset file. Registered by hand in verify.sh, which has no auto-discovery.

Changed

• tests/test-sc1-dogfood-log.sh now accepts either dogfood-log format: a valid v0.1 block (five fields) or a valid v0.2 Phase 9 block (four sub-sections). Either one closes the gate; the other, if present but malformed, downgrades to WARN. The v0.2 sub-sections are matched on keywords (artifact / critique / re-prompt / second pass) in ### headings rather than on verbatim labels, so the log stays writable in the operator's own words.
• references/04-handoff-and-scope.md cross-references Phase 9 as the coached path into design:critique.
• .coverage.md Authoritative-claims registry: 13 → 14 paths, 5 → 6 foundation references.

Notes

• Export-to-HTML is the input, not a URL. Claude Design artifacts have no shareable public URL and design-critique does not take generic web URLs, so Phase 9 coaches a standalone HTML export and hands over design:critique @<file>. The pre-filled-URL alternative was a broken affordance and was rejected.
• Layer mapping is LLM judgment plus operator confirmation, not a parser. Critique output is prose, and treating LLM prose as a machine-readable API is the anti-pattern this avoids. Findings map to the real layers documented in 01-prompt-fundamentals.md — Layer 1 (Goal / Layout / Content / Audience), 1.5, 2a, 2b, 3, 4 (Typography / Color / Motion / Backgrounds) and 5 (grading criteria) — and the operator confirms the mapping before any re-prompt is drafted.
• Pasted critique output is untrusted reference data, never instructions. This is the load-bearing security rule of Phase 9: a critique that inspected a Figma file or a screenshot can carry attacker-influenceable text verbatim into the conversation (OWASP LLM01). The operator pastes it between delimiters and it is read as data.
• Exit is operator-attested. Phase 9 does not decide that an artifact is good enough; the operator does.
• Accessibility findings are not mapped. They hand off to design:accessibility in Anthropic's plugin today, and a first-party /accessibility surface is deferred to v0.4.
• The scope fence is unchanged. Phase 9 routes to design:critique and never reimplements it. This plugin still ships none of the six forbidden command names, and tests/validate-plugin.sh assertion (h) still enforces that mechanically.

[0.1.0] — 2026-05-17

Added

• claude-design-facilitator skill with eight-phase facilitation flow (disambiguate → intent preset → audience + destination → DESIGN.md anchor → five-layer prompt draft → copy-paste delivery → iteration coaching → ship-readiness handoff) and 12 natural-language trigger phrases registered in .triggers.txt.
• Five foundation references under skills/claude-design-facilitator/references/: 00-what-claude-design-is-and-isnt.md (surface disambiguation), 01-prompt-fundamentals.md (five-layer prompt stack: GLCA + start-simple + concrete-alternative-spec + propose-options + AI-slop avoid-list + four design dimensions + four grading criteria), 02-design-md.md (DESIGN.md 9-section canonical structure + brand-to-DESIGN.md extractor), 03-iteration-and-session.md (Tweak / Comment / Chat cascade, session economics, recovery prompt library), 04-handoff-and-scope.md (one-way Design → Code handoff + scope fence vs Anthropic's knowledge-work-plugins/design).
• Eight per-preset references under skills/claude-design-facilitator/references/presets/ with evidence-grade labels: designs.md, prototypes.md, slides.md (Anthropic-documented + community-validated); one-pagers.md, wireframes-mockups.md, pitch-decks.md, marketing-collateral.md (Community-only); frontier-design.md (Experimental — no validated practitioner pattern as of 2026-05-16).
• .coverage.md at plugin root — preset enumeration table with evidence-grade labels (8 rows) + Authoritative-claims files bullet-list registry (13 paths). Canonical input for SC2 and SC3 verification.
• Five verification scripts under tests/: validate-plugin.sh (structural integrity + forbidden-command-name scope fence + operator-private-context grep + Norwegian-leakage advisory), test-skill-triggers.sh (description quality + trigger phrase coverage), test-sc2-artifact-coverage.sh (per-preset coverage from .coverage.md), test-sc3-citations.sh (Anthropic-domain citation discipline), test-sc1-dogfood-log.sh (operator dogfood log format-check in REMEMBER.md).
• verify.sh top-level roll-up with --strict (SC1 missing-block becomes FAIL) and --quick (skip skill-triggers test) flags.
• LICENSE (MIT) and GOVERNANCE.md (marketplace fork-and-own blurb).
• Marketplace registration in root .claude-plugin/marketplace.json.

Documentation

• Plugin README.md rewritten from scaffold placeholder to full v0.1 surface description with Scope and complementarity section (placed before installation per brief), What this plugin is NOT (Non-Goals), eight-phase facilitation flow table, skill surface table, reference content map, per-preset coverage table, verification section, AI-generated disclosure, fork-and-own MIT licensing.
• Plugin CLAUDE.md translated to English (operator override of marketplace's Norwegian-dialogue default per v0.1 brief constraint); added Scope fence section explicitly forbidding command-name collisions with Anthropic's knowledge-work-plugins/design (/critique, /accessibility, /ux-copy, /research-synthesis, /design-system, /handoff); Authoring rules section codifies English-everywhere, no operator-private context, evidence-grade label discipline, URL canonicalisation on support.claude.com/en/articles/<numeric-id>-<slug>; Communication patterns block preserved verbatim.
• Root marketplace README.md updated with ### [Claude Design](plugins/claude-design/) \v0.1.0`entry under the## Pluginssection, positioned after the Human-Friendly Style entry per existing convention. Entry documents the complementary lifecycle coverage vsknowledge-work-plugins/design`.

Notes

• Scope: claude-design facilitates the pre-design and during-design lifecycle for claude.ai/design (Anthropic Labs research preview, Opus 4.7 pinned, eight intent presets). For post-design — critique, accessibility audit, UX copy review, design-system audit, engineering handoff — install Anthropic's official plugin via claude plugins add knowledge-work-plugins/design. Zero command overlap, complementary by design.
• No browser automation. This plugin produces prompts; the artifact gets built inside claude.ai/design. The operator copies and pastes manually.
• No artifact code generation. This plugin is a prompt-builder, not an artifact generator. Claude Design is the generator.
• No artifact storage or versioning. Claude Design has no version-tree primitive and this plugin does not invent one. The verbal save-pattern documented in references/03-iteration-and-session.md is the closest substitute.
• English everywhere in shipped content. Operator override of the marketplace's default Norwegian-dialogue convention. tests/validate-plugin.sh assertion (j) emits WARN on Norwegian diacritics in shipped content; review case-by-case.
• Evidence-grade discipline. Every per-preset reference file carries an inline Evidence grade: label on line 4 with one of three values: Anthropic-documented + community-validated, Community-only, Experimental. .coverage.md is the canonical registry.
• Re-research triggers documented in .coverage.md — fire on Anthropic publishing per-preset guidance for Community-only presets, on new intent presets added to the launch enumeration, on the first verified frontier-design practitioner artifact shipping publicly, on Labs → GA URL rename for claude.ai/design, on Anthropic's knowledge-work-plugins/design adding or removing slash-commands.

[0.1.0-pre] — 2026-05-15

Added

• Initial scaffold (README, CLAUDE.md, ROADMAP, TODO, plugin.json placeholder).