claude-design

End-to-end facilitator for prompting Claude Design (claude.ai/design) — idea to copy-paste-ready prompt with iteration coaching, citing Anthropic primary sources.

A Claude Code plugin that ships one skill (claude-design-facilitator) plus a reference tree for prompting Anthropic's claude.ai/design workspace. The skill auto-fires on natural-language triggers, walks the operator through a nine-phase facilitation flow, and produces a prompt grounded in Anthropic's verbatim Goal / Layout / Content / Audience framework and the published per-preset prompt patterns. For anyone who keeps getting the convergent default aesthetic back from Claude Design. Output is the prompt — the artifact gets built in Claude Design.

Table of Contents

• Install
• Requirements
• Why this exists
• Scope and complementarity
• Non-goals
• What you can do with it
• Workflow example: from idea to prompt
• Skill surface
• Reference content map
• Per-preset coverage
• Verification
• Compatibility
• Re-research triggers
• Changelog
• License

Install

Add the marketplace once, then install the plugin:

claude plugin marketplace add https://git.fromaitochitta.com/open/ktg-plugin-marketplace.git

In Claude Code:

/plugin install claude-design@ktg-plugin-marketplace

Or enable directly in ~/.claude/settings.json:

{
  "enabledPlugins": {
    "claude-design@ktg-plugin-marketplace": true
  }
}

The skill auto-discovers; no further configuration needed.

Requirements

Claude Code with plugin support, and access to claude.ai/design. No npm packages, no Python, no external tools — see Compatibility for the full matrix.

---

Solo-maintained, fork-and-own. This plugin is a starting point, not a vendor product. Issues are welcome as signals; pull requests are not accepted. See GOVERNANCE.md for the full model and what upstream provides.

AI-generated: all content produced by Claude Code through dialog-driven development. Full disclosure →

---

Why this exists

Claude Design has a strong gravitational pull toward convergent middle-ground output. A one-line prompt like "make me a slide deck for Q1 results" reliably produces what Anthropic's own cookbook for prompting frontend aesthetics names as the failure mode: Inter or Roboto typography, white-to-purple gradients, evenly-spaced cards, cramped layouts that read as AI-generated. The convergence is not random — it is what the model defaults to when prompts are underspecified.

The fix is in the prompt itself, not in the artifact. Anthropic publishes a five-layer prompt scaffold across three primary sources — the Goal / Layout / Content / Audience framework in the Claude Design launch post and Get started article, the DESIGN.md anchor in the design system article, and the AI-slop avoid-list plus cultural-reference anchoring in the aesthetics cookbook. Assembling a prompt that actually uses all five layers, with the right per-preset pattern, in the right order, takes deliberate scaffolding most operators do not do unprompted.

This plugin does the scaffolding interactively. The claude-design-facilitator skill walks the operator through nine phases, surfaces the questions that produce a workable Goal / Layout / Content / Audience answer, anchors on DESIGN.md if one exists or extracts one if not, composes the five layers in the right order, and outputs a copy-paste prompt the operator pastes into claude.ai/design. The artifact gets built in Claude Design; this plugin produces the prompt.

The output is honest about what it is. Every authoritative claim cites an Anthropic primary source inline. Community patterns are labelled and attributed. The frontier-design preset is flagged Experimental rather than dressed up as canonical. The plugin recommends Anthropic's official knowledge-work-plugins/design for everything that happens after the artifact is generated — there is zero command overlap by design.

---

Scope and complementarity

This plugin covers the pre-design and during-design lifecycle for claude.ai/design: idea → intent-preset selection → prompt engineering → copy-paste delivery → iteration coaching → ship-readiness check.

For post-design work — critique, accessibility audit, UX copy review, research synthesis, design-system audit, engineering handoff guidance — install Anthropic's official plugin:

claude plugins add knowledge-work-plugins/design

Anthropic's plugin operates on existing artifacts (Figma URLs, screenshots, copy snippets) and ships six slash-commands: /critique, /accessibility, /ux-copy, /research-synthesis, /design-system, /handoff. There is zero command overlap with this plugin and complementary lifecycle coverage — the two plugins are designed to be installed together. See skills/claude-design-facilitator/references/04-handoff-and-scope.md for the full coverage map.

---

Non-goals

By design, this plugin does not:

• Drive the browser. No automation of claude.ai/design itself; you copy and paste the prompts the skill produces.
• Generate the artifact code. Claude Design is the artifact generator. This plugin produces prompts that go into Claude Design.
• Store artifact history or version artifacts. Claude Design has no version-tree primitive and this plugin does not invent one.
• Cover adjacent Anthropic surfaces. Classic Artifacts at claude.ai, Live Artifacts in Claude Cowork, custom visuals embedded in a chat reply are out of scope — see skills/claude-design-facilitator/references/00-what-claude-design-is-and-isnt.md for the disambiguation reference.
• Duplicate Anthropic's knowledge-work-plugins/design plugin. No /critique, no /accessibility, no /ux-copy, no /research-synthesis, no /design-system, no /handoff. The post-design lane belongs to Anthropic's plugin.

tests/validate-plugin.sh enforces the forbidden-command-name list mechanically.

---

What you can do with it

The skill claude-design-facilitator walks the operator through nine phases. The phases are scoping + grounding (1–4), drafting + delivery (5–6), iteration + ship-readiness (7–8), and an optional critique-iterate loop (9) that runs only when the operator asks for it.

Phase	What happens
1. Disambiguate the surface	Confirm claude.ai/design is the intended surface, not classic Artifacts, Live Artifacts, custom chat visuals, or knowledge-work-plugins/design. Read references/00 when signals are mixed.
2. Name the intent preset	Pick one of eight Claude Design presets: designs, prototypes, slides, one-pagers, wireframes-mockups, pitch-decks, marketing-collateral, frontier-design. The per-preset reference file shapes the prompt pattern. Evidence-grade labels are surfaced.
3. Audience and destination	Capture audience (internal team / external stakeholder / investor / customer) and destination (PDF / PPTX / HTML / Canva / Code-handoff / share-link). Flag PPTX-export traps for pitch-decks early.
4. Anchor on DESIGN.md	Read references/02. If the operator has no DESIGN.md, point at the copy-paste brand-to-DESIGN.md extractor prompt.
5. Draft the prompt	Compose layers 1–5 from references/01: Anthropic's verbatim Goal / Layout / Content / Audience framework + start-simple-layer-complexity + concrete-alternative-spec + propose-options-before-building + AI-slop negative constraints + four design dimensions + four grading criteria + the per-preset pattern.
6. Deliver	Output a single copy-paste-ready fenced markdown code block. Add a one-line caption and three to five expected follow-up turns.
7. Iteration coaching	Read references/03. Coach which surface to use next — Tweak panel (zero-token, surgical), inline comments (component-scoped), or chat (full regen). Session-break heuristics + recovery prompt library when iteration gets stuck.
8. Ship-readiness	Run the export validation checklist. If shipping to engineering, confirm the Design → Code handoff bundle is complete. Recommend installing knowledge-work-plugins/design for downstream critique / accessibility / handoff.
9. Critique-iterate (optional)	Read references/05. Coach the export and the design:critique run (9a), consume the pasted critique as untrusted reference data (9b), synthesize a re-prompt aimed at the one prompt layer that produced the problem (9c), then loop or exit (9d). Opt out at any point by saying skip. This phase routes to Anthropic's plugin; it never reimplements it.

The skill auto-fires on natural-language triggers like "I want to build a dashboard in Claude Design", "help me prompt claude.ai/design", "iterate on my Claude Design artifact". The full trigger list is in skills/claude-design-facilitator/.triggers.txt and tests/test-skill-triggers.sh validates each phrase appears in the skill description.

Explicit invocation works too: the skill registers as the slash command /claude-design-facilitator for when the operator wants to start a clean facilitation session.

---

Workflow example: from idea to prompt

A realistic session against the slides preset — Q1 results deck for an internal engineering all-hands.

Operator: "I want to build a Q1 results slide deck for the engineering team in Claude Design."

The skill auto-fires (the phrase matches .triggers.txt). It walks the phases:

Phase 1 — Disambiguate the surface. The skill confirms claude.ai/design is the intended surface, not classic Artifacts or Live Artifacts. The operator confirms.

Phase 2 — Name the intent preset. Slide deck → slides preset. The skill notes this is one of three Anthropic-documented presets (evidence-grade label surfaced from .coverage.md). It opens references/presets/slides.md and surfaces the five canonical Anthropic patterns (Q1 results, executive roadmap, customer-prep briefing, partner proposal, all-hands announcement). The operator picks pattern 1.

Phase 3 — Audience and destination. Internal engineering team; deck stays in HTML preview during the meeting, optional PPTX export to share with adjacent leads afterward. The skill flags the PPTX-export trap from references/presets/slides.md section (e): fonts substitute, master slides drop, charts may flatten to images. If a brand-compliant PPTX template exists, upload it to Claude Design as a project asset before prompting — Claude reads the slide master, layouts, fonts, and colour scheme and respects them (PowerPoint-mode article).

Phase 4 — Anchor on DESIGN.md. The operator has no DESIGN.md yet. The skill points at the brand-to-DESIGN.md extractor prompt in references/02-design-md.md: paste a brand-guidelines URL or PDF into Claude.ai, get back a DESIGN.md filling the nine canonical sections (typography, colour, spacing, layout primitives, motion, voice, imagery, density, accessibility). The operator runs the extractor against the company's brand site, gets a DESIGN.md, drops it into the Claude Design project assets.

Phase 5 — Draft the prompt. The skill composes the five layers from references/01-prompt-fundamentals.md: Goal / Layout / Content / Audience (Layer 1) → start-simple-layer-complexity (Layer 1.5) → concrete-alternative-spec house style (Layer 2a) → AI-slop negative constraints (Layer 3), composed with the per-preset pattern from presets/slides.md. The skill produces a single copy-paste fenced block. The operator inspects it, optionally edits the Goal sentence, then proceeds.

Phase 6 — Deliver. The skill outputs the prompt block. The structure of a realistic delivery for this scenario looks like:

**Goal:** Communicate Q1 engineering results to the all-hands —
where we started, what changed, what we shipped, what we learned,
what's next for Q2. The deck should land as confident but honest:
real numbers, named risks, no victory-lap framing. Audience is
~80 engineers across three teams. Density should be high enough
to skim later, low enough to follow live.

**Layout:** 10–12 slides, slide-by-slide:
1. Title + Q1 in one sentence
2. The three things we shipped (one per row, screenshot + metric)
3. The two things that slipped (named honestly, with the why)
4. Hiring update (count + retention)
5. Reliability (incident count, MTTR trend)
6. Customer signal (NPS + 2 verbatim quotes)
7. Engineering health (PR throughput, review latency)
8. The big bet for Q2 (one slide, named)
9. Risks for Q2 (3 bullets, ranked)
10. Asks from the all-hands (1–3 specific asks)
11. Q&A placeholder

**Content:** Use the metrics in DESIGN.md's `tone` section — direct,
specific, no marketing voice. Numbers are placeholder; I'll edit
before the meeting.

**Audience:** Internal engineering all-hands, 80 people, ICs through
EM/Director level. They want to know: did we ship what we said, what
broke, what's next, can I help.

**Avoid:** Inter or Roboto, white-to-purple gradients, evenly-spaced
generic card layouts, "exciting Q1!" framing, congratulatory tone,
stock-photo gradients, generic icon library defaults.

**Anchor:** Match the DESIGN.md uploaded as a project asset. If our
brand voice reads as understated technical, push the deck that way —
not the convergent SaaS-marketing deck aesthetic.

**Reference:** Treat this as the Q1 results pattern from
https://claude.com/resources/tutorials/using-claude-design-for-presentations-and-slide-decks
(pattern 1), with the Layout above overriding the tutorial's slide count.

That block is what gets pasted into claude.ai/design. The skill also surfaces three to five expected follow-up turns (e.g., "the headline slide is too marketing, push it more technical", "slide 5 reliability — show the MTTR trend as a sparkline, not a bar chart") so the operator knows what iteration looks like before starting.

Phase 7 — Iteration coaching. Once Claude Design produces the first version, the skill points the operator at the Tweak → Comment → Chat cascade in references/03-iteration-and-session.md: Tweak panel for surgical zero-token edits (spacing, font size, colour), inline comments for component-scoped changes (rewrite slide 5), full chat regeneration as a last resort. Plus the session-break heuristic (after 4 substantive screens, start a fresh session with a verbal save-pattern carrying state forward) and the recovery prompt library when iteration gets stuck.

Phase 8 — Ship-readiness. Before the all-hands, the skill runs the export validation checklist for the chosen destination (HTML preview → keep in Claude Design; PPTX → check fonts and master, charts may flatten). If the deck is being handed off to engineering for any reason, it recommends installing knowledge-work-plugins/design for design:critique, design:accessibility, and design:handoff — the post-design lane.

Phase 9 — Critique-iterate (optional). If the deck comes back with problems that surface tweaks cannot fix — a slide that does not earn its place, hierarchy that fights the narrative — the skill offers the critique-iterate loop from references/05-critique-iterate.md. It coaches the export, hands over design:critique @<file>, reads the pasted critique by judgment rather than by parser, and drafts a re-prompt aimed at the single prompt layer responsible — for slides, deck-wide drift usually means the house-style spec (Layer 2a) was too loose, not that any one slide was laid out wrong. The operator confirms the mapping before anything is handed back. Saying skip at any point ends the phase.

The full output of the session is a single fenced markdown block (Phase 6) plus a short follow-up-turns list and an iteration-coaching pointer. That is the entire user-facing deliverable.

---

Skill surface

Skill	Triggers	Output
claude-design-facilitator	12 natural-language phrases (full list in .triggers.txt); also explicit /claude-design-facilitator slash command	A copy-paste-ready Claude Design prompt block composed from the five-layer stack and the per-preset pattern, with follow-up-turn expectations

No commands, no agents, no hooks, no MCP servers at v0.1. The single skill is the entire user-facing surface.

---

Reference content map

The plugin ships 14 reference files in skills/claude-design-facilitator/references/:

Foundation references (6):

• 00-what-claude-design-is-and-isnt.md — Surface disambiguation against Artifacts, Live Artifacts, custom chat visuals, and Anthropic's knowledge-work-plugins/design.
• 01-prompt-fundamentals.md — The five-layer prompt stack: GLCA framework + start-simple-layer-complexity + concrete-alternative-spec + propose-options + AI-slop negative constraints + four design dimensions + four grading criteria. Anchored on four Anthropic primary sources.
• 02-design-md.md — DESIGN.md 9-section canonical structure + brand-to-DESIGN.md extractor prompt + failure modes.
• 03-iteration-and-session.md — Tweak / Comment / Chat cascade, session economics, 4-screen inflection, recovery prompt library (break-default-aesthetic, fix-the-system, edit-previous-message, 3-failed-comment escalation, model downshift, verbal save-pattern).
• 04-handoff-and-scope.md — Design → Code one-way handoff, bundle contents, lifecycle-stage coverage map vs Anthropic's knowledge-work-plugins/design, downstream tool recommendation.
• 05-critique-iterate.md — Phase 9 shared mechanics: operator-as-bridge handoff rationale, the design:critique preconditions, the untrusted-input rule for pasted critique output, and the judgment-based mapping from critique findings to prompt layers. Each per-preset file carries its own thin Critique-iterate hint block on top of this.

Per-preset references (8):

• presets/designs.md — Anthropic-documented + community-validated
• presets/prototypes.md — Anthropic-documented + community-validated
• presets/slides.md — Anthropic-documented + community-validated
• presets/one-pagers.md — Community-only
• presets/wireframes-mockups.md — Community-only
• presets/pitch-decks.md — Community-only (with explicit PPTX-export caveat)
• presets/marketing-collateral.md — Community-only
• presets/frontier-design.md — Experimental — no validated practitioner pattern as of 2026-05-16

---

Per-preset coverage

The canonical coverage manifest is .coverage.md. Below mirrors that file.

Preset	Evidence grade	Anthropic anchor
designs	Anthropic-documented + community-validated	launch post
prototypes	Anthropic-documented + community-validated	prototypes tutorial
slides	Anthropic-documented + community-validated	slides tutorial
one-pagers	Community-only	launch post
wireframes-mockups	Community-only	launch post
pitch-decks	Community-only (with PPTX-export caveat)	launch post
marketing-collateral	Community-only	launch post
frontier-design	Experimental — no validated practitioner pattern	launch post

When Anthropic publishes per-preset guidance for a Community-only or Experimental preset, .coverage.md and the affected preset file refresh — re-research triggers are documented inline.

---

Verification

bash verify.sh

Nothing runs this automatically — there is no CI in this repository. The command above is the entire suite, and it runs from a clean clone with no setup step.

Runs five test scripts under tests/ in dependency order:

Script	Verifies
validate-plugin.sh	plugin.json + SKILL.md frontmatter + LICENSE + GOVERNANCE.md + README.md + CLAUDE.md + .coverage.md presence; forbidden-command-name scope-fence check; operator-private-context grep; Norwegian-leakage advisory
test-skill-triggers.sh	SKILL.md description >=400 chars; every phrase in .triggers.txt appears in SKILL.md
test-sc2-artifact-coverage.sh	Each preset in .coverage.md has >=1 file hit in plugin content
test-sc3-citations.sh	No unsourced-attribution placeholders (citation-stub markers, verification-flag markers, vague second-hand phrasing); each Authoritative-claims file has >=1 Anthropic-domain URL. The script enforces the exact patterns it bans — see the script source for the regex.
test-sc1-dogfood-log.sh	Format-check the operator dogfood log in REMEMBER.md (gitignored) — 5 fields well-formed

Flags:

• --strict — pass-through to test-sc1-dogfood-log.sh. Without --strict, missing dogfood block is advisory. With --strict, it is the release gate.
• --quick — skip test-skill-triggers.sh for fast incremental runs.

Exit codes: 0 = all pass; non-zero = at least one sub-test failed.

---

Compatibility

Requirement	Version
Claude Code	Recent versions with plugin support
Anthropic surface	claude.ai/design (Labs research preview launched 2026-04-17)
Platform	macOS, Linux, Windows
Network	None for the skill itself; the artifact-generation lives in claude.ai/design
Dependencies	None — no npm packages, no Python, no external tools. Bash 3.2 compatible for test scripts.

---

Re-research triggers

The reference tree carries Anthropic citations that may decay. Re-research is triggered by:

• Anthropic publishing per-preset guidance for a Community-only or Experimental preset
• Anthropic announcing material changes to the Goal / Layout / Content / Audience framework, the AI-slop avoid-list, or the design grading criteria
• Anthropic adding or removing an intent preset from the launch enumeration
• A first verified frontier-design practitioner artifact shipping publicly
• Anthropic's knowledge-work-plugins/design plugin adding or removing slash-commands (scope-fence implications)
• Labs → GA URL rename for claude.ai/design

When a trigger fires, run bash verify.sh --strict after the update to confirm SC2 and SC3 still pass.

---

Changelog

Full release history: CHANGELOG.md. The plugin follows Semantic Versioning.

v0.1.0 — 2026-05-17. Initial public release. Single skill (claude-design-facilitator) with eight-phase facilitation flow, 12 natural-language trigger phrases, 13 reference files (5 foundation + 8 per-preset with evidence-grade labels), .coverage.md preset manifest plus Authoritative-claims registry, five verification scripts under tests/ enforcing structural integrity / scope fence / skill description quality / per-preset coverage / Anthropic-domain citation discipline / operator dogfood log format, top-level verify.sh roll-up with --strict and --quick flags, MIT license, GOVERNANCE.md fork-and-own model.

The path from v0.1 to v1.0 is dogfood-driven. The v1.0 readiness criteria are multi-preset breadth, auto-fire validation in real natural-language requests, and two consecutive dogfood sessions with zero critical patches. The dogfood log itself is operator-local and never published.

---

License

MIT. Fork it, modify it, ship your own version under your own name.