Overview
A three-phase process skill that upgrades AI from a "statistically mediocre default designer" to a "world-class designer." Theoretical benchmark: "How to Turn Your AI into a World-Class Designer" (Anshu Chimala's AI Double Diamond model, former 12-year lead of Apple's AI prototyping team; repository archive path docs/reference/articles/how-to-turn-your-ai-into-a-world-class-designer.md).
Golden Formula：
$$\text{World-Class AI Design} = \text{External Random Seed} + \text{Independent Critic Loop} + \text{Multimodal Enhancement} + \text{Ruthless Subtraction}$$
This file is therouting layer: it only declares trigger conditions, iron rules, and phase pointers; for phase details see references/。
When to Use / When Not to Use
Use(if any is met):
- The user requests the design/redesign of interface artifacts such as landing pages, web apps, mobile interfaces, game UI, dashboards, etc.;
- The user explicitly requests "has a sense of design / has taste / doesn't look AI-generated / a bit more unique / a premium feel / give an art direction";
- An existing interface needs "design critique / visual vibe-check / polish";
- Frontend construction needs creative direction beyond purely functional layout.
Don't use: purely functional changes with no visual requirements, backend/data logic tasks, or when the user explicitly requests "just use the default template."
Iron Rules (Invariants)
- Randomness must come from outside the model: before starting any design, first run the seed string procedure (
scripts/generate-seed.sh); verbal instructions like "make it a bit random / a bit more unique" don't count—the model will only predict tokens that "sound random." - Critic separation discipline: the Critic intervenes each time with afresh context + fixed template + seeing only the artifact—by default it receives only the current artifact (screenshot, downgrade if none);in blind-comparison mode, additionally attach a de-identified, randomly left/right previous version. In no mode may it carry implementation code, iteration records, or the full text of previous critiques.
- Three-signal convergence gate: D2 completion must simultaneously satisfy — A Structural Checklist(7 AI Tells + hierarchy/spacing/state/copy audit, item-by-item verifiable) + B Blind-Comparison Improvement(previous version vs current version forced choice: which is better + winning dimension,no scores output）+ C Human Sign-Off. For structural checklist extensions see critic-loop-protocol.md Gate A；Artifact-Level Acceptance(responsive and mobile, interaction state matrix, accessible keyboard paths, browser coverage, performance budget) with the
frontend-qa-gateskill as the primary path, whose acceptance baseline is an inline minimal copy, with the source version in that skill's acceptance-matrix (Option A: Gate A is the single source of truth for the acceptance baseline). Absolute scores (e.g., x/10)are for telemetry only, and must never be used as a pass criterion;gate not met ≠ delivery complete, report blocked truthfully pending adjudication. - Budget has a cap: first run 1–2 rounds to verify convergence; the full-process iteration cap is 5 rounds, and when exceeded, trip the circuit breaker and report to the user.
- Subtraction first: The Deliver phase removes first, then adds—any element that cannot be removed without losing the value explanation (decorative gradients/extraneous cards/decorative labels) is removed outright; the AI Tells audit must go through the checklist item by item (
references/ai-tells-audit.md）。 - Copy is placeholder: AI-generated copy is treated uniformly as typographic Lorem ipsum and must be rewritten before delivery into concrete, restrained, conversational human language (blacklist glossary see
references/ai-tells-audit.md）。 - Capability gating: Screenshot/image/video capabilities are enabled according to the host's actual availability; if screenshots are unavailable, use degraded review—never pretend to have seen a screenshot.
- Humans are creative directors: Direction choices, taste benchmarks, and trade-off decisions belong to the user; the agent is responsible for providing speed, breadth, and execution.
- References are anchors, not copies: moodboards/reference images are only used to calibrate the quality baseline; copying them is prohibited.
- Facts are written down immediately: Direction decisions and deletion lists are written into the design brief to avoid context rot.
Phase Routing
| Phase | Goal | Core techniques | Artifacts / Pointers |
|---|---|---|---|
| D1 Discover | Break down statistical mediocrity and explore divergently | Seed strings inject randomness + ambitious cross-boundary Prompt | Design brief (templates/design-brief-template.md); detailed rules see discover-phase.md |
| D2 Define | Inject taste and personality, approaching world-class execution standards | Independent Critic closed loop (blind A/B comparison + reference ranking) + (optional) image/video enhancement + artifact self-check embedded in each round(state matrix and responsive spot checks, minute-level, no Critic dispatched) | Review ledger + structure checklist + self-check records; detailed rules see critic-loop-protocol.md、multimodal-enrichment.md |
| D3 Deliver | Converge and subtract, eliminate AI traces | Ruthless subtraction + AI Tells audit + copy rewrite | AI Tells audit checklist + deletion records; detailed rules see ai-tells-audit.md |
Phase discipline: D1 may not enter D2 without producing a multi-direction brief; D2 may not enter D3 unless it simultaneously satisfies "all structure checklist items pass + blind comparison shows improvement (or is judged converged) + human sign-off"; D3 may not claim delivery unless the audit goes through the checklist item by item.
Quick Checklist
D1 ┌ generate-seed.sh 生成种子 → 提取子模式 → 定义创意方向（配色/布局/字体/质感）
├ 多方向并行：产出 2~4 个 Brief（必要时派发子智能体分别执行）
└ 与用户对齐方向，选定 1 个（兼修时取 2 个）
D2 ┌ 实现初版 → 捕获产物（截图 / 降级文本）→ 附版本 nonce
├ 派发 Critic（全新上下文 + 固定模板 + 设计意图与已决原则；必须回执 nonce）
├ 盲比：上一版 vs 当前版强制选择（不输出分数）→ 有改进则记账续跑
├ 结构清单逐项核验（7 AI Tells + 层级/间距/状态/文案 + 状态矩阵 / 视口抽查 / 无障碍基线 / 证据三态）→ 全过才谈放行
└ 人类签收 → 进入 D3；分数仅记遥测（可选增强见 multimodal-enrichment.md）
D3 ┌ 逐项过 AI Tells 审计清单（7 反模式 + 减法 + 文案黑名单）
├ 删除无价值元素 → 重写全部文案 → 人工复核方向一致
├ 交付说明：方向来源、盲比结论与签收、删除了什么，并标注证据三态（已实现 / 已运行验证 / 未验证）
└ 产物级完整验收（响应式 / 状态 / 无障碍 / 浏览器 / 性能）由 `frontend-qa-gate` 独立执行，本技能不替代该验收
Fallbacks & Budgets
| Scenario | Handling |
|---|---|
| Host has no screenshot capability | Use degraded review: provide the Critic with a structured artifact summary (semantic tree + visual parameters + key sections), see critic-loop-protocol.md §8 |
| Host has no image/video generation | Skip the enhancement layer and replace it with code-level texture (shaders/CSS textures), see multimodal-enrichment.md §4 Gating Matrix |
| Blind comparison shows no improvement for 2 consecutive rounds | Circuit breaker: stop reskin-style iteration, show the user the blind comparison records and attempted fixes, and request a direction decision |
| Discrimination self-check fails (inputs substantively different, outputs word-for-word identical) | Judge the Critic as failed: stop the loop immediately and escalate to a human; do not consume further rounds |
| 2 consecutive rounds of review are invalidated (receipt mismatch / blank artifacts) | Stop the loop: investigate the artifact pipeline (screenshot and receipt chain) and escalate to a human; invalidations do not count against the budget, but consecutive invalidations are a failure signal |
| Critic suggestions conflict with already-decided principles | Principles first + record in the ledger, hand over to human adjudication; silent reversal is prohibited (anti-ping-pong) |
| The review ledger has reached the 5-round limit | Circuit breaker: report the current best state and the gap list, ask the user to decide whether to continue |
| Gate not met (checklist not fully passed / no improvement / not signed off) | Truthfully report blocked pending adjudication,must not claim delivery is complete |
Artifact-level acceptance not passed (frontend-qa-gate has failing or unverified assertions) |
Deemed delivery incomplete: route through that skill's reflow routing before discussing delivery |
Auxiliary scripts (optional)
scripts/generate-seed.sh Inject external randomness into the Discover phase (String Seed of Thought):
SKILL_DIR="<taste-driven-designer 技能实际所在目录>"
bash "$SKILL_DIR/scripts/generate-seed.sh" # 64 位真随机种子
bash "$SKILL_DIR/scripts/generate-seed.sh" -l 128 -c 3 # 3 条种子并行探索
bash "$SKILL_DIR/scripts/generate-seed.sh" -s team-2026 -l 48 # 团队可复现
See the complete construction steps and inspiration extraction Prompt in seed-string-procedure.md。
Reference navigation
- discover-phase.md — D1 seed string protocol and the three-step ambition Prompt forging method
- critic-loop-protocol.md — D2 three-signal gate, blind comparison and ranking protocol, version receipt, discriminative power self-check, and review ledger
- ai-tells-audit.md — D3 list of 7 major anti-patterns, subtraction rules, and copy blacklist
- multimodal-enrichment.md — optional enhancement layer and capability gating matrix
- Template suite:
templates/design brief, Critic fixed prompt, seed construction three-step template