AI-Assisted UX · Delivery Phase (Using AI for Layout Design · Layout-to-Code Figma · Design Review · Showcasing AI Value + Will's Custom UX Metric Teardown Skill)

Will's UX workflow is divided into 6 phases (Discover / Define / Ideate / Experience Design / Deliver …), and this article covers the final phase Deliver.

First, understand what "deliver" means in this phase: It is not the kind of delivery where "you hand off Figma specs to developers" (that is covered in Google UX Certificate / Testing · Iteration · Design Review · Delivery §4 handoff). Will's Deliver here refers to distilling a completed project into a "portfolio page + interview talking points" to deliver—to recruiters/reviewers. Therefore, the common theme of all AI techniques in this phase is two things:(a) Rearranging information into a layout/visual that "tells a story and is instantly understandable",(b) Clearly explaining "how the work was split between me and AI". Understanding this theme will prevent each technique below from feeling scattered.

Significance for PMs: 90% of these techniques can be directly applied to your daily work—feeding a bad PPT page to AI to rearrange into several versions, converting competitor screenshots into editable Figma files, using a skill to conduct a "metric-driven review" of your own product interface, and writing project retrospectives into reporting talking points structured as "problem → judgment → solution → result".It is essentially a prompt library for "using AI for information visualization + value articulation", regardless of whether you work in UX or not.

1. Layout Generation: From One Old Layout to Multiple New Options

Task: You have a mediocre PPT page / magazine layout / infographic, and you want to use AI to rearrange it into several very different new versions, then select one for final output. Will breaks this down into 4 steps,the key is "adjust content first, then layout, and finally visuals"—keep these three things strictly separate, do not ask AI to do all of them in one sentence.

Step 1 · 1:1 Extract Text and Structure from the Old Layout (No missing words, no added words)

First, have AI extract all text from the old image verbatim and clarify the hierarchy. This step only transfers, does not create.

Reusable prompt skeleton: Task: Extract and restructure all text content in this image. Strict rules: Do not add any new words / do not delete any existing words / do not rewrite or restate meanings / only reorganization and translation are allowed / all original information must be fully retained. Process: ① Extract all visible text → ② Identify hierarchy (title / section heading / supporting points / label annotations) → ③ Infer relationships (group related content, parent-child structure, detect step sequences). Output: Clear hierarchical indented structure, maintain original information density, no explanations, no comments, no structures beyond the implicit structure of the original text.

Why "only transfer, no creation": The output of this step is the "content foundation" for all subsequent steps. If AI has already secretly changed wording/deleted key points in the foundation, no matter how beautiful the later layout is, it will be wrong. Lock down the content first, then talk about form.
Will's real-world testing: For the final HTML output, Claude produces better results than ChatGPT's image2 (when restoring layouts).

Step 2 · Content Direction Exploration (Only adjust narrative, no touching layout/visuals allowed)

Feed the plain text from Step 1 back in, and have AI come up with 3-4 sufficiently different restructuring directions—the differences must be reflected in "fundamentally different narrative frameworks", not differences in level of detail or wording.

Reusable prompt skeleton: Content direction exploration (only do this step in this round, no touching layout/visuals allowed). ① First, state in one sentence: what is the "one thing" the original text really wants to convey. ② Judge where the key point of this content lies (too much data? weak conclusion? logical gaps? graphics not clarifying hierarchy?), and design 3-4 sufficiently different directions based on this—directions must grow out of the content itself, not apply fixed templates. ③ Available editorial methods (use as needed, not all required): add missing context/causality, cut repetition and secondary content, rewrite titles and narrative frameworks, distill core data to the extreme, switch to more understandable graphic language, use hierarchy/scaling to create visual focus. ④ Output for each direction: one-sentence strategy name + 150-200 word explanation (why it suits this content, what to keep/cut/add, why viewers will understand it instantly) + preview of the restructured actual copy (new title/core numbers/descriptions/insights). ⑤ At the end, ask me which one to choose (combinations allowed), do not make any layout descriptions.

Why two separate steps: People tend to immediately get stuck on "put it on the left or right, what color to use". This step by Will forces you to first hash out "what this page is really trying to say, and what narrative framework to use"—once the narrative is set, the layout makes sense. This is the same logic as writing a PRD where you first define "what problem to solve" before talking about features.
Counterexample: Directly saying "help me make this PPT page look nicer" → AI gives you the same layout with a different font, because you didn't force it to rethink the narrative.

Step 3-4 · Select direction → Output image

Select a direction in the conversation (you can add extra requirements, such as "try a funnel shape for data visualization"), have AI produce an image generation prompt for that solution; then use Create image + 16:9 to generate the new layout image.

2. Optimize Layout / Migrate Visual Style (Only change the skin, do not move the bones)

Task: You already have a layout solution, and you want to "switch to a different visual style"—but information structure, text, and element positions must not be changed at all, only migrate colors/fonts/spacing/shape language. Two use cases:

Use Case A · Provide a visual reference image and migrate its style over

Reusable prompt skeleton (style transfer): Task: Redesign the current layout by applying the visual style of the reference image. Input: Image 1 = current layout (to be redesigned) / Image 2 = style reference image. Strictly retain (must not change): Layout structure / information hierarchy / all content and text / element positions and spacing logic / content grouping and relationships. Style transfer (applied from Image 2): Color system (primary/secondary/background) / font system (style/weight/hierarchy) / spacing rhythm / shape language (rounded corners/geometry) / shadow hierarchy and sense of space / border divider styles / background style (solid color/gradient/texture) / overall composition style. Constraints: Do not add content / do not delete content / do not change layout structure / do not add decorative noise / do not output watermarks / do not add extra text. Output: High-fidelity visual redesign, fully retaining original structure and content, comprehensively applying the new visuals.

Applicable boundaries (Will's reminder): This set of techniques is only suitable for "layout optimization", not for UI design(UI has interaction states/component rules, which cannot be solved by just changing skins). If font sizes are too small, you can manually adjust them later.

Use Case B · No reference image, directly adjust to the restrained feel of "polished by a senior editorial designer"

When you just feel it's "too crowded, too noisy, not premium enough", you can use a set of specific, visible polish instructions (don't vaguely say "make it more premium"):

Reusable prompt skeleton (restrained polish): Please boldly and visibly apply the following adjustments: Reduce overall font size by 1-2 levels (smaller text = more negative space = more confident, more editorial feel) / tighten font size hierarchy ratio (titles closer to body text, restrained hierarchy rather than flashy contrast) / significantly increase whitespace between blocks (at first glance even slightly "too loose") / increase container padding to let the layout breathe / remove all non-structurally necessary dividers (replace with whitespace) / unify remaining borders to a thin 0.5px / simplify background to low-contrast micro-layering / set title line height to 1.1-1.15 / body text line height to 1.6 / add slight negative letter-spacing to titles larger than 24px. The final result should be quieter, more spacious, and more restrained in typography than the original—as if polished by a senior editorial designer.

Why give "visible specific instructions": AI has no consensus on adjectives like "premium/refined", but it responds precisely to "reduce font size by 1-2 levels, line height 1.6, border 0.5px".Translating aesthetics into executable parameters is the key to getting AI to consistently produce output that "looks expensive".

3. Layout-to-Code & Figma Mockup (Turn an image into an editable source file)

Task: You see a nice competitor page/magazine page and want to turn it into an editable Figma file or HTML(instead of just a screenshot). Will's workflow:Single-page image → restoration prompt → HTML → reverse-engineer to Figma.

Step 1 · Have AI write a "restoration mode" prompt from the single-page image

The output of this step is a section of English restoration instructions for Figma Make, the core of which is "prohibit AI from free-styling, describe each element precisely".

Reusable prompt skeleton (generate restoration prompt): Analyze this layout screenshot and generate a restoration prompt that can be directly pasted into Figma Make. Requirements: ① Output in English (Figma Make is more accurate with English) ② First identify the format specification: judge the aspect ratio → output as a scrollable long page, with each layout stacked vertically as an independent Section, each Section maintaining the original ratio, width 100vw ③ Text must be fully retained: copy original text verbatim, no changes, no omissions, no placeholders; describe the font size hierarchy (Display/Heading/Subheading/Body/Caption/Label), weight, hex color, alignment, and position of each piece of text ④ Precisely describe container forms: rounded corner cards? color block backgrounds? thin line borders? borderless whitespace? And indicate background color/border color/border thickness/rounded corners/shadows ⑤ Information Graphic restoration: arrows (direction/line type/color/style), connecting lines, icons (shape + color, SVG or Unicode approximation), charts (structure/labels/color scheme); leave blank placeholder frames marked with aspect ratio for background photographic images ⑥ Extract color values: page background/container background/borders/primary text/secondary text/brand accent color ⑦ Spacing system: Section padding, gap/margin relationships between elements, padding from text to container edges ⑧ Add a restoration mode declaration at the beginning: prohibit free-styling/redesign, prohibit replacing container forms, prohibit changing text—make it clear this is layout restoration, not UI design Output: Directly provide the English prompt body, chunked by Section, one paragraph per page.

Step 2-3 · Generate HTML first, then reverse-engineer it into Figma

Step 2: Give the prompt above to Claude,and generate HTML first.
Step 3: Two ways to import into Figma Make——Method 1 Feed the prompt from Step 1 to Figma Make to produce an editable Figma file;Method 2 (more accurate) Feed the HTML from Step 2 to Figma Make to reverse-engineer an editable Figma source file.
Why go the extra mile with HTML: If you go directly "image → Figma", the AI tends to hallucinate details;First finalize it as HTML (structured, verifiable), then let Figma Make process the HTML, which is equivalent to giving it a precise structural contract, resulting in higher fidelity restoration.
Limitations: It can only restore the layout portion; distinctive photos/illustrations cannot be restored (placeholder frames will be left instead). It is recommended that all steps be performed within the same Claude conversation to maintain contextual continuity.

IV. Infographic Semantic Extraction + Visual Redraw (Strip Logic → Replace Skeleton → Apply Style)

Task: You have an infographic (flow/relationship diagram) and want to redraw it in a completely different visual style while preserving its logic. The difficulty: if you directly "swap styles", the AI will copy the original image's arrow styles/shapes along with it, resulting in an incomplete style change. Will's solution is to completely separate "logic" from "visuals" first.

Step 1 · Extract semantics only, no visual terms allowed

Reusable prompt skeleton (semantic extraction): Analyze this infographic and distill its semantic structure for use in redrawing a visually completely different version. Extract: ① Core entities: What "roles/nodes" are in the diagram? Only describe what they represent (e.g., "a user", "a platform"),do not describe shape/icon style/size/color. ② Relationships and flow: Who points to whom? Is it one-way or two-way? What is the nature of the relationship (transmission/blocking/diversion/convergence)? Only describe logic, do not describe arrow styles. ③ Overall narrative: The core conclusion this diagram intends to convey, in one sentence. ④ Layout orientation: Approximate positions of entities (left/center/right, top/bottom) as composition reference, no dimensions or spacing involved. Output: Structured text, do not output any visual descriptors (no colors/shapes/line styles/fonts).

Step 2 · Have the AI propose 5 vastly different styling directions

Reusable prompt skeleton (styling divergence): Based on the following semantic structure [paste Step 1 result], generate 5 completely different styling creative directions. Requirements: Each direction gets 2-4 coherent descriptive sentences (like a director explaining a concept to an art director), focusing on overall styling language, layout composition logic, and spatial relationships between entities; Prioritize abstract structures/geometric language/system diagrams/mechanical metaphors/information architecture/minimalist layouts; cultural or natural metaphors may be referenced but stay at the styling level (no people or figurative narratives); Differences between directions must be sufficiently large (not minor tweaks of the same style); do not describe colors/line widths/fonts; give each direction a 2-4 character short name.

Step 3-4 · Produce line art → Apply visual style

Step 3 Produce line art: Pick one direction and produce black-and-white line art to a uniform specification——pure white background / 16:9 aspect ratio / all lines uniformly 2px black single strokes / no filled color blocks, convey information solely through line outlines and spatial relationships.
Step 4 Apply style:Redraw the shapes from Image 1 using the visual style and feel of Image 2(Redraw the form of Image 1 using the visual style of Image 2). Optional add-ons:Do not include any text(no text) or Use small-size text, express everything in English(small font size, all English).
Why strip everything down to pure logic first: This is the most critical trick in the entire technique.Degrade the diagram into a "visionless pure logic skeleton", then grow new visuals from it, so you can truly change styles instead of just reskinning; otherwise the AI will always be dragged along by the visual inertia of the original image. The same logic can also be seen in AI-Assisted UX · Experience/UI Design Phase (Prompts and Workflows for Using AI to Generate Aesthetic UI · Design Systems · Color Schemes · Interaction States · Microcopy · UI-to-Code Figma · Responsive Design) under the "structure first, visuals later" approach.

V. Metric-Driven Solution Review Using the will-ux-metric-teardown Skill

This section separately introduces a custom Claude Skill built by Will(attached in the source materials as a.skill file). It takes the act of "picking apart issues on an interface"and forcibly ties it to business metrics.

What it does

Input one or a set of interface screenshots (multiple screenshots = consecutive steps of the same flow), output a metric-driven, priority-tiered, element-specific, improvement-plan-included UX review report (teardown). Its core belief in one sentence:"The priority of an experience issue is determined by its impact on business metrics, not by a designer's aesthetic OCD." A minor layout flaw and a default option error that causes a 30% conversion drop should not be treated as equal priority.

How it runs (three-stage, strictly in order)

Reverse-engineer product context: Before picking out any issues, first reverse-engineer from visible evidence on the interface ① Business model (Freemium/transaction/subscription/advertising/B2B, must cite interface evidence) ② Business goals + specific KPIs (activation rate/first-order conversion/D7 retention/ARPU, provide 3-5) ③ Experience goals + UX metrics (task success rate/Time-to-Value/error rate, provide 3-5) ④ Core user groups (2-4) ⑤ User primary tasks (first-person JTBD).
Lock in the North Star Metric: From the business goals,pick the single metric most critical to the business right now (aligned with product stage: early stage focuses on activation/acquisition, mature stage focuses on retention/monetization), and clearly state "why this one". All Tier 1 issues later will anchor to it.
Find issues by three priority tiers(Exhaust one tier completely before moving to the next, no quotas, no padding to hit numbers):
- 🔴 Tier 1 — Directly drags down the North Star Metric: Each issue must map out a causal chain 具体 UI 元素 → 用户行为 → 北极星如何下降, estimate magnitude if evidence supports it. Ironclad downgrade rule:If you cannot state which KPI it harms, immediately demote it to Tier 2/3. Tier 1 issues are almost never color or font size problems, but rather defaults/information ordering/missing guidance/problem framing/hidden critical information/missing trust signals at decision moments.
- 🟠 Tier 2 — Affects other metrics: Name which secondary metric it affects.
- ⚪ Tier 3 — Pure design level: Usability/visual hierarchy/accessibility/consistency/microcopy, no direct metric association. If a tier has no issues,leave it blank, do not force entries.
- Numbering is continuous across tiers(01, 02, 03…), so readers can see at a glance which issues are more important. In the report this matches the source material note "Tier 1-x, the smaller the number the higher the priority and impact".

What makes it good / How to use it (Why)

Why reverse-engineer the business model before picking issues: Most "design reviews" are aesthetic reviews, vague and wishy-washy about what's good or bad. This skill forces you to first lay out "what metrics this business/this page is actually targeting", so every subsequent issue has a priority anchor——that is the fundamental difference between it and ordinary "nitpicking".
How to use: Install this .skill into Claude (or directly use its workflow as a prompt), then drop in interface screenshots;if you know the product background (business model/what metrics are being optimized/constraints), provide that as well——the skill explicitly states "users know their own product better than pixels do"; if background is available, use it to anchor; if background conflicts with the interface, flag the conflict itself as an issue.
Complementary use: This is a solo tool for "using AI as a senior reviewer"; how to run a design review meeting(3 roles / describe problems without giving solutions / turn feedback into actions) is an interpersonal collaboration issue, see Google UX Certificate / Testing · Iteration · Design Review · Delivery §III. The two complement each other: the skill helps you before the meeting break down the interface into a prioritized issue list, then you run the interpersonal process during the meeting.

VI. Demonstrating AI Capabilities and Value in Projects

Task: After finishing a project, you need to clearly explain "which work I handed off to AI, and which were my judgments"in your portfolio/interviews——you can neither pretend you did everything yourself, nor make people think "so AI did all that, what did you do?" Will uses a prompt to automatically split the project into three deliverables for each UX stage.

Reusable prompt skeleton (AI Value Expression Coach): You are a UX career expression coach, helping designers demonstrate AI collaboration capabilities in portfolios/interviews. [Step 1] I will provide project content (text/stage records/PDF), read it in full, do not ask follow-up questions, analyze directly. [Step 2] Automatically identify which UX stages the project covers, and process each one. [Step 3] Output three sections for each stage: ① What AI was involved in: Which actions in this stage were handed to AI? What type of AI tool was used to do what? (Integrate tool names into the description, do not list them separately) ②Interview verbal script: 1-2 natural conversational sentences that reflect the division of labor of "AI executes / I judge", sounding like a real person talking rather than a written report. ③Written evidence (what to put on this PPT slide):Only output evidence directly related to AI use——format (screenshot/comparison table/annotated diagram/flowchart/quoted snippet) + specific content (what to screenshot/annotate/compare, be specific) + one layout tip (how to place it so people can see at a glance "this is human-AI collaboration"). Judgment criterion (key): Would this evidence still need to exist if AI were removed? If yes → it does not meet requirements, do not output it; if no → keep it. If a stage had no AI involvement, directly write "No AI-related evidence slide needed for this stage". For parts of the project where AI is not mentioned, supplement according to general practice and note that.

Why the judgment criterion "don't include it if it still holds without AI" is the most valuable: It prevents the most common pitfall——passing off a bunch of ordinary design images as "made by AI".Only keep evidence that "would not exist without AI", to truly prove your AI collaboration ability instead of padding your work.
Supporting materials: Will also provided a capability map titled "What Capabilities UX Needs in the AI Era", which splits 5 major modules——AI cognition/research/design/validation/collaboration——into dozens of secondary capabilities, each described with the same sentence structure "what to input → what action AI performs → what output it produces"——this can serve as an inspiration list for the "what AI was involved in" section above (example: UI visual optimization = upload interface screenshot → AI analyzes hierarchy/white space/alignment → output optimization plan; ROI expression = input experience optimization actions → AI maps to business metrics → output ROI logic chain).

Regarding other materials in this phase (delivery examples, non-AI techniques)

The Deliver section of the source materials also contains a batch of"finished examples" for portfolios/interviews, which are outputs of the techniques above rather than the techniques themselves; they are only mentioned here, not expanded upon:

Verbatim interview script examples(such as the SaaS website personalization project "Breaking Through the 2.1% Conversion Ceiling", the Zhihu Q&A detail page "short container" restructuring project)——all follow the narrative structure of "problem → research insight → technical challenge → solution → data result → retrospective", which is exactly what the prompt in §VI is designed to help you produce.
"Non-Professional Challenges UX Encounters in Projects"——10 types of challenges (misaligned multi-party goals/unclear priorities/cross-cultural/budget resources/short-term vs long-term conflicts/trend pressure/emotional fatigue/compliance/platform constraints/long-term maintenance), each provided with "upfront problem breakdown + handling approach + interview verbal template". The recurring core theme is one sentence:Align goals first, then discuss solutions; judge the problem type first, then take action——this can be used as a soft skill question bank for interviews.

Scope note (honest labeling): The Cognition / Case Study / References subsections in the Deliver area (common interview questions, references for excellent/senior UX designers, pure layout schematic diagrams) lean more toward job-seeking expression and layout sample materials, rather than reusable techniques for "accelerating design with AI", so they have not been distilled item by item; this article focuses on 6 directly reusable AI techniques + 1 custom skill.