Curated Highlights of Medium UX Articles (Multi-Topic: GenAI UX Patterns / Research & Experiment Methods / Mobile Trends / Inclusivity / Information Architecture / Company Cases)

This piece is a distillation of Will's saved 102 Medium UX articles (59 topic folders) saved by Will into "one note that replaces an entire bookmarks folder." The topics are wide-ranging — from GenAI UX patterns to A/B testing, mobile trends, accessibility, information architecture, company cases, and career advice. How to read it:Read by the 7 topic clusters; for each cluster, only note down "the few frameworks/checklists/numbers worth taping to your desk + the Why", and the remaining articles are listed truthfully in the "coverage checklist" at the end, so you can go back to the originals easily.

Reader profile:PMs building AI products, without deep technical specialization. That is why Why / when to use / counterexamples matter more than What here; jargon is explained inline on first appearance.

Important: This repository already has several dedicated wikis covering AI UX patterns (Shape of AI: UX Pattern Library for AI Products (seven major categories: Wayfinders/Inputs/Tuners/Identifiers/Governors/Trust/Dark patterns + design guidelines)), AI interaction paradigms (UX Interaction Paradigms in the AI Era (seven types: AI embedded in traditional interfaces / Agentive / conversational / human-AI collaboration / hyper-personalization / multimodal / trends)), Laws of UX (Laws of UX: 28 UX Laws (Fitts's / Hick's / Miller's / Jakob's / Tesler's / Doherty / Peak-End / Von Restorff / five Gestalt principles / Aesthetic-Usability … in Chinese and English, organized by 6 intent clusters)), UX golden quotes (NNGroup UX Principles / 19 Golden Quotes (You Are Not the User / Keep It Simple / Less Is More / Why Beats What / It Depends … clustered into 5 major themes + mindsets/counterexamples)), information architecture (NNGroup Information Architecture (IA) (three IA models · IA vs sitemap · flat vs deep hierarchy · full taxonomy workflow · left vertical navigation · 5 major pitfalls of privacy policy pages)), persuasive psychology (NNGroup Persuasive Design and Behavioral Psychology (six principles of persuasion · cognitive biases · attention limits · information foraging and wayfinding · mental models · trust hierarchy · self-determination motivation · privacy and ethics boundaries · 26 articles)), UX metrics and ROI (UX Metrics and ROI (HEART/SUS/NPS… metric systems + methods for converting UX value into monetary terms + reporting talking points)). This compilation does not rewrite those; it only adds the checklists, numbers, and company cases unique to this batch of Medium articles, with jump links provided where relevant.

1. GenAI / AI UX Patterns (the largest block in this collection)

One-sentence positioning:Shape of AI: UX Pattern Library for AI Products (seven major categories: Wayfinders/Inputs/Tuners/Identifiers/Governors/Trust/Dark patterns + design guidelines) already provides the "encyclopedia" of AI UX patterns (a seven-category pattern library),UX Interaction Paradigms in the AI Era (seven types: AI embedded in traditional interfaces / Agentive / conversational / human-AI collaboration / hyper-personalization / multimodal / trends) provides the "seven interaction paradigms." What this cluster adds are several more operational checklists — decision gates, automation tiers, specific ways to express confidence/limitations, and a layout gallery of "what AI interfaces look like."

1.1 21 GenAI UX Design Patterns (the most worth-memorizing checklist)

The value of this checklist is not "listing 21 items," but that it strings together "whether to use AI, how deep to go, what to do when it makes mistakes, how to build trust" into one end-to-end chain from decision to implementation. The groups most useful for PMs are selected below:

▸ First pass the decision gate: whether to adopt GenAI (Pattern 1)

Suitable for use: creative tasks (writing/summarizing/drafting), complex content generation (sketch-to-code), when users have difficulty expressing needs through ordinary interfaces.
Not suitable for use: scenarios requiring 100% accuracy (tax filing, legal documents), scenarios requiring fixed, consistent answers (software usage instructions).
Why: GenAI produces probabilistic outputs, which naturally come with hallucinations and inconsistency; in scenarios where "it must be correct," its probabilistic nature is a bug, not a feature. This shares the same root as NNGroup UX Principles / 19 Golden Quotes (You Are Not the User / Keep It Simple / Less Is More / Why Beats What / It Depends … clustered into 5 major themes + mindsets/counterexamples)'s "Stop Solutioneering" — first confirm the problem really needs AI, don't adopt it just because it's trendy.

▸ Augment vs Automate (Pattern 3) + three tiers of automation (Pattern 4)

Automation = AI completes the task for the user (suitable for repetitive, time-consuming, low-value tasks); Augmentation = AI helps the user do the task better (suitable for tasks where the user needs to retain control, or that are creative/highly subjective).
Three tiers:No automation (only gives suggestions, user decides, e.g. Grammarly) → Partial automation/assistance (gives a first draft, user edits, e.g. GitHub Copilot) → Full automation/agent (AI completes the task on its own, e.g. an agent automatically doing competitive analysis and sending emails).
When to use which tier: determined by task risk — low-risk tasks (reminders, filtering) can be highly automated; high-risk tasks (financial transactions) must retain a human oversight gate.
Counterexample: making high-risk tasks fully automated black boxes, where users cannot intervene in time once an error occurs → trust collapses. That is exactly why UX Interaction Paradigms in the AI Era (seven types: AI embedded in traditional interfaces / Agentive / conversational / human-AI collaboration / hyper-personalization / multimodal / trends) emphasizes "explainable + controllable + intervenable" in the "Agentive" section.

▸ The four-piece set for building trust (Patterns 7/8/10/11) — the group AI product PMs should internalize most:

Pattern	How to do it	Counterexample
Express product limitations Convey Limits	State capability boundaries clearly (Claude marks "knowledge cutoff date"; Rufus directly says "I can't help with that" when encountering non-shopping questions)	Marketing hypes it as omnipotent, and users only find out it can't do things after using it → disappointment
Display chain of thought Display CoT	Show processing status ("Retrieving" "Analyzing") + unfold reasoning step by step + mark information sources (Perplexity)	Spinning for a long time without explanation, making users anxious and thinking it is frozen
Provide data sources Data Sources	Answers include expandable citations (NotebookLM links back to original text), especially critical in high-risk domains (medical/financial/legal)	Giving conclusions without evidence; no one dares use it in high-risk scenarios
Express model confidence Model Confidence	Mark confidence levels in high-risk scenarios (Grammarly uses "likely"); for low-risk scenarios (generative art), don't interrupt the user	Stating uncertain answers with absolute certainty, misleading users

Why this four-piece set is integrated: The fundamental risk of GenAI is that "users cannot tell whether a given output is trustworthy." Limitations + CoT + sources + confidence are essentially ways to attach credibility labels to outputs, letting users judge for themselves whether to verify.

▸ Dual-track error handling (Patterns 16/17):

User input error → automatic correction (fuzzy matching when confidence is high) + ask clarifying questions (guide supplementation when input is ambiguous, don't guess blindly) + provide a quick "edit" entry point.
AI system error → three error types: system failure (hallucination/false output), system limitation (cannot answer due to lack of data), contextual error (technically correct but misunderstood by the user); use vague wording to reduce trust risk ("This may not be fully accurate") + provide a recovery path + collect feedback to feed iteration.

▸ Quick notes on the remaining 8 items(one sentence each): Translate user needs into data requirements (work backward from feelings/goals to figure out what data is needed), progressive onboarding (start with basics then advanced features, explain benefits not technical details), leverage mental models (Copilot works like traditional autocomplete), diverse outputs (offer multiple options to spark creativity), memory and recall (remember preferences but provide access/edit/delete entry points), contextual input parameters (auto-fill history / smart autocomplete), Co-Pilot co-editing (AI acts as an inspiration accelerator, not a replacement), automated user control (switches/sliders + undo paths), capture feedback (explicit likes/dislikes + implicit skip/delete signals), model evaluation (LLM-as-Judge + human review; the hybrid approach can push accuracy to 99%), guardrails (filter harmful content, comply with the EU AI Act), data privacy transparency (Slack explicitly states "customer data is not used for training").

Decision-oriented (for AI product PMs): Among these 21 items,1/3/4/7/8 are design decisions that must be answered before launch(whether to build it, how automated it should be, how to label confidence levels and limitations);16/17/18 are the post-launch operational closed loop(what to do when errors occur, how to collect feedback). Going through it as a checklist for AI features is more useful than memorizing each item one by one. For the comprehensive pattern library, see Shape of AI AI Product UX Pattern Library (seven major categories: Wayfinders/Inputs/Tuners/Identifiers/Governors/Trust/Dark patterns + design guidelines).

1.2 Seven Emerging AI UI Layouts (An Illustrated Guide to "What AI Interfaces Look Like")

Categorized by the AI's "position" in the interface, each corresponds to a different AI role and applicable scenario.Which one to choose when depends on whether the AI is a supporting role or the lead:

Layout	AI Role	Suitable for	Not suitable for	Representative examples
Bottom-right floating chat window	Service-oriented assistant	Quickly solving simple problems (order lookup / password reset)	Complex/creative tasks, as the space is too small	Zendesk · Intercom
Inline prompt overlay	Detail assistant	Rewriting/summarizing/autocompleting, for the user to accept or reject	Long-form thinking / creative writing	Notion · Grammarly
AI on an infinite canvas	Creative partner	Generating/renaming/adjusting layout for selected elements	Tasks requiring holistic understanding or strict version control	Figma · Miro · TLDraw
Centered conversation interface	Multi-functional assistant	Free-form exploratory Q&A / creation	Multi-step/structured tasks (form filling / app building)	ChatGPT · Perplexity
Left-side collaboration panel	Strategic partner	Chat on the left, edit on the right, organizing complex content (building websites)	Small screens / feels too heavy for simple tasks	ChatGPT Canvas · Lovable
Right-side collapsible panel	Professional consultant	Helps at key moments without interrupting the main workflow	AI-centric workstyles	M365 Copilot · Cursor
Grid/table interface	Research assistant team	Research/analysis/summarization, with one small assistant per cell	Feels complex to people unfamiliar with tables	Elicit · AnswerGrid

Why this diagram is useful: The most common mistake when building AI features is "defaulting to shoving a chat box in the bottom-right corner". This diagram reminds you——Whether the AI is the lead (centered / left panel) or a supporting role (inline / right panel / floating window) dictates a completely different layout, and choosing the wrong one makes the experience awkward.
and UX Interaction Paradigms in the AI Era (seven types: AI embedded in traditional interfaces / Agentive / Conversational / Human-AI Collaboration / Hyper-personalization / Multimodal / Trends) are cross-referenced: that article covers the interaction paradigms of "AI embedded in traditional interfaces / conversational / human-AI collaboration", while this table shows the physical on-screen placement.

1.3 Predictive AI UX: Use "Context Management Controls" to Reduce Typing for Users

Core insight:A good AI interface does not force users to describe their needs clearly all at once; instead, it provides a set of "context management controls" (time selectors, style options, tone sliders, source filters) for users to click and select, so the system proactively narrows down the scope.

Three context capture scenarios:Search flow(when searching for running shoes, automatically highlight "road shoes" "daily training" and provide dropdowns for terrain/brand),After file upload(after uploading a PDF, automatically show "Summarize/Analyze/Ask a question" buttons),During content creation(when writing an email, provide a tone slider for "Casual/Persuasive/Formal").
Reactive vs Predictive distinction: Claude/Copilot use "default settings + filters" (broad feature-level selections, still requiring users to manually type full questions); true predictive experiences (Apple Reminders suggesting contacts when you type "Call", Gmail Smart Compose completing full sentences, Grammarly correcting errors in real time) are behavior-based, predictive, proactive, and personalized.
Why / Counterexample: Making users describe their needs from scratch facing an empty input box = "blank canvas anxiety". Control-based context management downgrades "describing" to "clicking and selecting", greatly reducing cognitive load. The counterexample is piling on too many controls, which creates a new overly complex interface——refer back to NNGroup UX Principles / 19 Golden Rules (You Are Not the User / Keep It Simple / Less Is More / Why Beats What / It Depends … clustered into 5 major themes + mental models/counterexamples) "Less Is More".

1.4 A Five-Step Framework for Designing AI Products (Lightweight Process)

明确目标(Why)→ 定成功指标 → 团队内快速原型 → 真实用户测试迭代 → 上线后持续监控学习.

There is no essential difference from the general design process,with two AI-specific points: Step 2 metrics must include "ethics/fairness/transparency" (not just technical metrics); Step 5 requires continuous monitoring of accuracy/response time/satisfaction, because models drift over time.
This point aligns with Shape of AI AI Product UX Pattern Library (seven major categories: Wayfinders/Inputs/Tuners/Identifiers/Governors/Trust/Dark patterns + design guidelines) Pattern 19 (Model Evaluation) and the feedback loop in Section 1.1 of this cluster.

II. Research and Experiment Methods (A/B Testing is the Main Course of This Cluster)

One-sentence positioning: The quantitative/qualitative research methods in this repository are mainly covered in the NNGroup series. This cluster supplements a very down-to-earth five-step A/B testing method + two real-world cases with actual monetary amounts, as well as a reminder not to be fooled by vanity metrics.

2.1 Five-Step A/B Testing Method + Two Real-World Numbers

① 定义成功 → ② 找瓶颈 → ③ 建假设 → ④ 排优先级 → ⑤ 上线测试

Step 1 Define success: First answer "Why does the website exist in the first place? What is the single goal that most needs optimization?"——not a vague "optimize everything". Different sites have different metrics: e-commerce = revenue per visitor, media = read time/bounce rate, lead generation = form completion, donations = donation amount per visitor.
Step 2 Find bottlenecks: After aligning on "what counts as success", look at the funnel to find the step where users drop off the most.Obama 2007 campaign: Traffic was not large, but the funnel showed the bottleneck was the "get people to leave their email" step——after the redesign, they gained 2.8 million more email addresses, 10% converted to volunteers (280,000 people),and raised an additional $57 million.
Step 3 Form hypotheses:Haiti earthquake relief donation page(2010)——the original page looked like a tax form, and the hypothesis was "adding photos of the disaster area will make it more empathetic". The result was that donations actually decreased after adding photos (the images were too large and pushed the form below the fold requiring scrolling); after changing to a "two-column text-and-image layout that does not block the form", they ultimately raised an additional $1.02 million, testing 8 versions in 48 hours.
- Why this failure case is the most valuable: It proves that "you must have a clear hypothesis before testing". Blind testing that abandons an idea when results look bad misses the learning of "why it performed poorly"; with a hypothesis, failure itself points to the next clearer question.
Step 4 Prioritize: With limited people, budget, and traffic, prioritize by "which change is most likely to bring a large improvement × most cost-effective"; don't start with something overly complex for your first test (you also need to win boss support).
Step 5 Test: Randomly assign groups, and only draw conclusions after reaching statistical significance.

2.2 Macro Conversion vs Micro Conversion vs Vanity Metrics (Don't Be Fooled by Good-Looking Numbers)

Macro conversion: Directly tied to core goals (completing a purchase, signing up for a free trial).
Micro conversion: Small user actions (clicking a button, watching a video, leaving a comment) that do not directly generate revenue but have indirect value.
Vanity metrics: Look good but are actually useless——a typical example is "headline click volume" (sensational headlines trick people in who leave without reading anything). Truly valuable metrics are comments/shares/bottom clicks/repeat visits.
Why: Without clear goals, it is extremely easy to be misled by "good-looking but useless" numbers and misallocate resources. This point cross-references UX Metrics and ROI (HEART/SUS/NPS… metric systems + methods to convert UX value into monetary terms + presentation talking points) metric systems, and the NNGroup UX Principles / 19 Golden Rules (You Are Not the User / Keep It Simple / Less Is More / Why Beats What / It Depends … clustered into 5 major themes + mental models/counterexamples) "Show Me the Data" principle.

Also in this cluster are "How to conduct user interviews" and "Using AI role-play for UX research"——the core methodology is covered in NNGroup UX Research Method Selection (quick reference for 23 methods / four-quadrant positioning / selecting by question-stage-budget / five major families: qualitative exploration, quantitative validation, expert evaluation, IA, behavioral analysis); the unique aspect of these two pieces is the new practice of using AI as interviewees/sparring partners, which is exploratory.

III. Mobile and Design Trends (2026 Trend Report)

One-sentence positioning: Trend articles only need to be read once to know industry directions; don't memorize them as frameworks. The value of this report on nine major 2026 mobile trends is that it points out a main thread——mobile design is shifting from "interface design" to "system design".

Nine 2026 Mobile Trends (one-sentence summary each + one number)

AI personalization and predictive interfaces: Shifting from "reacting to behavior" to "predicting intent", design moves from drawing "pages" to defining "rules" (Reactive = you do X, give X; Predictive = you might want Y, give Y in advance).
Zero-UI / Conversational: UI becomes invisible, relying on voice/gestures/environmental awareness; the prerequisites are trust + accuracy (41% of users worry about privacy, Siri accuracy is around 83%).
Agentic UX (AI acts on behalf of the user): The system makes decisions for the user → design focus shifts to trust ("Does it understand my intent? Why did it choose this?"), requiring explainability + controllability.
Immersive/Spatial/3D (AR): Driven by Vision Pro;applicable boundaries——only worth it when product value is strongly tied to "spatial awareness" (home/automotive/wearables), otherwise limited.
Glassmorphism / Layered depth: Making a comeback as GPUs mature, but only used to emphasize key elements + ensure contrast; essentially using materials to reinforce information structure, not for decoration.
Microinteractions / Motion feedback: Motion is the "punctuation" of the interface——confirming actions, guiding attention, reinforcing behavior; the key is timing not flashy trajectories, keep it short, clear, and ignorable.
Context-adaptive UI: Responds to environment and behavior (e.g., maps automatically switching to dark mode at night). Data: adaptive UI can improve task completion rate by 22% and engagement by 31%.
Sustainability/Ethics/Inclusion: 80% of a product's environmental impact is determined at the design stage; dark mode saves OLED power; approximately 15% of people globally have accessibility needs.
Passwordless/Biometrics/Security: FIDO2 passkeys support 15 billion accounts, bringing a 70% improvement in login success rate. The UX insight from the Face ID case is brilliant——"Exception-first design": don't pursue 100% perfect recognition, instead anticipate failure scenarios (wearing a mask/glasses) and adapt in advance, making failure "almost imperceptible".

Why this trend report is useful for AI PMs: It ties three AI-related trends ("AI/Zero-UI/Agentic") to a recurring design philosophy——true competitiveness lies in reducing friction, enhancing prediction, and making system behavior understandable and trustworthy, rather than fancier visuals.

IV. Inclusivity / Accessibility (11-Tip Checklist)

One-sentence positioning: For systematic accessibility knowledge, see NNGroup Accessibility / Inclusive Design (5 articles · Three distinctions in inclusive design · Dark mode science and 8 major pitfalls · Screen reader user research · 16 guidelines for phone voice trees). The value of this 11-tip list is that it is extremely concrete and actionable——Every item can be directly used for self-audit, and the opening provides revenue evidence that "accessibility = profit", which is perfect for convincing your boss.

4.1 Convince your boss with money first: Accessibility is not charity, it is revenue

There are approximately 13.9 million people with disabilities in the UK (20% of the population), and over 1 billion globally.
Legal & General After improving accessibility, online life insurance sales doubled;Tesco After the change, annual grocery delivery sales increased by £13 million (and the main site was replaced with the accessible version);Virgin After the change, online sales increased by 68%.
Why: Accessible websites generally have better SEO, faster loading speeds, and are easier to use for ordinary users too — this is a typical case of "doing one thing right and incidentally improving the experience for everyone".

4.2 11-item self-audit checklist (ready to use directly)

Integrate accessibility into daily work: Plan before development, don't wait to patch it during the testing phase (fixing it later is expensive).
Sufficient contrast: WCAG requires a text/background contrast ratio of 4.5:1≥ 4.5:1; don't use light gray text on a dark gray background. Approximately 8% of men and 0.5% of women have color blindness.
Multi-channel coding: Don't rely solely on color — add icons + headings + bold/underline + patterns (Trello color blindness mode). Self-check method: Print it in black and white to see if you can still recognize the content.
External form labels: Don't tuck prompts inside input boxes (they disappear when you start typing + screen readers skip them); place labels outside the boxes.
Image alt text: Describe the content + its relationship to the context; use empty alt for purely decorative images so screen readers skip them (don't let it read the file name).
Organize content with heading hierarchy: Use h1→h2→h3 to represent structure (screen readers rely on this for navigation), don't misuse headings for styling; add a "skip to main content" link.
Obvious interactive elements: Clickable elements must look clickable + have focus indicators (required for click/keyboard/remote control).
Support keyboard operation: Tab order should follow the visual flow (left→right, top→bottom); if the entire site works without a mouse, it passes.
Large enough touch targets: Recommended ≥ 7–10 mm², so a finger can tap comfortably (on in-car screens especially, place frequently used buttons within easy reach).
Test, test, and test again: Turn on a screen reader to listen to content, zoom in and out, use only the keyboard — build empathy first; but testing with real disabled users is irreplaceable.
Make good use of ready-made tools: Apple/Android accessibility guidelines, inclusive-components.design pattern library, WebAIM contrast checker.

Decision-oriented: Among these 11 items, 2/4/8/9 are the most frequently missed(contrast, form labels, keyboard accessibility, touch targets); PMs working on AI products should prioritize these four during reviews.

V. Information Architecture / Process / Complexity (System Perspective)

One-sentence positioning: For systematic IA knowledge, see NNGroup Information Architecture IA (Three IA Models · IA vs Sitemaps · Flat vs Deep Hierarchies · Full Taxonomy Process · Left Vertical Navigation · 5 Pitfalls of Privacy Policy Pages). This cluster supplements three system-level thinking tools——how team collaboration breaks down silos, how complexity accumulates quietly, how value propositions align, and why globalization does not equal translation.

5.1 How complexity accumulates + how to reduce it (the most useful piece for PMs)

Complexity does not come from one big mistake, but from the stacking of countless small decisions that "seem reasonable at the time", and it accumulates exponentially(10 components = 45 interaction relationships, 100 components = 4950).

Four sources of accumulation: Feature bloat (every added feature increases code/interactions/testing/documentation/training), process stacking (adding approvals when errors occur but never removing old processes, like geological strata), informal workarounds (using Excel to fill system gaps → forming a "hidden system" that collapses when people leave), technical debt (not only in code, but also in organization/training/reporting relationships).
Why people can't control it: Human working memory can only process about 7 chunks of information; when a system grows to hundreds of components, no one can understand the whole picture → decisions can only be based on local perspectives. There is also "emergent behavior" (traffic jams = the sum of individual driving choices, organizational chaos = the sum of local optima across departments).
Real blowups: Healthcare.gov (2013, modules worked individually but failed in a chain after integration), Knight Capital (2012, 7 out of 8 servers had updated code, lost $440 million in 45 minutes), Boeing 787 (supply chain coordination complexity caused delays).
How to reduce it (ready to use directly): Use dependency graphs to find sources → Pareto analysis (20% of features bring 80% of value, delete low-frequency/edge ones) → complexity budget(adding 1 feature requires deleting 3 old ones) → simplification audits (regularly proactively delete, not optimize) → modular splitting → standardization and platformization.
Cultural layer: Misaligned incentives are the root cause (PMs are rewarded for "adding features", processes only increase); "deletion and simplification" must also be included in rewards, with dedicated simplification projects."Saying No" is one of the most effective control methods.
Why / For PMs: This piece directly addresses your biggest daily temptation — "just add one more feature". Treating the "complexity budget" and "add one, delete three" as hard rules is the most practical guardrail against feature bloat. It is consistent with NNGroup UX Principles / 19 Golden Rules (You Are Not the User / Keep It Simple / Less Is More / Why Beats What / It Depends … Clustered by 5 Major Themes + Mindsets/Counterexamples) "Keep It Simple / Featuritis" and the Spotify/Evernote cases in §6.1 of this compilation.

5.2 Collaboration framework: From centralization to decentralized shared Commons

An organizational design article that uses the "Agricultural Revolution" as an analogy, explaining how design teams evolve from "closed and siloed" to "shared collaboration". Three structural evolutions:

Centralization (concentric circles): The boss sets direction → middle management passes messages → frontline does the work; innovation can only start from the center, outer circles have no decision-making power → like an "ancient farm", information does not flow, talent is constrained. Many modern companies actually operate this way.
Decentralization (product-centric): UI/UX/Product/CX/Service Design act as equal nodes around "products and services", operating through innovation flow (red lines) + feedback systems (yellow lines).
Shared core resources (Commons): The center is a shared library — design systems, research document libraries, strategy templates, values, compliance templates. Teams do not start from zero, but take reusable assets from the Commons (Ant/Material Design are Commons; Google OKRs are cultural Commons; Airbnb's experiment platform is a feedback Commons).

Includes a practical comparison table: Differences in focus, typical tasks, deliverables, and pain point types across five roles — UI / UX / Product Design / CX / Service Design — PMs can use it directly to align who is responsible for what when coordinating cross-functionally.
Why: Especially relevant for AI product teams — role boundaries are blurring in the AI era (see §7), and a shared research library/design system/evaluation standard library can prevent teams from repeatedly stepping into the same pitfalls.

5.3 Value Proposition Canvas (aligning "why users buy from you")

Business Model Canvas (BMC)looks at "how the company survives" (9 elements: partners/activities/resources/value propositions/customer relationships/customer segments/channels/costs/revenue);Value Proposition Canvas (VPC)is a magnifying glass for the two blocks of "value proposition × customer segment" in the BMC, looking at "why users choose you".
Two halves of the VPC: The right side is the customer profile = Jobs (what they want to accomplish) + Pains (what they hate) + Gains (what they want to get); the left side is your response = Pain Relievers + Gain Creators. A match between left and right = a unique value proposition.
When designers use it: Early research (aligning real needs, don't brainstorm features out of thin air), product iteration (updating Pains with user feedback), before development (verifying whether a solution "is worth doing").
Why / Counterexample: Before designing a fitness app, you must first figure out whether users want to "lose weight" or "become confident/socialize more" — the VPC forces you to distinguish surface needs from real motivations. It is intertextual with NNGroup Persuasive Design and Behavioral Psychology (Six Principles of Persuasion · Cognitive Biases · Attention Limits · Information Foraging and Wayfinding · Mental Models · Trust Hierarchies · Self-Determination Motivation · Privacy Ethics Boundaries · 26 Articles), Shensi Circle / AI B2B Vertical Industries (8 Major Track Unicorn Cases)'s demand insights.

5.4 Global UX: A system challenge, not a translation challenge

This is the most information-dense article in the "Multilingual/Localization" sub-cluster (about 10 articles), with the core argument:Internationalization ≠ Localization ≠ Translation.

Internationalization (i18n) = building infrastructure that supports multiple languages/countries (multilingual interfaces, time formats);Localization (l10n) = deeply integrating into local culture so that information is "truly accepted". Most companies only do translation.
Compression/expansion ratio: English to German averages35% expansion("Food Poisoning" is 13 characters → German is 22 characters, +69%); German to Chinese can compress by 40-50%. Therefore input boxes/buttons/prompts must leave flexible space.
Three cases: Chevrolet Nova sounded like the Spanish "No va (doesn't go)" in Latin America (misunderstandings still happen within the same language across cultures); a bank debit form was only expanded to multiple countries late in the process (you should select the country first then show fields, otherwise rework is needed); Domino's (understood 70% of Indians are vegetarian, set up a vegetarian zone + paneer flavor, annual revenue $870 million) vs Kellogg's (in 1994 rejected the local "hot food" culture to promote cold cornflakes, market share <0.01%, later launched upma instant meal packets for +30% growth).
Core reminder:"UX Debt — Assuming Universality": Not considering global users at the start leads to accumulated experience debt from later rework, delaying development, damaging trust, and losing users. This is the same "do it right early to save trouble later" logic as complexity accumulation in §5.1 of this compilation.

Other localization articles in the same cluster (Netflix pseudolocalization, UI/UX Localization guide, Wix global users, designing for internationalization, product localization tips, multilingual interface experience, language and literacy) — when making overseas products, refer to the original texts as needed; the main viewpoints are all covered by the article above.

VI. Company Case Teardowns ("Product Autopsies" in the Bookmarks)

One-sentence positioning: Will's bookmarks contain a batch of folders named after companies (Airbnb/Calm/Canva/Duolingo/Gmail/LinkedIn/Otter/Slack/Spotify/Uber), mostly collections of UX/growth teardowns for those companies. Here we select the two most information-dense representatives + one recurring negative example; the rest are listed in the end-of-article list.

6.1 Recurring negative example: Feature bloat kills the core experience (Spotify / Evernote)

Spotify: Expanded from music to podcasts, audiobooks, AI playlists, short videos, event marketplaces — every individual feature makes sense, but the overall experience becomes more complex, and many users explicitly say they "just want the core music experience".
Evernote: Expanded from notes to multiple directions and even sold physical merchandise (socks and backpacks in 2017), downloads fell for 6 consecutive years by a cumulative-82%, and the new team had to cut features to return to the core.
Why / For PMs: This is ironclad evidence that "feature expansion ≠ user growth, and can even lead to churn". It is a different slice of the same lesson as complexity accumulation in §5.1 and the NNGroup UX Principles / 19 Golden Rules (You Are Not the User / Keep It Simple / Less Is More / Why Beats What / It Depends … Clustered by 5 Major Themes + Mindsets/Counterexamples) "Less Is More" principle.

6.2 Duolingo ("Product Autopsy" collection on growth + retention)

The Duolingo folder is the thickest company collection in this batch (8 sub-articles), covering: GPT-4-powered Duolingo Max, company strategy (mission-driven growth), onboarding testing (personalization/animation/speed), reviving user growth, evolution of the energy system, 20-day recall messaging blitz ("emotional blackmail"), 6-step reactivation experience, and freemium→premium guidance.

Use for AI product PMs: Duolingo is a textbook example of "gamification + recall + free-to-paid conversion" — when designing for retention and monetization, this is the most worthwhile set to read article by article (this compilation only indexes it and does not elaborate, as it deserves a dedicated deep read on its own).
Intertextual with NNGroup Persuasive Design and Behavioral Psychology (Six Principles of Persuasion · Cognitive Biases · Attention Limits · Information Foraging and Wayfinding · Mental Models · Trust Hierarchy · Self-Determination Motivation · Privacy Ethics Boundaries · 26 articles) (persuasion/motivation), UX Metrics and ROI (HEART/SUS/NPS… metric systems + methods to convert UX value into monetary terms + reporting talking points) (retention metrics).

6.3 Other company case studies (index)

Airbnb(Airbnb design language and historical releases (2021–2025 quarterly release timeline + DLS design system + trust/belonging design) (already has a dedicated wiki), Calm (meditation app experience), Canva (ease of use/templating), Gmail (email interaction), LinkedIn, Otter.ai, Slack, Uber — when doing benchmarking analysis, look up the original source documents by company name.

VII. Career / Soft Skills (What Designers Are Worth in the AI Era)

One-sentence positioning: For the UX career development system, see NNGroup UX Upward Influence · Metrics · Strategic Tools · Roadmaps · Career (13 articles: Proving UX Value Upward How to Sell UX "If…then…" formula · Rhetorical triangle ethos/pathos/logos · UX & Marketing collaboration / UX metrics aligning organizational goals workshop · UX Goals vs OKRs vs KPIs · CASTLE must-use software 6 dimensions / Stakeholder analysis power-interest matrix + attitude · SWOT four quadrants / Roadmap 6 steps · Prioritization 5 methods · Journey map time cost 73.8h / Career STAR · METEOR interview · Portfolio) / Google UX Certificate / Portfolio · Resume · Interview · Job Search. This cluster adds two sharp articles directly addressing "where is the value of designers/PMs in the AI era", which are very useful for people who are wondering "will AI replace me".

7.1 Three types of designers companies want in 2026 (traditional senior designers are not among them)

Based on the hiring criteria of Anthropic design lead Jenny Wen + a survey of 906 designers. Core points:Hiring shifts from "years of experience" to "capability profile".

Data context: Designers' evaluations of AI — makes me faster (89% agree), better collaboration (80%), better design (91%), slows me down (only 25% agree). Conclusion:"AI makes me faster" is already a consensus-level basic capability, no longer an advantage.
The three types companies actually want:
1. Block-shaped generalist: Reaches ≈80 points across design, product, and code, able to drive implementation across roles on their own — suitable for current organizations with "blurred boundaries".
2. Deep specialist: Top 10% in a certain field, because AI makes "passable design" cheap, and what is truly valuable is "extreme quality".
3. Cracked new grad: Extremely strong learning ability, no baggage from old processes, directly builds products hands-on.
Not on the list: Traditional senior designers (8 years of experience + strong portfolio + mature team) — the problem is not weak ability, but reliance on old organizational structures (PMs provide requirements, researchers provide data), while these supporting roles are being compressed or eliminated by AI.
Why / For PMs: This applies equally to PMs — what is valuable in the AI era is "judging what to build" rather than "execution".

7.2 The real skill of UX in the AI era: restraint and judgment > drawing UI

Big change: prototypes that used to take weeks can now be generated in hours, and "working software" is no longer scarce. Design value shifts from "making it" to "deciding what to make and how to make it". Historical pattern: after printing became cheap, writing became important; after photography became widespread, aesthetics became important; after code became easy,judgment becomes important.
Winning design vs functional design: Functional = clearly shows 50 bugs; Winning = identifies 48 of them as sharing the same root cause, and directly provides a "Fix all" button —front-loading complex decisions and digesting them for the user (this is agentic design).
Three tricks for experience speed: Reduce cognitive load (Arc lets users manage tasks instead of tabs), compress time (Linear/Superhuman use keyboard commands to shrink minute-level operations to seconds), build intuition (if users have to stop and think "what's next", the design has failed).
Designers' real work = restraint + choice: As AI-generated interfaces become increasingly homogeneous, the key becomes "does the product have a clear point of view" — not offering 50 sorting methods, but giving "the most reasonable one"; not showing all possibilities, but clearly telling users "what they should do".Shift from "providing options" to "defining the path", from "execution" to "judgment".

7.3 "When building is free, what is worth building" (companion piece to §7.2)

Data: GitHub Copilot speeds up development by approximately55%; about25% of companies in the 2025 YC Winter batch have 95% of their codebases generated by AI; "vibe coding" became Collins' word of the year.
The Michelin question: The difficulty of high-level creation is not "how much you can make", but "deciding what not to make" (Robuchon's minimalist cuisine relies on restraint). When costs approach zero, features are extremely easy to stack infinitely.
What low-cost software truly unlocks: Internal tool revolution (operations teams build usable tools in a day), prototypes going from days to minutes (alleviating whiteboard fear + more realistic testing), the design-development wall disappearing ("design engineers" deliver runnable components).
But two pitfalls for user experience: The "pixel-perfect trap" (AI-generated interfaces look good but lack validation/error states/navigation logic, and fail basic usability testing), the "Spotify problem" (feature bloat).
Conclusion: AI improves execution efficiency; the intention of "deciding what to build" still requires human judgment.Key capability shifts from "how to build" to "whether to build"; speed is only a tool, not a goal.
Why / For AI product PMs: These two articles are the "core thesis" of this collection — when AI makes "building things" cheap, the core value of PMs shifts completely from "can it be built" to "should it be built, and which one to build". This aligns with all the "subtract / understand the problem first" maxims in NNGroup UX Principles / 19 Golden Rules (You Are Not the User / Keep It Simple / Less Is More / Why Beats What / It Depends … clustered by 5 major themes + mindsets/counterexamples), and the complexity budget closed loop in §5.1.

VIII. List of articles covered in this compilation (59 topics, listed as-is)

Below, all 59 folder titles are listed categorized by the 7 clusters above.Those marked with ★ are the ones elaborated in the close reads above; the rest have been scanned and categorized by title; return to the sources to check the original texts when details are needed.

① GenAI / AI UX Patterns

★ 21 GenAI UX Patterns
★ Seven Emerging AI UI Layouts
★ Predictive AI UX
★ A Practical Framework for Designing AI Products
Navigating the new age of data visualization
Synthetic data for designers
Role Play with AI UX Research — User Interviews
Testing your UX ideas with vibe coding
Vibe Coding for mobile with Cursor AI
Google says "Vibe Design" is here, but it didn't…
Escaping AI sludge why MVPs should be delightful

② Research and Experiment Methods

★ 5 Steps of A/B Testing
How A/B Testing Helped Obama Win the Election
Email Testing: Obama Campaign's $500M Email Fundrai…
How to drive and conduct user interviews properly

③ Mobile and Design Trends

★ 9 Mobile App Design Trends for 2026
Creating Usability with Motion: The UX in Motion Manifesto
Why your brain rebels against redesigns — even good ones

④ Inclusivity / Accessibility

★ 11 tips for designing an inclusive user experience
How to Avoid Common Accessibility Issues

⑤ Information Architecture / Processes / Complexity / Visual Foundations

★ A collaborative framework that breaks down information barriers
★ How UX complexity accumulates
★ The Value Proposition Canvas
How to Create a Value Proposition Canvas
★ Global UX: A System, Not Language Challenge
Design for Internationalization
Designing Multi-Lingual UX
Multi-language UI Design Insights
Multilingual UI UX: Language & Literacy in App Design
Netflix Pseudo Localization
Product Localization Tips and Tricks
UI Localization: The Ultimate Guide for Beginners
UX Localization: Adapting the User Experience for I…
Wix's Global Users and Localization Design
Information hierarchy: high-density data design
Design principles for UI alignment
The Anatomy of Input Field
Navigation design
How to communicate errors
Color Psychology Facts and Statistics
How to be strategic when picking a typeface
What is design thinking

⑥ Company Case Breakdowns

★ Duolingo (8-article collection: Max / Strategy / onboarding / Growth / Energy system / Recall / Reactivation / Free-to-paid conversion)
★ Spotify (Feature bloat case study)
Airbnb (see also Airbnb design language and historical releases (2021–2025 quarterly release timeline + DLS design system + Trust/belonging design)) · Calm · Canva · Gmail · LinkedIn · Otter.ai · Slack · Uber
Others (miscellaneous collection) · The long and short of telephone progress
What happened to the car designed for women, by women
Designing useful ads

⑦ Career / Soft Skills / Industry Judgment

★ Companies want three types of designers in 2026
★ Which UX skills matter most in the AI era
★ When building is free, what's worth building
★ The Psychology-Backed Formula for Landing Pages…
What Does a UX Engineer Do, Exactly
How being a strategic advisor helps grow design influence
Innovation is not magic; it's technique
Taste is not a feature
Functionality comes first; a pretty website doesn't…
Low Traffic ≠ Failure — Low Conversion Is
What We Learned Doubling Client Conversions in 90 days
The Fallacy of "Anyone Can Build an App"

Sourced from Will's curated collection of Medium UX articles (102.md files / 59 topic folders, mixed Chinese-English layout). This compiled edition makes selections (ingest_quality: synthesized): for each cluster, closely read the few articles with the highest information density, and categorize the rest by scanning titles. It is highly intertextual with this repository's AI UX / UX fundamentals series: AI pattern library Shape of AI AI Product UX Pattern Library (seven major categories: Wayfinders/Inputs/Tuners/Identifiers/Governors/Trust/Dark patterns + design guidelines), AI interaction paradigms UX Interaction Paradigms in the AI Era (seven types: AI embedded in traditional interfaces / Agentive / Conversational / Human-AI collaboration / Hyper-personalization / Multimodal / Trends), UX laws Laws of UX 28 UX Laws (Fitts's / Hick's / Miller's / Jakob's / Tesler's / Doherty / Peak-End / Von Restorff / Five Gestalt principles / Aesthetic-Usability … Bilingual Chinese-English, organized by 6 major intent clusters), UX quotes NNGroup UX Principles / 19 Quotes (You Are Not the User / Keep It Simple / Less Is More / Why Beats What / It Depends … Clustered into 5 major themes + mindsets/counterexamples), information architecture NNGroup Information Architecture IA (Three IA models · IA vs sitemap · Flat vs deep hierarchy · Full Taxonomy process · Left-side vertical navigation · 5 major pitfalls of privacy policy pages), persuasive psychology NNGroup Persuasive Design and Behavioral Psychology (Six principles of persuasion · Cognitive biases · Attention limits · Information foraging and wayfinding · Mental models · Trust hierarchy · Self-determination motivation · Privacy ethics boundaries · 26 articles), UX metrics ROI UX Metrics and ROI (HEART/SUS/NPS… metric systems + Methods for converting UX value into monetary terms + Reporting talking points), Airbnb design language Airbnb design language and historical releases (2021–2025 quarterly release timeline + DLS design system + Trust/belonging design). Original image/video files are retained in sources (already gitignored).