Systems & primitives — Field Notes

BOOTCAMP

LDLR DESIGN FIELD NOTES

Prompts are a thin layer

Stripe, Fin, Anthropic, Vercel, and Linear turn design knowledge into context agents can act on, graded on a maturity scale.

OCT 5, 2026

how design gets done in 2026, part 3 / inside the AI infrastructure of Stripe, Fin (formerly Intercom), Anthropic, Vercel + Linear

The chats are disposable. The brief is the institution.

Why it matters

— Four maturity levels: human-readable, machine-accessible, machine-actionable, and self-evaluating

— Stripe's Protodash wraps AI prototyping in Sail so generated UI matches shipped product

— Linear's CEO built a custom skill drawing on 40,000 customer requests

— Fin committed every designer to shipping production code, then codified it in a framework

— Her warning: governance is lagging adoption across every org she profiled

Read on Bootcamp →

Systems & primitives Cross-functional & process

SMASHING MAGAZINE

LDLR DESIGN FIELD NOTES

A name everyone can use

Friedman tours naming conventions for tokens, components, layers, and features, built on the systems other teams already use.

OCT 6, 2026

A Practical Guide To Naming Things

The right name is the one that's well-understood and used by the team and users.

Why it matters

— Good names are short, meaningful, known by everyone, and free of visual properties

— Vodafone's token taxonomy climbs from primitives to semantic names that reveal their purpose

— Ask users to describe a feature in their own words, then adopt their words

— Check the Component Gallery before coining a name other teams already settled

— Low adoption can be a naming conflict, so log it in the backlog

Read on Smashing Magazine →

Systems & primitives Craft & detail

DISHANT SHARMA

LDLR DESIGN FIELD NOTES

Design systems as files

Google's Stitch team shipped DESIGN.md, an open alpha spec pairing tokens with prose intent so coding agents inherit visual rules.

OCT 4, 2026

Why DESIGN.md Broke the Internet in Under 24 Hours

Design systems are becoming text files.

Why it matters

— YAML tokens for color, type, spacing, and radii, with Markdown explaining the intent

— A CLI lints the file for broken references and contrast failures

— One open format could travel across tools instead of living inside one vendor

— The announcement reached 2.2 million views within a day of shipping

— Skeptics: a file generated from appearance records the visible choices, not the design intent

Read on DEV Community →

Systems & primitives

DIVE CLUB

LDLR DESIGN FIELD NOTES

Systems at Google scale

Rhiannon Bell, VP of UX for Google Search, on the decisions behind AI Mode, dynamic component systems, and the skills gaining value.

OCT 6, 2026

Systems & primitives Hiring & team

NIELSEN NORMAN GROUP

LDLR DESIGN FIELD NOTES

Curation beats prompting

NN/g’s power-user study lands one shift: context curation, not prompt craft, is where the leverage moved.

BY TANNER KOHLER / SEP 18, 2026

The 3 Roles of Context for AI Agents

Curating context now matters more than writing the perfect prompt.

Why it matters

— Every power user studied had built a context library before agents got good

— Three roles: global rarely changes, local tracks the task, ambient streams in raw

— Handling follows role: curate global, let the agent write local, let it sift ambient

— A deliberately lazy prompt worked fine because the MD file did the heavy lifting

Read on Nielsen Norman Group →

Research & validation Systems & primitives

BUILDING NUBANK

LDLR DESIGN FIELD NOTES

Written to be heard

A low-vision Nubanker wrote BLIND-CLAUDE.md: every response must survive being read aloud, top to bottom, in order.

BY NUBANK EDITORIAL / SEP 16, 2026

Designing Claude for ears, not eyes: An accessibility skill for blind and low-vision users

Every response must make sense read top-to-bottom, out loud, in order.

Why it matters

— BLIND-CLAUDE.md treats auditory comprehension as an engineering constraint, not a UI afterthought

— Tables become prose: a screen reader narrates cells as disconnected values with no structure

— Spatial references like 'see above' become restatements; every section carries its own context

— Long build logs collapse to the two or three lines that matter

— One person's validated workflow, honestly flagged: no broader user feedback collected yet

Read on Building Nubank →

Craft & detail Systems & primitives

DESIGN SYSTEMS COLLECTIVE

LDLR DESIGN FIELD NOTES

Can, should, may

Ahmadi's test for every agent action: execution, judgment, governance. Readiness is legibility, not tooling.

SEP 16, 2026

A Design System Isn't AI-Ready Until Its Decisions Are Legible

Can the agent do it? Should it do it this way? May it decide this?

Why it matters

— AI readiness is a governance problem before it is ever a tooling problem

— An agent invents the missing button variant and commits it to the shared library

— Decision legibility: a non-participant sees what was decided, why, and who owns it

— Source of truth says what exists; source of authority says why it binds

— Agent experience design: UX for the repository the agent has to navigate

Read on Design Systems Collective →

Systems & primitives Critique & decision-making

DESIGN SYSTEMS COLLECTIVE

LDLR DESIGN FIELD NOTES

The bill under the demo

45 seconds to generate the screen, three hours to make it honest. Singh's metric: time to trusted output.

SEP 18, 2026

The 'Fix-it Tax' of Generative UI

It is a fast proposal with a bill folded underneath.

Why it matters

— Green rectangles for switches, raw hex for surface.base, focus states stripped for cleanliness

— Generated UI delivers the appearance of decisions without the decision trail

— Cleanup is archaeology: find the assumption, trace the rule, replace it without breaking structure

— Her fix: a pattern contract written before generation, tokens and boundaries included

— One innocent token had failed contrast testing six months earlier; no model knows that

Read on Design Systems Collective →

Systems & primitives Prototyping & exploration

BUILDING NUBANK

LDLR DESIGN FIELD NOTES

Skills are dependencies

Palma and Martins gate the marketplace: 2,000 skills vetted, 1,600 risks flagged, before one reached a developer.

BY LUCAS PALMA AND PAULO MARTINS / SEP 2, 2026

When AI skills become supply-chain dependencies

A prompt can suggest behavior, but a control outside the model can enforce it.

Why it matters

— Skills read like documentation, but their operational effect is a dependency

— Skill Vetter gates the marketplace: 2,000 scanned, 1,600 risks, 1,000 remediated

— Deterministic rules catch known patterns; an LLM reads intent across instructions

— Ask-before-running lives in the prompt, so the model approves its own action

— Local checks plus CI make the safe path the easiest path

Read on Building Nubank →

Systems & primitives Cross-functional & process

DESIGN SYSTEMS COLLECTIVE

LDLR DESIGN FIELD NOTES

Who checks the checker

Williams reads seven August writers landing on one gap: enforcement graduated into furniture, which nobody owns.

BY SHANE P WILLIAMS / SEP 11, 2026

Who Checks the Thing That Checks

A recorded reason with no owner is just a better-documented assumption.

Why it matters

— Custody evaporates when a rule hardens into infrastructure; pipelines become furniture

— Negreiros, in a comment thread: refuse to let enforcement graduate into infrastructure

— Encoded governance looks identical to working governance until someone checks

— Ovcharenko: drift surfaces when someone new asks why docs and file disagree

— The apprenticeship that produces that noticing person is being automated away

Read on Design Systems Collective →

Systems & primitives Critique & decision-making

DESIGN SYSTEMS COLLECTIVE

LDLR DESIGN FIELD NOTES

Make breaking it hard

Kumar’s five layers for agent-built UI: tokens, components, context, discovery, and enforcement wired into CI.

BY RAKESH KUMAR / SEP 11, 2026

AI-Generated UI Needs a Design System That Can Enforce It

Don’t ask AI to remember your design system. Make violating it difficult.

Why it matters

— The agent’s dashboard compiled, then shipped a second table component anyway

— Context tells AI what the system is; enforcement tells it what it cannot violate

— An MCP server exposes the contracts; compliance still lives in the pipeline

— Raw hex, arbitrary spacing, and duplicate components all fail the build

— Compliance becomes a metric: token usage, custom components per agent PR

Read on Design Systems Collective →

Systems & primitives Cross-functional & process

UX COLLECTIVE

LDLR DESIGN FIELD NOTES

Eras keep the wrong half

Neeman walks five eras since Bell Labs, each keeping the disposable habit and forgetting the durable lesson.

SEP 11, 2026

Five eras of UX design and the lessons to keep in the AI era

Information architecture didn’t die. It went underground and waited for the machines to need it.

Why it matters

— Karlin’s touchtone keypad and Deininger’s 1960 grid study open the lineage

— Norman’s 1993 Apple title covered industrial design, interface, even the manual

— Design system ceremony only made sense while execution stayed expensive

— UXPA logs the worst net staffing reduction on its record, below the 2009 trough

— Leverage moved upstream to rules, structure, and the data agents execute against

Read on UX Collective →

Systems & primitives Storytelling & narrative

LINEAR

LDLR DESIGN FIELD NOTES

Three of every four PRs

Ramp built Inspect with six years of Linear as its memory. The agent now writes three of every four merged PRs.

AUG 31, 2026

The coding agent behind 75% of Ramp’s merged PRs

The speed has moved Ramp’s bottleneck from writing code to reviewing it, a constraint any company running agents at this scale will meet.

Why it matters

— Inspect started as a casual idea; three engineers shipped the first version in two weeks

— Built in-house so it fits Ramp’s lifecycle exactly: one API key, their own log schemas, their sandboxes

— A designer takes a dashboard rework most of the way from Figma, then hands the live session to an engineer

— In Slack, “@Inspect, fix this” is often all it takes to open a PR

— A central team defines what agents can access, so functional teams automate safely on shared rails

Read on Linear →

Cross-functional & process Systems & primitives

MURPHY TRUEMAN

LDLR DESIGN FIELD NOTES

Forks look like adoption

Import counts can’t tell reuse from copying, and AI made forks cheap. Trueman watches prop shapes and wrappers instead.

BY MURPHY TRUEMAN / AUG 21, 2026

Every fork looks like adoption

Forking has always looked like adoption. It just used to be expensive enough that someone actually noticed.

Why it matters

— GitClear, 623M code changes: cross-file reuse down 35%, refactoring down 70%, duplication up 81% since 2023

— A ProductCard wrapping Card counts as adoption while quietly becoming a parallel API

— Signal one: a long tail of one-off prop combinations means teams work around the API, not with it

— Signal three is the one he wants most: a Button that never imports Button, caught by structural similarity

— The check doesn’t need to block the merge — it needs to make the twenty-minute conversation happen

Read on Murphy Trueman →

Systems & primitives Critique & decision-making

DESIGN SYSTEMS COLLECTIVE

LDLR DESIGN FIELD NOTES

The rule breaks first

Negreiros publishes the ADR, the permissions, and the Safari bug that broke production twice while every rule held.

BY GUILHERME NEGREIROS / AUG 27, 2026

Agents Don't Break the Rules. The Rules Break First.

The rule was applied exactly as written. The regression came back the same day it left.

Why it matters

— ADRs agents read before proposing anything: context, decision, rejected alternatives

— Three permissions in the repo: observe, generate, never overrule a ratified decision

— WebKit blocks var() inside :visited on purpose; the token never resolved

— A compliance audit removed the hex fallback, correctly, and reintroduced the bug

— Even the post-incident note drifted; records need review, not write-once trust

Read on Design Systems Collective →

Systems & primitives Critique & decision-making

LINEAR

LDLR DESIGN FIELD NOTES

Fence the cascade, too

Skovhus turns a 1,000-PR styling migration into a boundaries argument: contracts enforced by tooling, built for agents.

BY KENNETH SKOVHUS / AUG 26, 2026

Styling Linear for the future with StyleX

Many of the patterns that are hardest to migrate are the ones that made the old system difficult to reason about at scale.

Why it matters

— Non-negotiables: build-time styles, explicit contracts, deterministic merging, safe long-term bet

— Styling at a distance made deliberately difficult, not discouraged by convention

— Deterministic codemod first; agents took more load as models improved mid-project

— A dev-bar counter and per-component highlighter made progress visible and social

— Lint rules and a type-aware repo checker enforce the conventions afterward

Read on Linear →

Systems & primitives Craft & detail Cross-functional & process

DESIGN SYSTEMS COLLECTIVE

LDLR DESIGN FIELD NOTES

Pieces, never the puzzle

Donald, a career systems hand, suspects we’ve outgrown design systems and sketches the one that absorbs the product.

BY DAN DONALD / AUG 27, 2026

Do we still need design systems or do they evolve to become something more?

We have the pieces but rarely do we see the puzzle — the products, the features or the journeys they're being used for.

Why it matters

— Design systems stop at patterns; products, features, and journeys stay out of scope

— His coinage: Digital Product Systems, owning business context down through tokens

— Making docs clearer for machines made them better for humans too

— If the system generates the product, the job shifts to system integrity

— New problems, not pattern-following, are where the people stay vital

Read on Design Systems Collective →

Systems & primitives Cross-functional & process

LINEAR

LDLR DESIGN FIELD NOTES

Delete the IDE

Turakhia took the editor away from an entire org for two weeks: the old way of working lives there.

AUG 11, 2026

Coinbase's bet on agent-first development

Every conversation now starts with the same question. Why can't an agent do this, and how do we make it possible?

Why it matters

— Chintan Turakhia had every engineer delete their IDE for two weeks

— Coordination knowledge scattered across heads and Slack is a wall for agents

— Forge turns a Slack bug report into a labeled issue, a fix, and a PR

— Speedruns spread the workflows: one session, 500 PRs in fifteen minutes

— The tracked metric is autonomous operation time, and it keeps climbing

Read on Linear →

Cross-functional & process Systems & primitives

DESIGN SYSTEMS COLLECTIVE

LDLR DESIGN FIELD NOTES

Local first, system later

Local components as signals, not violations: designers build in place, and the system team governs what earns promotion.

BY VEDANT / AUG 21, 2026

Let Designers Build. Let Design Systems Govern.

The maturity of a Design System isn't measured by how few local components exist.

Why it matters

— Local components are experiments; system components are governed contracts

— Six steps: build, validate in product, review, assess, contribute, publish

— Five teams building the same empty state is feedback, not violation

— Custodian becomes governor and enabler; designer becomes contributor

— Not every successful local pattern deserves promotion to the system

Read on Design Systems Collective →

Systems & primitives Cross-functional & process Critique & decision-making

LINEAR

LDLR DESIGN FIELD NOTES

Fence the field

Wolting swapped predictability for boundaries: system prompt, tools, run scope, harness, and the agent finds its way.

BY MATTHIJS WOLTING / AUG 10, 2026

How we built Linear Agent

So rather than engineering a fixed path for Linear Agent, we defined the boundaries within which it could find its own.

Why it matters

— Five boundary surfaces: system prompt, tools, product model, run scope, custom harness

— Constraints live in tool design, not prompt prose — the tool can't overreach

— Coding agents get few deep tools; Linear Agent gets many shallow ones

— System skills disclose prompt fragments and tools progressively, preloaded by inference

— They refused the agent the SDK and GraphQL API — breadth traded for blast radius

Read on Linear →

Systems & primitives Critique & decision-making

DESIGN SYSTEMS COLLECTIVE

LDLR DESIGN FIELD NOTES

Three truths, no liar

Figma holds intent, Storybook behavior, production reality. Williams says reconciling them is a job nobody posted yet.

AUG 5, 2026

The Job Nobody Is Hiring For Yet

Evidence gets faster and cheaper. The verdict still needs someone who can be asked why, in public, and answer.

Why it matters

— The opening incident: Figma showed four dropdown actions, Storybook four, production two

— None of the three lied — engineering removed actions that generated support tickets

— AI makes cross-checking the layers near-free; deciding which layer governs stays human

— Smarter or rotting? Telling those apart is judgment, and judgment needs an owner

— Most real design-system teams run three to five people, understaffed by their own report

Read on Design Systems Collective →

Systems & primitives Critique & decision-making Hiring & team

PROOF OF CONCEPT

LDLR DESIGN FIELD NOTES

The next user isn't human

Hoang names the discipline forming around agents as users: access, context, tools, and parallel paths for two minds.

AUG 2, 2026

Agent Experience

Why force an octopus to open a jar with two tentacles because you as a human have two hands?

Why it matters

— Agents are users now: design for access, context, tools, orchestration

— Same goal, parallel paths: search bar for humans, direct query for agents

— Agents don't complain; observability must surface retries, dead ends, silent failures

— Documents go bidirectional: hierarchy for people, structure and metadata for agents

— He asks his agents to run a retro and file feedback

Read on Proof of Concept →

Systems & primitives Prototyping & exploration Cross-functional & process

DESIGN SYSTEMS COLLECTIVE

LDLR DESIGN FIELD NOTES

The rulebook compiles

Williams, from Morales’s talk: latency was your reviewer. When agents erase the wait, governance ships as an output.

JUL 13, 2026

Governance Becomes an Output

Governance doesn't have to be a manual fight. It just has to be encoded.

Why it matters

— Slow production looked like discipline; mostly it was latency doing review

— Remove the clock and the checking goes undone, loudly, all at once

— Morales's build: encode conventions into agents so governance ships as output

— Infrastructure can only encode what the team actually agreed on first

— Ask first: does your system know what it is?

Read on Design Systems Collective →

Systems & primitives Critique & decision-making

NIELSEN NORMAN GROUP

LDLR DESIGN FIELD NOTES

Context, not documents

Every prompt is now a design decision. Alicea’s answer: stop persuading humans and curate context that steers models.

JUL 24, 2026

UX-Context Design: Using UX Knowledge to Inform AI-Generated Design

Its success is measured by whether AI output improves, not by whether stakeholders are convinced by it.

Why it matters

— When PMs and engineers prompt, every generation is a design decision

— Personas persuade humans; models need the underlying reasoning instead

— Proposes UX.md beside DESIGN.md: research synthesis, glossary, user and world models

— Research findings become generation constraints, not insights in a report

— Continuously curated beside the code; there is no handoff moment

Read on Nielsen Norman Group →

Systems & primitives Research & validation Cross-functional & process

LINEAR

LDLR DESIGN FIELD NOTES

Wound up once

Loop engineering becomes team infrastructure: describe a recurring job in plain language and the agent runs it.

BY NAN YU / JUL 20, 2026

Introducing Loops

Describe the job in plain language and then choose whether it should run on a schedule or in response to an event.

Why it matters

— Loop engineering jumps from individual developer practice to team infrastructure

— Plain-language job descriptions become scheduled or event-driven agent operations

— Runs from full workspace context: issues, projects, documents, connected tools

— Triage loops can review code and start a fix before an engineer arrives

— Shared visibility: anyone can read instructions and inspect every run

Read on Linear →

Cross-functional & process Systems & primitives

DESIGN SYSTEMS COLLECTIVE

LDLR DESIGN FIELD NOTES

Execution got cheap

A designer watches AI finish outcomes without touching why they exist, and calls interpretation the scarce skill.

JUN 22, 2026

AI Didn't Break Design. It Exposed It.

Execution is becoming cheap and abundant. The real scarcity is understanding.

Why it matters

— AI slipped into design's own layer: pattern recognition and structured decision-making

— Treating AI as a tool locks attention on prompts; it is a meaning-making layer

— Context can be engineered; intent stays unspoken and is harder to construct

— Design systems shift from systems of appearance toward systems of understanding

— The value question isn't which tool next, it's where your value sits now

Read on Design Systems Collective →

Systems & primitives Critique & decision-making

BEN CALLAHAN

LDLR DESIGN FIELD NOTES

The skipped conversation

Callahan splits product quality in two: the design system sets the floor, and how product teams use it sets the ceiling.

JUL 5, 2026

What is product craft in the age of AI and design systems?

The worst product teams can make awful experiences with the best design systems.

Why it matters

— Craft is the skill to deliver higher quality over time; quality is the thing you'd measure

— A system raises the floor, but restraint and knowing when to deviate set the ceiling

— Four scenarios: only high craft on both sides produces a high ceiling

— He reverses himself, since enable-over-police left system teams out of the quality conversation

— The loop: co-define quality, craft toward it, evaluate, bake the wins into the floor

Read on their blog →

Systems & primitives Craft & detail Cross-functional & process

DESIGN SYSTEMS COLLECTIVE

LDLR DESIGN FIELD NOTES

Encode intent

When agents author your design system, intent must live in tokens and metadata, not only in a practitioner's head.

JUN 29, 2026

The System Knows What You Built. Does It Know What You Meant?

AI has lowered the floor on execution. The ceiling on judgment is still yours.

Why it matters

— The question moved from can AI use your system to can it write it.

— Agents connected to metadata make decisions about what to reach for.

— Discoverability isn't documentation coverage; it's intent legibility where decisions get made.

— Tokens should carry purpose, not just values; metadata should explain what's for.

— Unresolved ambiguity gets filled by the agent's best guess, everywhere, unnoticed.

Read on Substack →

Systems & primitives Critique & decision-making

FIGMA

LDLR DESIGN FIELD NOTES

Agents people can trust

Sheta Chatterjee walks the design calls behind Gemini Enterprise: a shared AI inbox and people kept in control.

BY SHETA CHATTERJEE / APR 29, 2026

How to design agentic tools for work

the user's focus should always be on their goal, not on managing AI

Why it matters

— One brand language up top, real differentiation down at the feature level.

— The guiding rule: keep focus on the goal, not on managing AI.

— An AI Inbox shows what each agent is doing and needs.

— Chat threads become shared project spaces, so the AI joins the team.

— Transparency and an obvious way to intervene, since the stakes are real.

Read on Figma →

Systems & primitives Cross-functional & process Craft & detail

JAMIESON ROTHWELL

LDLR DESIGN FIELD NOTES

Shipping binaries now

After three decades built on the handoff, Jamieson Rothwell ships working software and treats builds as design.

MAY 11, 2026

Product Design in 2026: Am I still a designer?

Shipping is the new designing.

Why it matters

— The old ritual, design then hand off blueprints, is over for him.

— LLMs and coding agents turn static design into active construction.

— Local-first as a design choice: his tool Warren guards data ownership.

— A small cross-platform system, Sitka, keeps React, HTML, and Swift aligned.

— Barrier to entry collapsed; the barrier to greatness is still taste.

Read on Medium →

Prototyping & exploration Systems & primitives Craft & detail

GEORGE XING

LDLR DESIGN FIELD NOTES

Shipping solo

A solo builder lays out the agent framework, test automation, and remote setup that made his product workflow far faster.

MAY 6, 2026

Prototyping & exploration Systems & primitives

JAKOB NIELSEN

LDLR DESIGN FIELD NOTES

Redesign the workflow

Citing a 515-startup field experiment, Nielsen argues AI pays off when teams redesign whole workflows, not single tasks.

APR 9, 2026

Cross-functional & process Systems & primitives

VIOLETTA BONENKAMP

LDLR DESIGN FIELD NOTES

Design you can execute

A founder argues the shift that matters is encoding brand rules and taste into files non-designers can run.

JUN 1, 2026

Systems & primitives Craft & detail

DESIGN MAVERICKS

LDLR DESIGN FIELD NOTES

Leadership literacy

Forty design leaders in Amsterdam land on one conclusion: the role dissolves while the capacity becomes foundational.

JUN 4, 2026

When AI Builds Everything, Design Becomes Everything

We are training people for a job. We should be cultivating a capacity.

Why it matters

— Intercom now ships over 90% of pull requests authored by Claude Code

— When execution costs nothing, judgment about direction becomes the product

— Atlassian runs 64% ICs, every designer a platform designer, demo before memo

— Frigieri's upstream conditions: workflows, data pathways, decision rights, controls

— The Frozen Middle: everyone experimenting alone without strategic guidance from above

Read on Design Mavericks →

Hiring & team Critique & decision-making Systems & primitives

GITHUB

LDLR DESIGN FIELD NOTES

Accessibility, shipped

GitHub’s five-year program turns outward: audit data, design tooling, and AI agents make inclusion operational.

BY ED SUMMERS / MAY 21, 2026

Building GitHub's next chapter in accessibility

Accessibility is never done.

Why it matters

— Audit analysis found 48% of issues preventable in the design phase

— Figma Annotation Toolkit documents accessibility intent before a line of code

— Copilot triage auto-populates 80% of metadata, cutting resolution time 62%

— Pull request files-changed page rebuilt accessibly across seven shipped updates

— Copilot CLI shipped screen reader mode and colorblind themes from day one

Read on GitHub →

Systems & primitives Craft & detail Cross-functional & process

GOOGLE LABS

LDLR DESIGN FIELD NOTES

Vibe design gets a canvas

Stitch grows into an AI-native infinite canvas: design agents, voice critique, and DESIGN.md moving between tools.

BY RUSTIN BANKS / MAR 18, 2026

Introducing "vibe design" with Stitch

Close the gap from idea to reality in minutes rather than days.

Why it matters

— Start from business objective and intended feeling, not a wireframe

— An agent manager tracks parallel explorations across the infinite canvas

— DESIGN.md exports design rules as agent-friendly markdown between tools

— Voice critique makes real-time updates while you stay in flow

— Stitch screens together and press play to preview full journeys

Read on Google Labs →

Prototyping & exploration Systems & primitives

POSTHOG

LDLR DESIGN FIELD NOTES

When FOMO ships early

PostHog's CEO on shipping AI before it worked, the architecture fix that turned it around, and betting the product on agents.

BY JAMES HAWKINS / MAY 1, 2026

PostHog's next chapter

FOMO came ahead of progress.

Why it matters

— A one-shot assistant shipped early, returned empty graphs, and felt half-useful.

— The fix was a single query loop so the agent could iterate.

— Agents now create most new dashboards, with MCP usage roughly doubling monthly.

— The bet: a headless, proactive product that ships fixes while you sleep.

— Hawkins is candid about UX churn, and why the short-term pain is worth it.

Read on PostHog →

Prototyping & exploration Critique & decision-making Systems & primitives

BOOTCAMP (MEDIUM)

LDLR DESIGN FIELD NOTES

Design relief

An early-career case for building leverage: encode taste into tools and remove friction over polishing more screens.

DEC 18, 2025

Stepping Into 2026 as a Designer (When You're Still Early in Your Career)

I wasn't designing outputs anymore. I was designing relief.

Why it matters

— Polished screens are cheap now, so execution alone no longer carries value.

— He stopped refining artifacts and started building small tools that remove friction.

— A junior's edge: no fixed identity, so quiet experiments come easier.

— Taste still matters, but it scales only once you encode it.

— The move is from producing screens to shaping how they get produced.

Read on Bootcamp (Medium) →

Systems & primitives Prototyping & exploration

DESIGN SYSTEMS COLLECTIVE

LDLR DESIGN FIELD NOTES

Readable by machines

Indeed’s build story: structured metadata over prose, and JSON five times cheaper than Markdown for the same query.

MAY 14, 2026

Fully Machine-Readable Design Systems

Why it matters

— AI reads your design system differently; structured metadata beats prose documentation.

— Foundations should be always-on; components served on demand through an MCP.

— At Indeed, JSON cost five times less than Markdown for the same query.

— Progressive disclosure maps what AI needs without flooding its context window.

— The system you wrote for designers won't work for agents without a rewrite.

Read on Design Systems Collective →

Systems & primitives Craft & detail Research & validation

DESIGN SYSTEMS COLLECTIVE

LDLR DESIGN FIELD NOTES

Legibility is the work

Agents expose every gap you papered over. Williams asks whether your system can explain itself to an LLM.

MAY 18, 2026

Your System Can't Explain Itself. That's the Problem.

Why it matters

— Most design systems are a shared palette and a Confluence page nobody trusts.

— A real system enforces its contracts; agents expose every gap you papered over.

— Can your system explain itself to an LLM without hallucinating your brand?

— AI didn't create this problem; it's revealing the documentation debt you already had.

— Legibility is the work. Not the components. Not the tokens. The clarity.

Read on Design Systems Collective →

Systems & primitives Craft & detail Storytelling & narrative

BOOTCAMP (MEDIUM)

LDLR DESIGN FIELD NOTES

shadcn won by default

shadcn sits behind every AI-generated UI. A developer built it, and most design teams never noticed the competition.

MAY 11, 2026

The most important Design System in 2026 that designers missed was built by a developer.

Every shadcn project is a fork.

Why it matters

— AI tools converged on shadcn as the default shape of generated UI.

— It lives in GitHub and React templates, not Figma, so designers missed it.

— The MCP server turned accidental convergence into installed infrastructure.

— The work isn't fighting the default. It's reading it, then deciding what to override.

Read on Bootcamp (Medium) →

Systems & primitives Critique & decision-making

DESIGN SYSTEMS COLLECTIVE

LDLR DESIGN FIELD NOTES

Agents don’t read docs

Agents read structure and naming, not docs or goodwill, and the informal contract becomes a hard boundary.

MAY 11, 2026

Systems & primitives Craft & detail

MICROSOFT DESIGN

LDLR DESIGN FIELD NOTES

The system is the handoff

Microsoft folded Copilot into four surfaces tied by one ‘Throw and Catch’ pattern: the pattern is the system.

BY JON FRIEDMAN / MAY 12, 2026

A simplified system

Intelligence without continuity can feel like an interruption instead of a thought partner.

Why it matters

— Four canonical Copilot surfaces, not scattered features per app.

— 'Throw and Catch' governs handoff between the four surfaces.

— Context persists across apps. The conversation does not restart.

— On-canvas narrows from exploration to document focus on selection.

Read on Microsoft Design →

Systems & primitives Prototyping & exploration Craft & detail

GITHUB (ENGINEERING & DESIGN)

LDLR DESIGN FIELD NOTES

Instant is the spec

GitHub Issues reframed performance around distribution shape, then paid 4.7% server/cache divergence to feel instant.

BY ALEXANDER LELIDIS / MAY 14, 2026

From latency to instant: Modernizing GitHub Issues navigation performance

We are willing to serve data that may be slightly stale if that allows the navigation to complete near instantaneous

Why it matters

— Reframed the goal: shape the distribution, not chase p99.

— Five layers: in-memory cache, IndexedDB, preheat, service worker, code-split.

— Accepted 4.7% server/cache divergence as the operating envelope.

— Cache hit went from 33% to 96%. P10 dropped to 70ms.

Read on GitHub (Engineering & Design) →

Craft & detail Systems & primitives Critique & decision-making

DESIGN SYSTEMS COLLECTIVE

LDLR DESIGN FIELD NOTES

AI didn't add to the spec

AI-readiness is the old spec finally enforced: tokens, layout intent, and governance now have to read to a non-human.

APR 27, 2026

Legible by Design: What AI Readiness Actually Demands of a Design System

The best systems are the ones that remain coherent even when the people who built them have moved on

Why it matters

— AI doesn't need a new system. It exposes the broken one.

— Five load-bearing properties: tokens, layout intent, a11y, governance, handoff.

— Legibility to machines reveals what was always opaque between humans.

— Stewardship beats generation. The system has to outlive its authors.

Read on Design Systems Collective →

Systems & primitives Critique & decision-making Cross-functional & process

WEBFLOW

LDLR DESIGN FIELD NOTES

Where AI debt lands

AI code defers cost instead of cutting it: seven debt categories land far from where the velocity gain was booked.

BY ANKIT AGRAWAL / MAY 12, 2026

The cleanup cost of ungoverned AI-generated code

It costs a contributor five minutes to generate a low-quality AI pull request, and hours for the maintainer to verify and reject it.

Why it matters

— Seven debt categories, segmented by who actually pays the bill.

— Time-to-exploit window collapsed from months to days.

— Five minutes to generate a PR. Hours to reject one.

— Triage PRs by risk class. Output is the wrong metric.

Read on Webflow →

Critique & decision-making Cross-functional & process Systems & primitives

FIGMA

LDLR DESIGN FIELD NOTES

The canvas reads code

McMillin reframes canvas, code, and PRD as one system: an agent translates, and judgment moves through both.

BY BRETT MCMILLIN / APR 30, 2026

Workflow lab: Expanding the canvas with Figma MCP

The artifact isn't a bug report. It's a shared reference for a design conversation.

Why it matters

— Agents surfaced ten product states the original canvas missed.

— Visual diff replaces the bug ticket as a design artifact.

— On canvas exploration stays cheap. In code every direction costs.

— Source of truth becomes connected artifacts, not one canonical file.

Read on Figma →

Prototyping & exploration Cross-functional & process Systems & primitives

MICROSOFT DESIGN

LDLR DESIGN FIELD NOTES

Paper earns its geometry

Microsoft spent years replacing plastic with paper. Committing was easy; making every millimeter behave was the work.

APR 21, 2026

Paper, please: The box that did more with less

A variance of one to two millimeters risked scratching the device or allowing damage during transit.

Why it matters

— Foam absorbs impact by being soft. Paper does it by structure.

— Over 1,000 prototype boxes to get coatings, paper, and inks working.

— Smaller boxes per pallet linked sustainability and shipping cost.

— Measurement ran as a second project tracking progress across thousands.

Read on Microsoft Design →

Cross-functional & process Craft & detail Systems & primitives