docs: add generative UI explainer article

Articles/generative-ui-text-output-not-enough.md

@@ -0,0 +1,198 @@
+# What Is Generative UI? And Why Text Output Is No Longer Enough
+
+For the last decade, most software automation has treated the interface as a
+fixed shell. A user clicks a known button, fills a known form, opens a known
+table, and the system responds inside layouts that were designed long before
+the request arrived. Even when language models entered the product stack, the
+default answer stayed the same: put a chat box next to the application and let
+the model return text.
+
+That was a useful starting point. Text is universal, cheap to render, and easy
+to stream. But it is also a narrow output format for many real tasks. When a
+user asks for a budget comparison, a migration plan, a sales pipeline review,
+a booking workflow, or a compliance checklist, they do not only need prose.
+They need structure, controls, state, and visual priority. A paragraph can
+describe those things. A generated interface can let the user act on them.
+
+Generative UI is the shift from models that only generate words to systems
+that generate usable interface states. Instead of responding with a wall of
+text that says “here are three options,” the application can render three
+option cards. Instead of describing filters, it can produce filter controls.
+Instead of summarizing a workflow, it can create a stepper with actions,
+warnings, and progress. The model is no longer just a narrator. It becomes
+part of the interface composition layer.
+
+## The problem with text-only output
+
+Text-only AI has three persistent problems in product workflows.
+
+First, text hides hierarchy. A model can write “high priority” or “low risk,”
+but users still have to scan paragraphs to understand what matters. Tables
+help, but plain markdown tables quickly collapse under complex data: nested
+statuses, actions, evidence links, confidence, responsibility, and exceptions.
+
+Second, text separates recommendation from action. A chatbot can advise a user
+to update a dependency, compare a plan, or inspect an invoice. The next step
+still lives somewhere else in the product. The user must translate the
+recommendation into clicks. That translation is a common source of abandonment
+and error.
+
+Third, text does not remember UI state naturally. Users expect software to
+preserve selections, expanded rows, input values, validation messages, and
+partial completion. A chat transcript is a poor container for that kind of
+state. The more interactive the job, the more awkward a pure chat UI becomes.
+
+These constraints are not cosmetic. They directly affect task completion. If a
+user has to copy details out of a model response, paste them into another
+screen, then manually verify whether the action succeeded, the model is doing
+only part of the work.
+
+## What generative UI changes
+
+Generative UI treats the response as an interface specification rather than
+only a message. The generated output can include components such as cards,
+tables, forms, timelines, charts, callouts, comparison grids, accordions, and
+action panels. Those components can be rendered by the application in a
+controlled design system, so the result still feels native rather than like
+arbitrary HTML pasted into a chat window.
+
+The key idea is not “let the model draw anything.” That is risky and
+inconsistent. The better model is constrained generation: the product exposes
+approved components and schemas, and the model chooses how to compose them for
+the user’s task. This gives teams a middle path between rigid hand-coded
+screens and unstructured text.
+
+For example, a support tool could expose components like `IssueSummary`,
+`RefundDecisionCard`, `EvidenceTimeline`, and `EscalationForm`. When a user
+asks “what should I do with this refund request?”, the model can render a
+structured decision view: the likely policy outcome, the evidence that
+supports it, the risk flags, and a pre-filled form for the next action. The UI
+is generated for that case, but the parts are still governed by the product.
+
+## Why now?
+
+Several trends make generative UI practical now.
+
+The first is stronger structured output. Modern models are better at following
+schemas, producing JSON-like structures, and respecting component contracts.
+That makes it possible to generate interface descriptions that a renderer can
+validate before showing them to users.
+
+The second is streaming. Users do not want to wait for a full page to be
+invented before anything appears. A generative UI renderer can progressively
+hydrate pieces of the interface as the model produces them: a summary first,
+then cards, then detailed rows, then optional actions.
+
+The third is design-system maturity. Many teams already have component
+libraries with buttons, forms, tables, badges, modals, charts, and layout
+primitives. Generative UI can reuse those assets rather than bypass them. That
+means product teams can keep brand consistency, accessibility rules, and
+interaction patterns while making interfaces more adaptive.
+
+The fourth is workflow pressure. Users increasingly expect software to
+collapse research, decision, and action into one flow. A static dashboard may
+show what happened. A generated interface can respond to what the user is
+trying to do next.
+
+## Where generative UI fits best
+
+Generative UI is most valuable when the user’s request is specific, the data
+is structured or semi-structured, and the next action benefits from visual
+organization.
+
+Good examples include:
+
+- Data analysis summaries that need charts, filters, and anomaly cards.
+- Internal operations tools where each case has different evidence and
+  required steps.
+- Developer tools that can render diffs, test failures, dependency risks, and
+  suggested commands.
+- Sales or customer success workflows that need account timelines, renewal
+  risk, and next-step forms.
+- Finance and compliance tools that need decision traces, confidence, and
+  audit-friendly outputs.
+- Education products that can create exercises, hints, progress checks, and
+  visual explanations.
+
+Generative UI is less useful when a response is truly simple. If a user asks
+for a one-line definition, a generated component tree adds unnecessary
+overhead. The point is not to replace every sentence with a widget. The point
+is to stop forcing complex work through a paragraph when an interface would be
+clearer.
+
+## The product design shift
+
+The biggest change is how teams think about screens. Traditional product
+design asks: “What pages do we need?” Generative UI asks: “What components and
+actions should the system be allowed to assemble for a user’s intent?”
+
+That changes the unit of design. Instead of designing every possible page
+state manually, teams design a governed component vocabulary. They define what
+a `RiskCard` can show, what fields an `ActionForm` accepts, which charts can
+be rendered, what validation is required, and when human confirmation is
+mandatory. The model can then compose those components dynamically.
+
+This does not remove designers or engineers. It gives them a different control
+surface. Designers define the visual grammar. Engineers define schemas, state
+handling, permissions, and validation. Product leads define which intents map
+to which workflows. The model fills the gap between rigid navigation and user-
+specific context.
+
+## Risks and guardrails
+
+Generative UI also introduces real risks.
+
+A model may choose the wrong component, omit an important warning, overstate
+confidence, or generate a flow that looks actionable but lacks the right
+business rules. That is why generative UI needs guardrails: schema validation,
+component allowlists, server-side permission checks, deterministic fallbacks,
+and clear separation between display and irreversible action.
+
+The model should be able to suggest an action panel. It should not be able to
+bypass authorization. It can draft a form. It should not submit a payment,
+delete data, or change account settings without explicit product-level
+controls. The generated interface must still respect the same security model
+as hand-coded UI.
+
+Observability also matters. Teams need to log which component schemas were
+generated, which validations failed, which actions users took, and where users
+abandoned the generated flow. Without that, generative UI becomes impossible
+to improve safely.
+
+## Why text is still part of the answer
+
+Generative UI does not make text obsolete. Text remains the best medium for
+explanation, nuance, and narrative. The better pattern is hybrid: short
+textual reasoning paired with structured components. A generated pricing
+comparison might include a two-sentence recommendation, a comparison table, a
+sensitivity slider, and an action checklist. Each part does what it is good
+at.
+
+This is the practical future of AI interfaces: not chat versus UI, but chat as
+one input mode into an adaptive interface layer.
+
+## What teams should build first
+
+A good first generative UI project is narrow and high-friction. Pick a
+workflow where users already copy model output into another tool or manually
+translate summaries into action. Define five to ten components that cover the
+common cases. Add strict schemas. Render progressively. Keep irreversible
+actions behind normal product confirmation. Measure whether users complete the
+task faster and with fewer errors.
+
+Do not start by asking a model to invent an entire application. Start by
+letting it assemble a better answer for one painful workflow.
+
+## Conclusion
+
+Generative UI is the next step after chat because many user intents are not
+naturally text-shaped. People ask for decisions, comparisons, plans, audits,
+and actions. Text can explain those things, but it often cannot carry them all
+the way to completion.
+
+The opportunity is to let models produce structured, validated, interactive
+interface states using the components teams already trust. Done well,
+generative UI makes software feel less like a set of static pages and more
+like a workspace that adapts to the job in front of the user.
+
+Text output was the demo. Generative UI is where the work starts to happen.