GLM (Z.ai) composer UX: Magic Design, Deep Think & Agent
GLM (Z.ai) is a builder, not a messaging box. The serif headline asks “What can I build for you?” Three pills under the bar (Magic Design, Full-Stack, Write Code) swap the example gallery. Search and Advanced Search hide behind the globe. Switching to GLM-5.2 replaces those pills with artifact categories and parks Deep Think Max on the right. Agent is a separate home with IM, Writing, Data Insight, and Discord / Lark / WeChat connectors.
Calm default
What works
- Headline names the job: build, not chat. Users do not have to guess whether this is a coding product or a general assistant.
- Three outcome pills sit under the bar: Magic Design, Full-Stack, Write Code. Those are deliverables, not model names.
- The card stays typeable. +, globe, and the atom sit inside the bar. Send stays disabled until there is text.
- GLM-4.7 in the header is already visible. Casual send does not require opening a picker.
What we would push on
- The subhead is slogan copy. “Boundless creative world” does not help first send.
- Globe and atom have no labels until hover. Search is easy to miss.
- ZCode and API in the header are extra product doors on a first-visit composer.
Business strategy
GLM (Z.ai) is selling itself as something that ships sites, decks, and code, not a cheaper ChatGPT. The home has to advertise those jobs before the first keystroke.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| Builder headline plus three job pills on a familiar chat bar | Artifact catalog visible without opening + | Slogan subhead; unlabeled icons; extra header products |
Takeaway
If the product ships artifacts, say build in the headline and put job pills under the bar. Keep the default card typeable.
Search vs Advanced Search
What works
- Two options with outcome copy. Search is “Single-round search, quickly get information.” Advanced Search is “Multi-round search, in-depth research and analysis.”
- Advanced Search is a toggle, not a second destination. The empty composer stays calm.
- Nesting this under the globe keeps research depth off the first-visit pills.
What we would push on
- No Always-on, no domain controls, no cost preview. Single vs multi-round is a coarse switch for a product that also sells Deep Think.
- The menu has no selected check on Search itself. Off vs on for basic search is implied, not labeled.
Business strategy
Default single-round keeps answers cheap. Advanced Search is the research upsell without teaching a Search chip on home.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| Nested Search / Advanced Search instead of an in-bar Search chip | Calm default; research depth exists | Easy to miss; only two settings |
Takeaway
Keep web search as a nested policy with outcome copy. Label the off state if basic search can also be disabled.
Pattern: Input Mode ToggleSearch is a nested policy with outcome copy: single-round vs multi-round, not a glowing chip on the default bar.
Pattern: Tool Switching in Composer
Advanced Search on
What works
- The globe turns blue when search is armed. That is the only in-bar signal, and it is enough if you already know the icon.
- The toggle sits next to the depth copy, so on/off and “multi-round” are one decision.
What we would push on
- No removable chip. Once the menu closes, Advanced Search is a colored icon. Easy to forget it is on.
- Magic Design / Full-Stack / Write Code stay on screen. Search depth and build mode are two independent systems with no explanation of how they combine.
Business strategy
A quiet armed state lets power users keep research on without crowding the builder pills. Casual users never see a Research mode they did not ask for.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| Blue globe as the only armed-search indicator | Bar stays uncluttered | No chip to dismiss; easy to leave Advanced Search on by accident |
Takeaway
If a nested toggle changes spend or latency, show a removable chip after the menu closes.
Pattern: Input Mode Toggle
Magic Design
What works
- The pill moves into the bar. Mode is now a chip you can see before send, not only a row underneath.
- The gallery swaps to visual work: neon campaign, recipe grid, hiring poster. Examples match the job.
- Placeholder stays “How can I help you today?” so typing still works if you ignore the cards.
What we would push on
- The cards have no obvious click target or “use this look” affordance. They read as marketing, not starters.
- Placeholder does not change. Kimi and Manus rewrite the prompt copy to the job. GLM (Z.ai) does not.
Business strategy
Magic Design is the consumer creation door. Featured work has to look like a studio, not a chat log, or GLM loses to Midjourney-shaped tools.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| In-bar Magic Design chip plus a visual gallery | Mode is visible; examples sell the job | Placeholder still generic; cards may not be prompt starters |
Takeaway
When a mode is visual, pull the pill into the bar and swap the gallery. Change the placeholder to a design job, not a chat greeting.
Pattern: Tool Switching in Composer
Pattern: Prompt Starters
Full-Stack
What works
- Full-Stack chip sits in the bar. Same pattern as Magic Design, so the three pills feel like one system.
- Gallery shifts to apps: personal blog, podcast generator, pixel-art game. The examples are products, not posters.
What we would push on
- Card titles truncate (“Personal blog web developm…”). You cannot tell stack or outcome without opening one.
- No Plan vs Build split. Lovable forces that before the agent runs. GLM (Z.ai) treats Full-Stack as a gallery theme.
Business strategy
Full-Stack is the Lovable/Replit fight. Showing shipped-looking apps under the bar is how GLM claims it can finish a site, not only talk about one.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| Full-Stack as an in-bar chip plus app gallery | Same disclosure pattern as Magic Design | Truncated titles; no scope controls before send |
Takeaway
If a mode claims to ship an app, show complete examples and set scope (pages, stack) in the bar, not only after generate.
Pattern: Tool Switching in Composer
Pattern: Prompt Starters
Write Code
What works
- The gallery becomes interactive pieces: tarot, Rubik’s cube, voxel scenes. Code mode sells toys and demos, not a blank IDE.
- Home still has a typeable bar. You can ignore the cards and paste a spec.
What we would push on
- The Write Code pill is not clearly in the bar in this state. Magic Design and Full-Stack get an in-bar chip; this mode looks closer to default home.
- No language, repo, or ZCode handoff in the composer. ZCode lives in the header as a separate product.
Business strategy
Write Code on chat.z.ai is a teaser for ZCode. The gallery has to look like shipped interactive work so the jump to the IDE product feels optional, not required.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| Code mode as a gallery of interactive demos | Sells GLM coding without leaving chat | Weaker in-bar chip than sibling modes; ZCode is a second door |
Takeaway
Sibling modes should share one chip pattern. If code has a real IDE, say when to stay in chat and when to open it.
Pattern: Tool Switching in Composer
Pattern: Prompt Starters
GLM-4.7 through 5.2
What works
- Each row has a version name and a one-liner. GLM-5.2 is “Flagship model, excels at coding and long-horizon tasks.”
- NEW on 5.2 and a check on 4.7 make current vs latest obvious.
- Vision gets its own row (GLM-5V-Turbo) instead of hiding behind a paperclip.
What we would push on
- Five GLM IDs. 5.2, 5.1, 5-Turbo, 5V-Turbo, 4.7. Casual users cannot tell Turbo from flagship without reading every line.
- No latency, cost, or rate-limit preview. “Previous flagship” is history, not a reason to pick it.
- The bar still says GLM-4.7 while 5.2 is marked NEW. The picker has to do all the upgrade work.
Business strategy
GLM ships version numbers as the brand. The picker is a launch surface for GLM-5.2, with 4.7 kept as the familiar default so existing users are not yanked onto a new flagship.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| Versioned GLM list with NEW badge and job one-liners | Flagship launch is visible; classic default stays | Raw IDs; no cost or speed; five near-duplicate names |
Takeaway
If version numbers are the brand, keep them, but lead with job copy and show what the new flagship unlocks in the bar, not only in the menu.
Pattern: Model Selection UITiers mix job copy (coding, long-horizon, chat, vision) with raw GLM version IDs and a NEW badge on the flagship.
Pattern: Cost Transparency
GLM-5.2 now available
What works
- The modal lists jobs, not parameters: AI Slides, Better Artifacts, Long-document Boost, Fluent Long-form Writing, Immersive Roleplay.
- “Model Upgraded / GLM-5.2” on the left is a poster. Next on the right is a single path through the news.
- The dimmed home still shows GLM-4.7 in the header, so you know what you are leaving.
What we would push on
- Next without Skip or “keep 4.7” on the same row. Closing the X is the only decline, and it is easy to miss.
- The bullets do not map to the three home pills. Slides and artifacts are new jobs the GLM-4.7 bar never named.
Business strategy
A model launch has to sell new artifacts or it is just a changelog. GLM uses the modal to re-pitch the product as slides, pages, and long docs, the same jobs Kimi puts on home pills.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| Blocking feature modal with artifact bullets and Next | Names what 5.2 is for | Interrupts first prompt; weak keep-current path |
Takeaway
When a flagship ships, list the new jobs in the modal and give a one-click stay-on-current control next to Next.
Pattern: Prompt Starters
GLM-5.2 home
What works
- Deep Think Max lands on the right of the bar. Reasoning depth is now a default control, not a hidden globe setting.
- The three builder pills are gone. In their place: Landing Page, Knowledge/Teaching Material, 3D Modeling, Mini Game, Personal Blog.
- Example cards match the selected chip. Landing Page shows fitness, editorial, and portfolio sites.
What we would push on
- This is a different product home, not a theme on the same bar. Magic Design / Full-Stack / Write Code disappear with no map back.
- Category chips are content types. Builder pills were jobs. Users who learned the GLM-4.7 row have to relearn the 5.2 row.
- Deep Think Max is jargon until you open the menu. Max of what is unclear on the collapsed label.
Business strategy
GLM-5.2 is positioned as the artifact model. The home stops advertising three studio modes and starts browsing finished page types, closer to a site generator than a chat composer.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| Flagship home replaces builder pills with artifact categories and Deep Think | 5.2 looks like a new product; reasoning is visible | Two homes, no legend; Max is unexplained on the chip |
Takeaway
If a model switch rewrites the composer, say what you gained and keep one familiar control so the bar still feels like the same app.
Deep Think High vs Max
What works
- High and Max sit in one menu with a check on Max. Effort is a two-rung ladder, not a model ID.
- Deep Think is a separate toggle above the rungs. You can kill reasoning without leaving GLM-5.2.
- The control lives in the composer, same place DeepSeek puts DeepThink and Claude puts effort.
What we would push on
- High and Max have no outcome copy, latency, or cost. Claude’s effort rows say what you get. These rows are names.
- Deep Think on plus Max selected is two dimensions (on/off and height) that collapse into one chip label. Easy to think Max is a model.
Business strategy
Reasoning is how GLM-5.2 justifies the flagship. Putting High/Max on the bar advertises depth at send time, when a hidden thinking toggle would not.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| In-bar Deep Think toggle with High / Max rungs | Reasoning is opt-out on the flagship; effort is one click | No outcome copy or cost; chip reads like a model name |
Takeaway
Put thinking effort in the bar on the flagship. Add one line of job copy per rung so High vs Max is not a coin flip.
Pattern: Model Selection UI
Pattern: Cost Transparency
Artifact category chips
What works
- The selected chip fills grey. The gallery swaps to lessons: Rust systems, Axioma math, calculus notes.
- Cards look like real pages, not prompt tiles. That matches the “Better Artifacts” promise from the launch modal.
- Landing Page stays in the row, so you can jump categories without leaving home.
What we would push on
- Five chips already overflow into Personal Blog. Mini Game and 3D Modeling are easy to skip.
- Clicking a chip does not change the placeholder. You still type into “How can I help you today?” while looking at a calculus textbook.
Business strategy
Category chips are a catalog of artifacts GLM-5.2 can ship. Teaching material is a wedge for long-document and tutoring work that 4.7’s three studio pills never showed.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| Category chips that swap example galleries | Browse by artifact type; selected state is obvious | Crowded row; placeholder ignores the category |
Takeaway
If chips pick an artifact type, rewrite the placeholder to that job. Keep the row short enough that the last chip is still visible.
Pattern: Tool Switching in Composer
Pattern: Prompt Starters
Separate Agent home
What works
- Headline switches to “Create anything you can imagine.” The surface admits it is a different product.
- Four mode pills: IM (selected), Full-Stack, Writing, Data Insight. IM is the chat-app default.
- Connector cards sit under the bar: Discord, Lark, WeChat, Telegram. Connect is the action, Coming soon is honest on Telegram.
What we would push on
- Chat vs Agent is a sidebar icon swap with no label on the collapsed rail. Easy to land here by accident.
- Full-Stack exists on both homes. On Chat it is a gallery theme. On Agent it sits next to IM. Same word, two systems.
- IM as a pill name is insider copy. Discord users might get it. Everyone else sees an acronym.
Business strategy
GLM (Z.ai) wants to live in the apps people already use. Agent home is a connector storefront, not a better composer. That is a distribution bet, closer to Manus suggestion cards than to Kimi’s in-bar pills.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| Agent as a second home with IM modes and connector cards | Chat apps are visible; Coming soon is labeled | Two products, one URL; Full-Stack named twice; IM is jargon |
Takeaway
If Agent is a different job, label the rail and keep mode names from colliding with Chat. Put connectors on the Agent home, not buried in settings.
Pattern: Tool Switching in Composer
Pattern: Prompt Starters
Connect Discord
What works
- The modal is one field: Bot Token, with show/hide. Cancel and Save are standard.
- “Set up on the Discord Developer Portal” is an external link, so the hard part is not hidden inside GLM copy.
What we would push on
- Pasting a bot token is a developer ritual. Manus and ChatGPT connectors start from OAuth. This will bounce anyone who is not already running a Discord app.
- No preview of what the bot will be allowed to do. Save is trust-me.
Business strategy
A token modal ships fast and works for builders who already have a bot. It does not get GLM into a community the way one-click OAuth would. Fine for a preview. Wrong as the only Connect path.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| Discord Connect as paste-your-bot-token | Ships without OAuth app review; portal link exists | Developer-only; no permission preview |
Takeaway
If Connect is on the Agent home, use OAuth for the default path. Keep token paste as an advanced fallback.
Steal this
- What can I build for you as a builder headline, not a chat greeting
- Three outcome pills that pull into the bar and swap the gallery (Magic Design, Full-Stack, Write Code)
- Search vs Advanced Search as a nested globe menu with outcome copy
- GLM picker with NEW on the flagship and a one-line job under each version
- Launch modal that lists new artifact jobs (slides, pages, long docs), not parameter charts
- Deep Think High / Max in the flagship bar, with a toggle to turn thinking off
- Category chips that swap real-looking example pages
- Agent as a second home with connector cards and Coming soon labeled
Skip this
- Slogan subheads like boundless creative world
- GLM-4.7 and GLM-5.2 homes that rewrite the pills with no map between them
- Full-Stack as both a Chat gallery theme and an Agent mode
- IM as a pill label
- Connect Discord by pasting a bot token as the only path
- A blue globe as the only sign Advanced Search is on
- High vs Max with no outcome copy or cost
How others design the composer
Same job, different product bets, and what each tradeoff reveals.
Compare composer UX across products
ChatGPT, Claude, Perplexity, and Gemini side by side: default bar, tools, cost, and what to steal.
KimiKimi’s home pills rewrite the whole bar into Docs, Sheets, Websites. GLM-4.7 pills mostly swap the gallery and pull a chip into an otherwise unchanged card. GLM-5.2 then throws those pills away.
Read teardown
ManusManus puts connector cards on the assign-a-task composer. GLM (Z.ai) parks Discord, Lark, and WeChat on a separate Agent home, then asks for a bot token.
Read teardown
DeepSeekDeepSeek keeps Instant/Expert and DeepThink on one card for every user. GLM (Z.ai) hides Deep Think until you switch to GLM-5.2, then makes Max the default chip.
Read teardownUseful for a critique or spec? Share it.
Original gallery pages: Tool Switching in Composer