Claude's composer teardown
Updated June 12, 2026
Web search stays on by default, so you just type. Model and effort sit on the right, separate from + on the left. That keeps simple prompts simple and makes it clear what you are using before you send.
Calm default
What works
- No tool vocabulary on first load. You can just type.
- Starter pills (Write, Learn, Code, Life stuff, Claude choice) teach common jobs without opening the + menu.
- + on the left and model on the right keep behavioral toggles separate from which model answers.
- Sonnet 4.6 and Low effort show before the first keystroke. Price-sensitive users see what they are spending on; everyone sees model tier and upgrade path in one control. ChatGPT keeps upgrade as a separate upsell with no tie to what you are using now.
What we would push on
- Starter pills and the + menu are two front doors. Some users may never open + for skills, connectors, or screenshot.
- Mic and waveform sit as plain icons side by side on the right rail. Dictation and voice conversation look the same until you tap. Conversation needs a clearer affordance (blue fill, wave styling).
Takeaway
Calm card, explicit model on the right, depth in +. Visibility on model and effort is a feature, not clutter.
+ menu
What works
- One + menu groups attach, agent tooling, and behavioral toggles. You pick intent, not infrastructure.
- Web search defaults on with no chip in the card. You just prompt. Claude skips the "is this a search?" split ChatGPT forces with a separate mode.
- Keyboard shortcut on attach rewards power users without adding another bar icon.
What we would push on
- The list got long: screenshot, project, skills, connectors, plugins, style. Organizing well is not the same as feeling light.
- Skills, connectors, and plugins sit beside attach rows with equal visual weight. Agent vocabulary in the composer may confuse casual chat users.
Takeaway
The + menu scales capability without bar clutter. Web search is the model: default on, no mode chip. The length of the list is the harder problem.
Attachments
What works
- The image sits inside the composer card above the prompt. You see what Claude will read before send.
- Send affordance appears when content is ready. Clear signal that the message is shippable.
What we would push on
- Images render as inline preview, not filename chips with type and size. Harder to scan when you attach multiple files.
- Attach still lives inside +, not on the bar. Sending a file is not an edge case.
Takeaway
Steal keeping multimodal context inside the card. Question preview-only treatment when users stack several attachments.
Skills submenu
What works
- Flyout keeps the parent + menu open. You do not lose context exploring one level deeper.
- skill-creator, Manage skills, and Add skill cover discover, edit, and extend without a separate settings page.
What we would push on
- Skills next to connectors and plugins makes the composer feel like an agent workbench, not a chat box.
- No preview of which skill is active before send.
Takeaway
Good flyout pattern for overflow. Pair it with in-card feedback when a skill or connector is armed.
Model picker
What works
- Outcome labels per model (everyday, complex, quick) teach the choice in plain language alongside version names.
- Model picker on the right rail stays separate from +. Behavioral toggles do not compete with which brain answers.
- Effort is one tap away from the model row instead of buried in account settings.
- Fable 5 and upgrade gates sit beside workhorse models. Model, price tier, and reason to upgrade in one picker, not a disconnected upsell.
What we would push on
- The bar shows "Sonnet 4.6" and "Low" side by side. Low reads like low performance, not low effort. Model and effort are different choices and need clearer labels.
Takeaway
Right-rail model choice is the right split from +. Explicit model, effort, and upgrade path beats a hidden upsell. Label effort so it does not sound like a downgrade.
Effort and thinking
What works
- Effort copy explains the tradeoff in human terms: thorough answers vs hitting limits sooner.
- Thinking toggle sits next to effort in the same panel. Related controls stay grouped.
What we would push on
- Effort and thinking both change how hard the model works. The composer does not spell out when to use which.
- Low, Medium, High in the effort menu sound like quality grades. Spell out "effort" or token use so they do not read as model tiers.
Takeaway
Plain-language effort copy in the submenu is worth stealing. Fix the shorthand in the bar so Low never sounds like a weak model.
More models
What works
- Legacy and alternate Opus versions tuck behind More models. The first screen stays shorter.
- Same flyout pattern as skills: parent menu stays open while you browse.
What we would push on
- Overflow models show version numbers without the everyday or complex labels from the main picker.
Takeaway
Overflow keeps the first screen short. Match the outcome labels from the main list when you tuck models away.
Voice dictation
What works
- Transcription appears inside the composer card. You can edit before confirm.
- Waveform plus X and check give a clear review step instead of auto-send on silence.
What we would push on
- Dictation replaces the text area content. Attaching files or toggling web search mid-dictation is unclear.
Takeaway
Steal the confirm step and in-card transcription. Differentiate the two voice entry points before users have to learn by trial.
Voice listening
What works
- Listening state is unmistakable once you are in: blue treatment, label, and a dedicated Stop button.
- Conversation mode feels live. Dictation is transcribe-then-confirm. Different jobs once active.
What we would push on
- Blue glow only appears after you tap waveform. On the bar, both voice icons still look the same.
- Stop vs cancel vs check: three exit paths across two screens. Document the mental model or simplify.
Takeaway
Conversation mode earns its blue treatment in-session. The entry icons on the bar still need to teach the split upfront.
How it fits together
What works
- Calm card first, then + for attach, search, skills, and overflow agent tools.
- Web search on by default, no chip. Most prompts stay simple.
- Model and effort on the right rail, visible before send for spend control and upgrade clarity.
- Voice adds a review step before send instead of piping audio straight into the thread.
What we would push on
- Skills and style armed in + often have no in-card feedback before send.
- Thinking adds a third lever beside model and effort on the right rail.
- Starter pills and + are parallel discovery paths with no bridge between them.
Takeaway
Claude splits capability well: + for extras, right rail for model brain. Web search shows the right default: on, no mode chip. Skills and style still need clearer armed-state feedback.
Steal this
- Web search default on, in-menu toggle only, no chip splitting prompts into a search mode
- Flyout submenus that keep the parent open (skills, more models)
- Model and effort visible in the composer for spend control and upgrade clarity
- Outcome labels on models and plain-language effort copy
- Voice transcription with confirm before send
Skip this
- Forcing web search into a separate mode chip when most prompts should just work
- Long + menus with no in-card feedback when skills or style are armed
- Upgrade upsells disconnected from the model the user is on today
- Effort labels that read like model quality when shown beside the model name
- Two voice entry icons that look identical until the user taps
- Inline image preview only when users need multi-file chip scanning
Original gallery pages: Tool Switching in Composer · Context Chip Management