Insites Studio®Docs

Command Palette

Search for a command to run...

Use the AI prompt

Get the most out of Studio's AI editing surface.

The AI in Studio is a Claude model with two tools — write_file and edit_file — and a system prompt that knows about Insites' Liquid conventions and your project state. It's strong at code edits when you give it specific, scoped requests; it's weaker when you give it vague aspirational ones. This guide is a practical playbook.

Be specific about scope#

Bad: "Make the homepage better."

Good: "On the homepage hero, replace the 'Get started' CTA with a two-button group: 'See pricing' (primary) and 'Book a demo' (secondary). Keep the same alignment and spacing."

The AI's first job is to figure out what you want changed. The more specific you are about what to change and what to leave alone, the fewer iterations you spend correcting drift.

Reference files by path#

The AI sees your repo, but it reads your prompts more carefully than it reads the repo. Naming the file you're talking about saves it from guessing.

Bad: "Fix the navigation."

Good: "In app/views/partials/site/header.liquid, the dropdown menu doesn't close when you click outside it. Add an outside-click handler."

Use the selected element#

If you click an element in the preview pane, the editor passes that selection as context to the AI. Then "change this button's color to ins-blue" works without you naming the file.

Reject early, reject often#

A bad patch isn't a problem. A bad patch you didn't reject is. Reject as soon as you spot a misstep — your next prompt can correct course while the conversation context is fresh.

The Pending Patches card lets you reject patch-by-patch within a single AI message. You don't have to take or leave the whole turn.

Chain, don't dump#

When you have multiple changes, chain them in successive turns instead of cramming everything into one message. You'll get cleaner diffs and tighter scope.

Bad (one turn):

Change the hero copy, add a pricing page, remove the testimonials section, add a footer with social links, and refactor the nav to use the design system tokens.

Good (chained):

  1. "Change the hero copy to: 'Build apps that ship.' Make the subhead match in tone."
  2. "Add a pricing page at /pricing using the same layout pattern as /."
  3. "Remove the testimonials section from the homepage."
  4. "Add a footer with Twitter, LinkedIn, GitHub icons linking out."
  5. "Refactor the nav to use the design system tokens — --color-ins-* for color, font-display for the brand mark."

Each turn is a focused diff. Each is independently revertable.

Pick the right model#

  • Haiku 4.5 — fast, cheap, very capable for routine edits. Default for iterative work where you're cycling through many small changes.
  • Sonnet 4.6 — the workspace default. Best balance for code edits where reasoning about structure matters.
  • Opus 4.7 — slowest and most expensive, but worth it for the rare task that needs deep reasoning across a lot of context (cross-cutting refactor, complex multi-file logic).

The model picker is in the chat header. Per-turn switches are fine; the workspace default is set in Workspace settings → AI defaults.

Approval mode and iteration speed#

If you're in Confirm mode and constantly clicking Accept-all on safe additive patches, switch to Auto mode. Additive patches will apply themselves; destructive ones still wait. You'll iterate faster without losing the safety net.

For exploratory throwaway work — sketching an idea, rapid prototyping — Dangerous mode is fine. Just don't leave it on by default.

See Approval mode for the full breakdown.

Tell the AI what to stop doing#

Claude takes direct instructions seriously. If it keeps suggesting something you don't want:

  • "Stop putting the email signup in the hero. Move it to the footer or omit it entirely."
  • "Don't refactor unrelated files when I ask for a styling change."
  • "Don't add comments to my Liquid files unless I explicitly ask."

Direct instructions stick across turns within a conversation.

Use the chat as a working memory#

You can ask the AI to:

  • "Summarise what we've changed so far this session."
  • "What's left on the todo list we agreed on?"
  • "Walk me through the homepage hero file as it stands now."

Treat it like a pair-programmer who's been watching. Get it to verbalise state when you're losing track.

When the AI is stuck#

If three turns in a row produce wrong patches:

  1. Reject everything and reprompt with more context.
  2. Provide an example: paste a snippet of the kind of code you want.
  3. Switch to a stronger model (Sonnet → Opus) for that turn.
  4. If the file is large, focus the prompt on a specific function or block.

Common questions#

Can the AI run my tests? Not today. It edits files; you run tests yourself.

Can it deploy? No. Deploy is always human-initiated.

Does it remember across sessions? No. Each Studio session starts a fresh conversation. Long-lived facts about the project belong in the project brief or in CLAUDE.md if you have one.