Choosing between the six modes

Each mode routes through different services. See which feature uses which model.

The six modes

Chat Claude / GPT / Kimi / Gemini

The default for anything whose output is text: research, comparison, summarising, planning, drafting. When in doubt, start here.

Three things reliably improve the result:

• Specify the shape of the output — “as a comparison table”, “five bullets”, “as the body of an email”
• Hand over context as equipment — anything you keep re-explaining belongs in Knowledge
• Give it a role — equipping the reviewer role keeps critique consistent (AI team)

Image Prompt OpenAI gpt-image

For generating and editing images. Describing subject, background, mood and aspect ratio — in that order — works well.

While an image is being generated you see the same “Generating image…” card as for video, with the elapsed time. When it finishes, the card becomes the image. If the run ends without an image, or you cancel, the card settles to “no image was generated” / “cancelled” instead of spinning on.

1. Set the mode to Image.

2. Describe it including the ratio, e.g. “product shot against a blue sky, 16:9”.

3. If you have a reference photo, attach it in the composer before sending.

Generated images land in the media pool and can be used directly in the editor. See Create images, video and audio for more.

Video Claude Higgsfield

For generating video. Length, orientation, camera movement and mood all help.

Generation consumes Higgsfield credits. You'll see an estimate before it runs and approve it first. See Create images, video and audio.

Article Claude / GPT / Kimi / Gemini

For structuring and writing SEO articles. Give the reader and the goal alongside the topic — that's what lifts quality.

1. State topic, reader and goal
   e.g. “How to sleep better, for people in their 30s–50s. Draft an SEO structure and body.”

2. Lock the outline first
   Get the headings, fix the order and granularity, then move to the body.

3. Write the body
   For long pieces, work heading by heading — it holds together better.

4. Export it
   📤 Export this thread in the thread menu saves it as Markdown.

Audio ElevenLabs

For narration and music. What you type means different things for each:

• Narration — the exact script you want read aloud
• Music — a description of the track (e.g. bright and upbeat, piano-led)

Results go to the media pool and can be placed on the narration or music lane of a composition. Requires an ElevenLabs connection (setup).

Code Claude / GPT / Kimi / Gemini

For implementation, fixes and investigation. Equip the code project first.

1. Register the project under Code in the sidebar (Code projects).

2. Equip it to the thread and set the mode to Code.

3. Say what you want and what the result should be, e.g. “fix the bug in this function”.

Web research switches to a cheaper model on its own

Looking things up is about finding and gathering, so it needs a lot of attempts rather than the reasoning power of a top-tier model. Leaving Opus or Fable on for research burns through tokens fast, so research turns only switch to a lighter model automatically.

The switch lasts for that turn alone. Your thread and settings are left untouched, so the next request goes back to your usual model without you doing anything.

The badge under each answer tells you which model ran. When a switch happened it carries 💰 saved, and hovering it explains what changed to what, and why.

If you'd rather it didn't, turn off Use a cheaper model for web research under Settings → Models (it's on by default).

Modes versus engines

A mode says what kind of work; the engine says who does it. Keeping both in mind avoids surprises.

Default models per mode are set under Settings → Models.