Threads
Threads and the composer
A thread is one piece of work. The composer holds the mode, the equipment, attachments and how you send. This page covers three things: giving an instruction, stopping it, and deciding how much to delegate.
Creating and tidying threads
＋New at the bottom of the sidebar → Thread.
To rename it, open the thread menu and choose ✎ Rename.
To take it out of the list, hover the row and press ✕ (moves it to the archive). Deleting permanently is in the menu.
Deletion can't be undone. If the thread has children you'll be warned that they will be deleted too. When in doubt, archive it instead.
What's in the composer
The left side is about context; the right side is about execution.
| Element | What it does |
|---|---|
| Mode | Switch between chat, image, video, article, audio and code. Choosing a mode |
| 📚 Knowledge, roles, connections | Equip knowledge, skills, roles and external connections to this thread |
| Attachments | 💻 Choose from this Mac / 🗂 Choose from the media pool. Drag and drop also works |
| AI engine | Switch Claude / GPT / Kimi / Gemini for this thread only |
| Send | Send the instruction. Switch to 🚀 Run as a team to hand it to the AI team |
Everything you equip appears as a chip above the composer, removable with ×. That row is always the answer to “what is the AI actually seeing?”
Wander off to another thread mid-sentence and your text is still there when you come back. It also survives closing the app, and restarts from an update. Each thread keeps its own draft, so nothing gets mixed up. Once you send and the box is empty, the draft is dropped.
Sending, and the send options
By default Enter sends and Shift+Enter adds a line break. You can swap this under Settings → Appearance & input → Send key by choosing ⌘+Enter to send (Enter for a new line).
The options next to the send button change how the instruction is handled.
| Option | What happens |
|---|---|
| Plan before executing | The AI presents a plan first and only proceeds once you approve it |
| Branch from here | Creates a child thread carrying this conversation — and your unsent draft — with it |
| Hand over to the AI team | Passes the conversation and this request to a team, where a commander organises it |
| Turn this into a skill | Distils the thread's workflow into a reusable skill |
| Turn this thread into a loop | Builds a brief for a recurring run out of this conversation |
| Hand over to an agent | Teaches what this conversation settled to an agent, in a form it can use next time |
For anything hard to undo — mass edits, publishing externally — Plan before executing is the safe choice. You get to read the intent before anything happens, and can redirect mid-way.
One request, several engines
Name two or more engines in a single request and it is split into steps that run in turn, each on the engine you named.
| Example | What happens |
|---|---|
| “Prompt by Kimi, review by Claude, image by GPT” | Three steps run Kimi → Claude → GPT. Each step receives the gist of the one before |
Progress shows on a card in the thread as ✅ / ❌. Only steps that send something outside (publish, post, email) pause for your approval first.
A step like “generate the image with GPT” runs in image mode for that step only, even inside a chat-mode thread (the step label reads “GPT · image”). If a step ends without producing an image, it is not marked done — the run stops and says the engine never reached image generation.
Video steps run on Claude even if you name GPT or Kimi, because video generation is wired through Claude only (the step says so).
Queueing while it works
You can keep typing while the AI is busy. The field reads You can send while it's working (queued and sent automatically when it finishes; ⚡ to interrupt)…
- Sending normally adds it to the queue, to be delivered when the current task ends.
- Use ⚡ when it needs to jump the line.
It means the next instruction you think of doesn't have to wait in your head.
Stopping, undoing, resetting
| What you want | How |
|---|---|
| Stop what's running | Esc — cuts off the AI mid-task |
| Take back what you just sent | Esc within 5 seconds of sending. Your text returns to the field |
| Reset the context | ⟳ Reset session in the thread menu |
Esc prevents further work, but files already written and messages already sent externally stay done. For anything irreversible, gate it with approval mode rather than relying on stopping.
Approval mode: how much runs automatically
Set under Settings → Safety → Approval mode for tool execution. It applies to both Claude and ChatGPT.
| Mode | Behaviour | When to use it |
|---|---|---|
| Ask every time | A dialog appears before generation and before touching folders outside the working directory | While you're still getting used to it |
| Auto approve | Automatic inside the working folder; still asks for anything outside | Day-to-day |
| Bypass | No confirmation at all; other folders are readable and writable | Only when you fully understand the blast radius |
The same Safety tab also carries policies passed to the AI at generation time:
- Confirm before sending confidential data externally
- Add an AI-generated label by default
- Use only commercially cleared material and wording
To auto-approve temporarily for one thread, use ⚡ Auto approve in the thread menu.
Work mode: run to completion
Turn on Run to completion in the composer and the AI resumes on its own whenever a reply ends mid-task. If it needs a decision from you, it stops and waits.
Auto-continued messages are labelled 🏁 Continuing the work (automatic), so they're distinguishable from your own instructions. When it stops, it tells you why.
| Message | Meaning |
|---|---|
| 🏁 Work mode: stopped because the AI is waiting on your answer | A decision is needed. Reply and it resumes |
| 🏁 Work mode: the work is complete | Finished |
| 🏁 Work mode: stopped after reaching the auto-continue limit ({n}) | The runaway guard. Send anything to resume |
| 🏁 Work mode: stopped because plan usage reached {p}% | Paused on usage limits |
Change the limit under Settings → Models → Auto-continue limit for work mode (times). Broader automation is covered in Let it run on its own.
The thread menu
| Item | What it does |
|---|---|
| 📖 Continue in a new chapter | Breaks a long conversation, carrying the essentials forward |
| 🎖 Create an AI team from this thread | Forms a team using this conversation as the brief |
| 🪄 Turn this thread into a skill | Saves the approach as a reusable skill |
| ✓ Return results to the parent | Sends a child thread's conclusion back up |
| 📤 Export this thread | Writes the conversation out as Markdown |
| 📂 Open the storage folder | Reveals this thread's folder in Finder |
Chapters, branching and returning results are explained in Growing context.