Prompts

A prompt in Catalyst is more than a saved instruction. It’s a reusable bundle you can attach to any conversation that carries its own system instruction, a knowledge base of documents, optional data/SQL, and tools — so selecting a prompt sets the model up for a specific job in one click.

Selecting a prompt

Press ⌘P (or open the prompt picker) in chat and choose one. From that point the conversation runs with the prompt’s instruction, documents, and tools in place.

You can save your own (Add Prompt on the Prompts screen) or browse Public prompts shared across the workspace:

Attach a knowledge base — chat with an entire book

This is what makes prompts powerful: a prompt can carry attached documents, and Catalyst turns them into a searchable vector knowledge base. Drop in a PDF, a spec, a handbook — even a whole book — and the model can pull the relevant passages on demand while you chat.

Here a prompt called The Little Book of Deep Learning has the whole book (a 4.5 MB PDF) attached. Catalyst split it into 47 searchable chunks and embedded them. Now any chat that uses this prompt can ask questions about the book and get answers grounded in the actual text:

How the documents are handled, automatically:

• Small documents are included in full so the model always has them in view.
• Large documents (a book, a long report) are split into chunks and embedded, then retrieved by meaning only when a question is relevant — so a 200-page book costs you nothing on questions it doesn’t touch, and precise passages on the ones it does.
• A prompt can hold up to 50 documents, and a usage bar shows how much you’ve attached.

Study, research, support

This turns a prompt into a focused study or reference companion: attach a textbook and quiz yourself, attach your product docs to answer support questions, attach a contract and interrogate it. The model cites what’s actually in the documents instead of guessing.

What else a prompt can carry

A system instruction

The standing guidance — tone, rules, role — applied to every chat that uses the prompt.

Documents

The vector knowledge base above — inline for small files, chunked & searchable for large.

Data & SQL

Bind the prompt to a data source and queries, so the model works against the right data.

Tools

Pre-select the integrations the model should have available for this job.

Sharing prompts

Make a prompt public and others in the workspace can use it — and its attached documents come along. They can run the prompt and query its knowledge base, while the raw files stay owned by you. It’s how one person can package a book, a methodology, or a dataset into something the whole team can simply select and use.

Prompts vs. memory vs. skills

These three personalization tools are easy to confuse:

• Prompts — a reusable setup you choose for a conversation: instruction + documents + data + tools.
• Memory — what the model knows about you, recalled automatically across all chats.
• Skills — task-specific know-how the model loads only when a task needs it.