A field guide
to creative AI coding.
A few useful starting points for your first sessions. Short enough to keep the work moving; specific enough to make the next step deliberate.
01Getting oriented
Start with the material, not a perfect prompt.
Choose a folder you recognize and begin with a read-only request. Ask the agent to explain the structure, identify the important files and describe what it would inspect next.
02Choosing a model
Make the connection a deliberate choice.
The supplied reference supports configured endpoints. Add the API base URL, credentials when required and a model available on that server. Test a small request before a large task.
03Connecting tools
Add what the current question needs.
Use the connector settings to configure an MCP server and its required command, arguments or credentials. Dependencies and access are connector-specific. Avoid enabling tools only because they are available.
04Giving better feedback
Describe the gap, not just your reaction.
Say what the result should do for its audience, what currently gets in the way and what must not change. A focused revision gives you a result you can compare.
05Reviewing the work
Look at the result outside the conversation.
Read the files that changed and try the relevant interaction. For code, run appropriate checks. For content, read it as the intended audience. A confident answer is not a verification result.
06Knowing where context goes
Local workspace does not always mean local inference.
A remote endpoint processes the request context sent to it. Connected services can receive the inputs needed by their tools. Check their permissions and data policies before using sensitive material.
07About this preview
A direction, with room to develop.
The current site introduces a separate product direction. Screenshots are shared reference material and will be replaced. Windows installation becomes available when its release file is published.
No matching notes. Try "model", "tools" or "review".
Make your first session easy to evaluate
Start with a familiar folder and a read-only question. Ask for an explanation of the structure and compare it with what is actually present. This gives you an early view of how the chosen model uses context, before an edit or tool action makes the session more complicated.
When you get stuck, identify whether the problem concerns the brief, the endpoint or an external tool. These require different next steps. Refine an unclear request, test a connection separately or inspect a connector permission instead of changing the entire setup at once.