Skip to main content

Use capabilities in Chat

Before sending a task, prepare the model, workspace and connections it needs. Then state the action and permitted scope. The current Add (+) menu is for attachments, app context and workspaces; it does not contain Tools or Skills selectors.

Request an action​

  1. Select a model that supports tool calls.
  2. Select the intended workspace for local file work, or check Plugins → Manage connections for external services.
  3. Check Command authorization beside Add. Use Ask when you need to inspect controlled local actions.
  4. State exactly what Ghast should do, what it must not change, and the evidence to return.
  5. Read approval requests before deciding; verify the actual result after execution.

For example:

Read this project's test configuration and explain the test command. Do not modify files or install dependencies.

There is no per-tool checklist in the current Add menu. A prompt is not a security boundary, and Full access is not a read-only mode. Follow tool task procedures and permission guidance.

Use a reusable workflow​

Type / and search for the intended skill. Check that the result is a skill, not a command. Select it, then add your task to the draft. Skill selection does not send the message, but selecting a disabled skill enables it; some command results execute immediately.

Before using an unfamiliar skill, read its complete instructions and source in the Skills library. See Use skills for selection, scope and result checks.

A capability is missing​

  • Check model tool-call support and visible errors.
  • For a connected service, check connection diagnostics, account authorization and any requested tool-change review.
  • For a skill, check its source, enabled state and archived state in the library.
  • If Ghast answered without acting, clarify the required action and evidence.
  • After a failed call, inspect partial results before retrying; broader permissions are not a general fix.

See extension troubleshooting.

Check what happened​

Inspect the tool activity, input parameters and output in the response. Run summaries and More actions → Usage can record skills and tools used, but do not prove that every instruction succeeded.

Open changed files, read test results, or check objects in the external service. A completed tool call and a completed user task are not the same thing. See Read the output.