---

name: background-prompt description: Execute a task in the background with a prompt through the core engine LLM provider and return plain text output. Use when the user wants to run a prompt as a background task via the Agent CLI.

Background Prompt

Run a prompt as a background task through the local core/bin/agent-cli tool.

When to Use This Skill

• User wants to execute a prompt in the background
• User wants a prompt executed by the default core engine LLM provider
• The task is suitable for single-prompt inference with plain text output

Your Roles in This Skill

• Backend Developer (Engineer): Execute the tool command and capture output safely
• Technical Writer: Convert tool output into a clear user-facing response
• Security Engineer: Ensure prompt execution follows configured security flags

Role Communication

As an expert in your assigned roles, you must announce your actions before performing them using the following format:

As a {Role, and Role-XYZ if have more roles}, I will {action description}

This communication pattern ensures transparency and allows for human-in-the-loop oversight at key decision points.

Instructions

Follow these steps in order:

Step 1: Confirm intent

Ensure the user wants to run a background prompt. If not explicit, ask for confirmation before invoking.

Step 2: Build the prompt

Compose one concise prompt string from user instructions. Keep task scope explicit and include required output format details.

Step 3: Execute Agent CLI

Run from repository root:

core/bin/agent-cli "<prompt>"

Optional overrides only if user requests:

core/bin/agent-cli --provider <provider-id> --sandbox --auto --network "<prompt>"

Step 4: Return plain text result

Return the command output as plain text to the user. If command fails, include the error text and a concrete retry suggestion.

Expected Output

• A plain text response generated by core/bin/agent-cli
• Any execution error surfaced clearly when failures occur

Key Principles

• Use core/bin/agent-cli as the primary execution path
• Keep prompts specific and deterministic
• Respect configured security defaults unless user requests overrides