ask-copilot – Coding skill for AI agents

Ask Copilot

Overview

This skill allows the agent to interact with GitHub Copilot CLI (copilot) in a non-interactive (headless) mode. Use this skill when the user explicitly wants secondary advice, code reviews, explanations, or code generation from GitHub Copilot's models.

Use source: self and source_type: self when the skill is original to this repository and does not require README external-source credit.

Copilot is an external service. Treat prompts, file paths, snippets, repository content, command output, and generated suggestions as data that may leave the local environment.

When to Use This Skill

• User Request Only: Use this skill ONLY when the user explicitly asks to "consult Copilot", "ask Copilot", "review with Copilot", or explicitly requests a second opinion using Copilot.
• Do NOT Invoke Automatically: To comply with privacy policies, the agent must not invoke this skill automatically for its own second opinions or checks without explicit user consent.

How It Works

Step 1: Request Explicit User Consent

Before executing any command that references local files, repository paths, snippets, command output, secrets-adjacent config, or private project context, you MUST obtain explicit user consent to send that material to GitHub Copilot.

Ask for separate approval before allowing Copilot to run tools, execute shell commands, edit files, install packages, or mutate the workspace.

Step 2: Execute with Minimal Permitted Flags

To prevent TUI lockups, execute the copilot command with headless flags. Do not use blanket bypasses such as --yolo, --allow-all-tools, or --allow-all-paths for routine Q&A or review.

• For Read-Only / General Q&A: Send only the user-approved, redacted text in the prompt. Do not grant Copilot broad local-path access; it is not needed when the prompt already contains the approved context.
• For Trusted Mutation Tasks: Prefer a scoped permission flag if the CLI supports one. Use blanket mutation bypasses only after the user explicitly authorizes Copilot to execute tools and mutate the workspace for the specific task.

Step 3: Use Session Management (Optional)

To maintain conversation context, use --name and --resume flags, or pass a --session-id on subsequent calls.

Examples

Example 1: General Question (Read-Only)

Does not require repository path access or mutation permissions.

copilot -p "Explain how to implement a debounce function in TypeScript" -s

Example 2: Code Review (Approved File Excerpt)

Always confirm the exact file and excerpt with the user before executing. Keep the path in a

quoted variable; build the prompt from a static instruction plus the approved excerpt. Shell

does not re-evaluate command-substitution output, so metacharacters inside the reviewed file

remain prompt text rather than shell syntax:

review_file="path/to/file.ts"
test -f "$review_file" || { echo "File not found: $review_file" >&2; exit 1; }
copilot -p "$(printf '%s\n\n' 'Review this approved excerpt for potential memory leaks:'; sed -n '1,220p' -- "$review_file")" -s

Never construct a shell command by interpolating user-controlled prompt text, paths, issue

content, or filenames into shell source. Use fixed command structure, quoted variables, and

approved file content only.

Example 3: Named Session Management

copilot -p "Remember this session label for follow-up questions." -s --name "my-session-name"
copilot -p "Summarize the prior advice in this session." -s --resume "my-session-name"

Best Practices

• ✅ Do: Ask for user consent before uploading any project files to third-party endpoints.
• ✅ Do: Send only the approved, redacted excerpt; keep Copilot out of the broader workspace.
• ✅ Do: Keep untrusted values in quoted variables or command input, never in shell source.
• ✅ Do: Use -s (silent) to suppress metadata and statistics, leaving only clean output.
• ❌ Don't: Automatically trigger this skill for background second opinions without the user's explicit ask.
• ❌ Don't: Send files, logs, environment details, or private repository context to Copilot without explicit approval.
• ❌ Don't: use --allow-all-paths for a review, or interpolate untrusted text inside copilot -p "...".
• ❌ Don't: add permission bypasses to avoid a prompt. Discover supported non-interactive scoped permissions; stop if the approved read-only invocation cannot run without broader access.

Limitations

• This skill does not replace environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, or safety boundaries are missing.
• Copilot responses may be incomplete, outdated, or wrong; verify any proposed code locally before using it.

Security & Safety Notes

• The --yolo flag bypasses all permission prompts and allows Copilot CLI to run arbitrary shell commands and mutate workspace files. It must be treated as a high-risk option and never used by default.
• Always check that the code/files being sent do not contain sensitive credentials, API keys, or private environment variables.
• Prefer redacted snippets over whole files when only a small context sample is needed.
• --allow-all-paths grants Copilot broader local visibility than a narrow review requires; it is not a read-only least-privilege flag.

Common Pitfalls

• Problem: The terminal hangs or the command times out.

Solution: Ensure both -p (or --prompt) and the narrowest required non-interactive permission flag are present in the command arguments. Without required permission flags, the CLI may prompt for confirmation and hang headless processes.

Related Skills

• @cli-assistant - How to interact with CLI tools in general.