run

Copilot CLI Skill Guide

GitHub Copilot CLI is an AI-powered terminal coding agent. This skill delegates tasks to it via copilot -p (non-interactive prompt mode).

Prerequisites

Confirm copilot is available:

copilot --version

If not found, install via one of:

• winget install GitHub.Copilot (Windows)
• brew install copilot-cli (macOS/Linux)
• npm install -g @github/copilot

Stop and report failures if copilot --version exits non-zero.

Running a Task

1. Ask the user (via AskUserQuestion) which model to use AND which reasoning effort level to use in a single prompt with two questions.

   • Models: claude-sonnet-4.6 (default), claude-sonnet-4.5, claude-sonnet-4, gpt-5.2, gpt-5.4
   • Effort: low, medium, high (default), xhigh
   • Skip asking if the user already specified these in their prompt.

2. Select the permission level based on the task:

   • Read-only analysis: --allow-all (default — safe for analysis)
   • Code edits / refactoring: --allow-all (needed for file writes)
   • Network or broad access: --allow-all + confirm with user first

3. Assemble and execute the command:

copilot -p "your prompt here" \
  --model claude-sonnet-4.6 \
  --effort high \
  --allow-all \
  --no-ask-user \
  --autopilot \
  -s \
  2>/dev/null

4. Summarize the output for the user, highlighting key findings or changes made.

5. After Copilot completes, inform the user: "You can resume this Copilot session at any time by saying 'copilot resume' or asking me to continue."

Required Flags for Non-Interactive Mode

Always include these flags when using -p:

• --allow-all — required for non-interactive execution (tools + paths + urls)
• --no-ask-user — Copilot cannot ask questions when running as a sub-process
• --autopilot — let Copilot continue working until the task is complete
• -s (--silent) — output only the agent response, no stats or UI chrome

Quick Reference

Use case	Key flags
Read-only analysis	-p "prompt" --allow-all --no-ask-user --autopilot -s 2>/dev/null
Code edits	-p "prompt" --allow-all --no-ask-user --autopilot -s 2>/dev/null
Research (web)	-p "prompt" --allow-all --no-ask-user --autopilot -s 2>/dev/null
Specific directory	--add-dir /path/to/dir (add before other flags)
Structured output	--output-format json (JSONL, one object per line)
Save session log	--share ./copilot-output.md
Resume last session	copilot --continue --allow-all --no-ask-user --autopilot -s 2>/dev/null
Resume specific session	copilot --resume=<session-id> --allow-all --no-ask-user --autopilot -s 2>/dev/null

Model and Effort Flags

# Model selection
--model claude-sonnet-4.6    # Default
--model claude-sonnet-4.5
--model claude-sonnet-4
--model gpt-5.2
--model gpt-5.4

# Reasoning effort
--effort low       # Quick, simple tasks
--effort medium    # Standard tasks
--effort high      # Complex analysis (default recommendation)
--effort xhigh     # Maximum reasoning depth

Resuming a Session

To continue a previous Copilot session:

# Resume the most recent session
copilot --continue --allow-all --no-ask-user --autopilot -s 2>/dev/null

# Resume a specific session by ID
copilot --resume=<session-id> --allow-all --no-ask-user --autopilot -s 2>/dev/null

When resuming, the previous session's context is preserved. You can provide additional instructions by using -i "follow-up prompt" with --resume:

copilot --resume -i "Now also fix the edge case for empty input" \
  --allow-all --no-ask-user --autopilot -s 2>/dev/null

Restate the chosen model and effort level when proposing follow-up actions.

Critical Evaluation of Copilot Output

Copilot is powered by its own AI models with their own knowledge cutoffs and limitations. Treat Copilot as a colleague, not an authority.

Guidelines

• Trust your own knowledge when confident. If Copilot claims something you know is incorrect, push back directly.
• Research disagreements using WebSearch or documentation before accepting Copilot's claims.
• Remember knowledge cutoffs — Copilot may not know about recent releases, APIs, or changes that occurred after its training data.
• Don't defer blindly — Copilot can be wrong. Evaluate its suggestions critically, especially regarding:
  • Recent library versions or API changes
  • Best practices that may have evolved
  • Project-specific conventions documented in AGENTS.md

When Copilot is Wrong

1. State your disagreement clearly to the user
2. Provide evidence (your own knowledge, web search, docs)
3. Optionally resume the session with corrections:

   copilot --continue -i "Correction: the API changed in v3. The correct approach is..." \
     --allow-all --no-ask-user --autopilot -s 2>/dev/null
   

4. Let the user decide how to proceed if there's genuine ambiguity

Error Handling

• Pre-flight check: Always run copilot --version before first use in a session. Stop and report if it fails.
• Non-zero exit: If copilot -p exits non-zero, report the error and ask the user for direction before retrying.
• Timeout: For long-running tasks, inform the user of progress. Copilot's --autopilot mode will continue until completion, but very large tasks may take significant time.
• Permission errors: If Copilot reports permission issues, suggest adding specific directories with --add-dir rather than immediately escalating.

Tips

• Use -s (silent) to avoid stats/UI output cluttering Claude Code's context
• Append 2>/dev/null to suppress stderr (thinking tokens, progress indicators)
• If you need to see thinking tokens for debugging, remove 2>/dev/null and tell the user you're showing verbose output
• For very large codebases, use --add-dir to scope Copilot to relevant directories rather than the entire repo
• Use --output-format json when you need to parse Copilot's output programmatically
• Use --share ./output.md to save a full session transcript for later review