Github Copilot Cli — 技能工具

Frontmatter Linting (Do This First)

YAML frontmatter is strict. A single extra space can break the skill.

Before committing or publishing:

# Basic sanity check (no output = good)
python - <<'PY'
import yaml,sys
with open('SKILL.md') as f:
    yaml.safe_load(f.read())
print('Frontmatter OK')
PY

Rules to remember:

• No leading spaces before keys (name, description)
• Use spaces, not tabs
• Keep frontmatter minimal (only name and description)

---

Mental Model

Treat Copilot CLI as a team of elite specialists coordinated by you:

• One Copilot instance can act as frontend engineer
• One as backend engineer
• One as tester / QA
• One as infrastructure or refactor specialist

Copilot is excellent at coding and architecture when given clear roles. You act as the CTO / conductor:

• Define goals and constraints
• Let Copilot instances propose solutions
• Observe trade‑offs and conflicts
• Escalate decisions or risks to yourself explicitly

---

Core Commands You Should Actually Use

1. Ask questions about a codebase

gh copilot explain "What does this service do?" --path src/

Use when orienting yourself or reloading context after a break.

---

2. Generate a focused change (most common)

gh copilot suggest "Add logging when translation fallback is used" --path services/translation

Best practice:

• Phrase the request as a delta, not a feature
• Always point it at a specific directory

---

3. Debug with constraints

gh copilot suggest "Why might this function return null under load?" --path src/choreo

Follow up manually by reading the code it points to.

---

4. Tests first, code second

gh copilot suggest "Write failing tests for punctuation correction on voice transcription" --path tests/

Then iterate toward the fix yourself.

---

Prompting Patterns That Work

✅ Good prompts (role-aware)

• "As a backend engineer, draft a minimal fix for X"
• "As a tester, add guards so Y never happens"
• "As infra, refactor this to be safer, not faster"

❌ Bad prompts

• "Implement feature X end-to-end"
• "Refactor the whole service"
• "Make this production-ready"

---

Multi‑Copilot Orchestration Loop (Recommended)

• Decompose (CTO)

- State the goal and constraints - Split into FE / BE / QA / Infra concerns

• Propose (Copilot roles)

gh copilot suggest "As a backend engineer, propose a minimal fix for mixed-language carryover" --path src/
gh copilot suggest "As a tester, write failing tests for mixed-language carryover" --path tests/

• Cross‑check (Copilot vs Copilot)

- Compare proposals - Look for disagreement or assumptions

• Escalate (to you)

- Surface trade‑offs - Highlight risk - Ask for decision

• Finalize (with you)

- Apply changes - Clean up naming - Merge intentionally

---

When NOT to Use Copilot CLI

Copilot CLI should not be the final authority in situations where:

• Product or organizational trade‑offs dominate over code correctness
• Cross‑repo or cross‑team coordination is required
• Security, privacy, or compliance decisions are involved
• Ambiguous state machines where correctness depends on real‑world behavior

In these cases, Copilot may still propose options, but you must explicitly review and decide.

---

Golden Rule

Copilot is a force multiplier, not a decision owner.

Use Copilot to:

• Generate competing implementations
• Surface assumptions
• Stress‑test ideas from multiple angles

You own:

• Final intent
• Risk acceptance
• Merge decisions

Copilot accelerates thinking — it does not replace judgment.