fix(skills): force MCP tool usage in all 15 skills

[HumanBean17]

Summary

• All 15 skills described MCP tool calls in prose/example format but lacked enforcement directives — agents could bypass MCP entirely and answer from training data or file browsing
• Added "MCP required" section to every skill with explicit YOU MUST call these MCP tools directive
• Added reasoning preamble requirement (Q-class / Pick / Why) before each MCP call (matching AGENT-GUIDE.md convention)
• Added reference to docs/AGENT-GUIDE.md for the full operating manual
• Added "Do not" guardrails: no answering from training data, no file browsing when MCP can answer, no fabricating ids
• Rewrote Steps sections with imperative language ("Call resolve..." instead of "resolve(...)")
• Updated skills/README.md with skill structure documentation

Root cause

Skills were written as documentation (worked examples, argument contracts) without any enforcement language. An agent reading a skill understood what to do conceptually but wasn't forced to actually make MCP tool calls. The fix adds three enforcement layers:

1. MCP required section — explicit dependency declaration + "YOU MUST" directive
2. Reasoning preamble — forces the agent to think about which tool and why before each call
3. Do not guardrails — explicitly prohibits the fallback behaviors (training data, file browsing, guessing)

Test plan

• All 15 skills contain "MCP required", "Do not", "reasoning preamble", and AGENT-GUIDE reference (verified via grep)
• ruff check . passes (docs-only change)
• Manual: invoke each skill and verify agent makes MCP tool calls

🤖 Generated with Claude Code