Skills hold know-how, MCP connects tools: a layered Claude workflow you can copy
cfregly · x · 2026-11-01
cfregly distills takeaways from Thariq Shihipar's (@trq212) Claude Agent SDK workshop at AI Engineer Code Summit, plus his own hands-on layering advice:
- Skills package reusable know-how but need upkeep; MCP provides a common interface, yet servers still require maintenance and access checks; for a fixed brief, a plain script is often enough. Add layers only when reuse/integration is genuinely worth maintaining.
- He published a full example in the learn-claude-workshop repo: a fictional Northstar Freight renewal task where the app reads via a read-only MCP lookup, Claude executes a version-pinned Skill (with a bundled renderer/self-check), and a local evidence.py compares tool results, mounted-file hashes, and output against the local pin — unresolved approvals stay unresolved and require human review.
- Rule of thumb: tools for bounded actions, scripts for composable work.
Related event: Claude Workshop Offers Layered Guidance on Choosing Skills, MCP and Agents(2 posts)→
More from coding & agent
- Compiler-backed coding agent Benzi hits 82.6% on SWE-bench Verified for $9.26 — DonkeyTheKing · 2026-11-01
- Ronie: free open-source AI agent that makes Ray-Ban Meta glasses actually smart — Odd-Sun6296 · 2026-11-01
- RSNA 2026 to teach radiologists vibe coding for building clinical AI tools — PRouzrokh · 2026-11-01
- GPT-6.1 Terra builds 3-level game in 6.5 min; DeepSeek V4.1 Flash slower but 36% cheaper — latent_lurker8587 · 2026-11-01
- Theo: switching Opus from xhigh to high reasoning actually works better — andersonbcdefg · 2026-11-01
- Solo dev ships Pokemon-like roguelite autobattler built with Claude Code; Steam playtest live — rainbooow · 2026-11-01