Skip to content

Black Belt Part A — Build the Platform

Status: drafted · Time: 8 min · Audience: platform-builder Outcome: Understand what Part A teaches, what Green-Belt muscle memory it assumes, and the order to walk it in.

Part A promise. By the end, you can author MCP servers, publish shared skills and plugins through the canonical repositories, prove every promised user surface, decide when to write your own agent with the SDK, orchestrate multi-agent workflows, and design tool contracts that compose.

Prerequisite. Green Belt awarded. Nomination registered with the cohort lead. The chapters assume comfort with G.6 / G.7 (skills), G.8 / G.9 (subagents, worktrees), G.26 (pre-ship-check), and the SKILL.md anatomy those chapters teach. Time budget. ~4h 15m for the six modules; ~8h for Quest B-1. Plan over a week or two.


Green Belt taught you to build with AI: craft, practices, and guardrails that ship into a product repo cleanly. Part A of Black Belt teaches you to build the platform AI runs on: MCP servers other PODs install, shared skills that compound through razorpay/agent-skills, plugins distributed through razorpay/claude-plugins, agent-SDK decisions for bespoke cases, multi-agent orchestration patterns, and tool-contract design.

Six modules, one quest. The boss fight (POD-uplift embed) is in Part C; Black Belt is awarded when all three Parts close.


§ChapterPillarTime
B.1Authoring an internal MCP serverHarness60 min
B.2Publishing a shared skillContext30 min
B.3Publishing a plugin — package once, prove every surfaceHarness45 min
B.4The Claude Agent SDK — when to write your own agentHarness45 min
B.5Multi-agent orchestrationHarness45 min
B.6Tool design — JSON schemas, output shapes, error contractsHarness30 min

Walk them in order on a first read. B.1 anchors the platform-builder voice; B.2 and B.3 are the publishing layer; B.4 is the build-vs-buy decision; B.5 and B.6 are the systems-design layer.


Quest B-1 — Publish a shared skill. Publish a repository-native skill so another POD can install it with one command. At least two installs from PODs outside your team.

The quest is the practical test of B.2. Read B.3 before starting so you can explain why the shared library, rather than plugin packaging, is the right distribution surface for your workflow.


“I author MCP servers, publish shared skills and plugins, prove each supported user surface, and design tool contracts other teams compose with.”


Part B (Push the Craft) covers progressive disclosure, memory systems, prompt evals, team-and-org-scale observability, effort settings and model routing. Part C (Shape the Org) covers running office hours, embedded sprints, AI RFCs, API Council contributions, and skill / plugin governance. Both are drafted; walk Part A first.


Previous: ← Black Belt README · Next: → B.1 Authoring an internal MCP server