Skip to content

Green Belt

Status: drafted · Time: 14 min · Audience: experienced-builder Outcome: Understand the Green Belt arc, what is drafted in Part A, and what comes next in Parts B and C.

Graded artefacts for this belt: 🎮 Quest G-1: Author a team skill · 🎮 Quest G-2: The Greenfield cross-over · 🏁 Boss Fight G-B: The double-ship · 🏆 Green Belt badge template

Promise. You don’t just use AI — you engineer around it. You shape your team’s context, author skills, orchestrate subagents, run multiple agents in parallel. You understand the three pillars and teach them to teammates.

Prerequisite. 🟡 Yellow Belt. Optionally, enrolment in the Ship-to-Learn track (3 phases over ~7 weeks: Foundations → Guided Build → Solo Ship), nominated via the Selection Framework. Time budget. 10–14 hours of playbook, plus the capstone feature. Cohort enrolment runs over a calendar quarter. Ramp L-level. L2 (team-workflow integration).


Green Belt is a three-part belt and ships in three drops. All three Parts are drafted. Boss Fight G-B and the Green Belt badge are drafted; the belt is now claimable end-to-end. The full belt is awarded when a builder completes Parts A + B + C, both quests, and the boss fight, per Appendix L’s reviewer protocol.

Cohort leads, see Appendix L — Certification for reviewer-protocol specifics.


“My team moves faster because of how I use AI.”


§ChapterPillarTime
G.1The Three Pillars — prompt × context × harnessAll20 min
G.2Why context windows fill — the single constraint everything else followsContext20 min
G.3CLAUDE.md for a real service — WHAT + WHY, under 200 linesContext45 min
G.4Hierarchical CLAUDE.md in a monorepo — root vs package vs localContext30 min
G.5CLAUDE.local.md — personal overrides, and what belongs thereContext10 min
G.6Skills — what they are, why they compoundContext25 min
G.7Writing your first SKILL.md — the anatomy, naming, disciplineContext45 min
G.8Subagents — when to delegate, how to pass intent cleanlyHarness30 min
G.9Worktrees — running 3–5 Claude instances in parallel, safelyHarness30 min
G.10Hooks + slash commands — when to automate the pre-flightHarness30 min
G.11Advanced prompting — goals, constraints, worked examplesPrompt40 min
§ChapterPillarTime
G.12E2E testing with Playwright + Claude CodeHarness45 min
G.13The Playwright Skill pattern — one-shot test qualityContext30 min
G.14tests/seed.spec.ts — saving the agent 10,000 tokensContext15 min
G.15Design-to-code — Figma + Blade + Code Connect, end to endContext60 min
G.16Blade deep dive — tokens, primitives, variants, accessibilityContext45 min
G.17The production-compiler skill — repair AI-Studio output to BladeContext30 min
G.18The daily loop — Node, pnpm, localhost, mobile viewportHarness30 min
G.19Branch-preview platform — branch → live URLHarness30 min
G.20Observability with AI — logs, traces, cost attributionHarness25 min
G.21Debugging the hard kind — when Claude is wrong, and you have to tellPrompt30 min
§ChapterPillarTime
G.22What never goes into a prompt — the Razorpay redlinesHarness20 min
G.23The LLM proxy — what LiteLLM does and why every call routes through itHarness25 min
G.24PII, PCI, RBI — the regulators, the reasons, the reflexesContext30 min
G.25Prompt injection + output classifiers — threats and mitigationsHarness30 min
G.26The pre-ship-check skill — 6-layer gate before every PRHarness20 min
G.27The Blade-compliance reviewer skill — file-granularity checksHarness15 min
G.28Using a subagent for security reviewHarness20 min

Quest G-1. Write, commit, and share a SKILL.md that captures one of your team’s repeated workflows. At least one teammate must successfully invoke it.

Success criteria: merged skill file, invocation log from at least one teammate, one-paragraph reflection.

Quest G-2. Pick a greenfield target (an internal tool, a new agent skill pack, a self-serve analytics surface, a plugin) and ship one meaningful change that exercises Part A and Part B together.

Boss Fight G-B. A real feature on a product repo plus the Quest G-2 greenfield PR. The product-repo PR must include a scoped CLAUDE.md, at least one Playwright test, a pre-ship-check pass, the PR-guardrail used to construct the PR, and a teammate’s sign-off on the craft of the prompts. Two merged PRs + retro + teammate sign-off.


Awarding Green Belt requires evidence per Appendix L:

  • Yellow Belt awarded date (referenced, not duplicated);
  • Quest G-1 evidence — merged skill file URL plus invocation-log link;
  • Quest G-2 evidence — merged greenfield PR URL plus retro;
  • Boss Fight G-B evidence: two merged PRs (one product, one greenfield), Playwright test link, pre-ship-check output, teammate prompt-craft sign-off;
  • a one-paragraph “what changed for my team” reflection from the builder.

The badge artefact mirrors the White and Yellow templates.


  • If you have just earned Yellow Belt: open Part A README and start with G.1.
  • If you are a cohort lead: read the Appendix L reviewer protocol; Green Belt’s reviewer requirements are stricter than Yellow’s (one out-of-team reviewer required).

Files: belts/03-green/ — organised into subfolders a-craft/, b-practices/, and c-guardrails/, all drafted, plus the Green Belt badge.

Previous: ← Yellow Belt badge · Next: → Part A README