This page compiles 1 file from the repository, verbatim, in reading order. The living version: this folder on GitHub.

level-2-intermediate/repos/README.md

Level 2 — Working code repositories

Four public GitHub repos back the four core walkthroughs in Level 2. Each is a real, running codebase — not a snippet — meant to be cloned, run, and modified as you work through the level.

They're intentionally scoped:

Repo	Walkthrough	What it teaches	Domain
`lease-abstractor`	02	Anthropic tool-use with schema-enforced output, side-by-side citation UI, eval against a gold reference	Real estate
`support-triage-agent`	03	3-agent pipeline (classify → enrich → draft), HITL approval dashboard, brand-voice config	Domain-agnostic (SaaS / e-commerce / local services / agency)
`diligence-agent`	04	RAG over a document folder, structured findings with citations, per-domain playbooks	Domain-agnostic (M&A / vendor eval / CRE / contracts)
`belle-mcp-server`	05	MCP server design, per-tool auth + rate limit + audit, HITL propose-write pattern	Real estate

Why this mix

You want to know these patterns, and you should not have to be a real-estate operator to learn them. Only two of four repos are real-estate-specific by design: the ones where a specific domain makes the pattern more concrete (lease documents for schema-enforced extraction, a property portfolio for MCP tool design). The other two are generic playbooks you point at any vertical.

How to use them

Read the code before the walkthrough. Skim the README and src/ tree of the relevant repo. 15 minutes.
Clone and run locally. Every repo has a Quick start section in its README that ends with a working local process. Set the env vars, run the seed, see it work end-to-end.
Follow the walkthrough. Open the corresponding module — the walkthrough refers to specific files and functions in the repo by path, not by copy-paste of the code.
Modify it on your data. Replace the fixture with something from your own life or business. That's the exit criterion for the module.

Personal use vs company use

Every repo README has both a docs/personal-use.md and a docs/company-use.md. Same code, two ways to deploy — as a solo operator or as a team. That reflects how you'll actually use these; nothing about the code changes based on scale, only the deployment shape and the governance around it.

Common tech stack across the four

Concern	Choice
Language	TypeScript for the three JS repos, Python 3.11 for `diligence-agent`
Model	`claude-sonnet-4-20250514` default across all four
Data	Supabase (Postgres) primary, local fallbacks (JSON file or ChromaDB) for zero-cloud demos
Validation	Zod (TS) / Pydantic v2 (Py) on every agent input and output
Deployment	Vercel for Next.js apps, Railway for services
Testing	Vitest / pytest, plus per-repo eval harnesses

If you're building your own version of any of these, start from these choices and only deviate when your constraints demand it.

What to build next

After you've cloned and modified at least one repo:

Combine two — e.g., feed lease-abstractor output into diligence-agent for a lease-portfolio review.
Add a new tool to belle-mcp-server for a domain you actually care about.
Fork support-triage-agent, swap the brand-voice config, and point it at your real inbound support channel.