Builder

From "I've never used AI to code" to "I ship production systems with observability and governance."

5 levels 8-12 weeks artifact: Deployed production application

Demonstrated, not asserted

  • Rally HQ live A live tournament platform built end to end with this method — the path's destination, running in production. verified 2026-07-15
  • specchain install The spec → tasks → implementation workflow this track teaches, as an installable tool. npx create-specchain verified 2026-07-15
  • Ask BC gated· BigCommerce merchants A production agentic assistant — proof the Level 3-4 patterns ship, not just demo. verified 2026-07-15

Who this is for

This path is for you if

  • ·You're a developer who wants to ship AI-powered software
  • ·You learn by building, not by watching tutorials
  • ·You want production-ready code, not demo projects
  • ·You're comfortable with JavaScript/TypeScript and web development

Probably not if

  • ·You want a curriculum to passively consume
  • ·You're collecting credentials — the artifact here is a deployed app, not a certificate

The levels

0 Orientation 1-2 days

Understand the landscape before you start walking.

Terms that matter

  • LLM — Large Language Model, the engine behind every tool here
  • RAG — Retrieval-Augmented Generation: giving AI context from your documents
  • Prompt Engineering — designing inputs that get useful outputs
  • Agentic AI — AI that takes actions, not just generates text

Read

  • Anthropic's prompt engineering guide — platform.claude.com/docs
  • AI SDK overview (how modern apps integrate LLMs) — ai-sdk.dev

Checkpoints

  • ·You can explain LLM, RAG, and agentic AI to a colleague
  • ·You understand what an "AI agent" does that a simple prompt doesn't
  • ·You can set up a new SvelteKit or Next.js project from scratch

Done when — You can explain key AI concepts (LLM, RAG, agents) and why AI-assisted development differs from chat interfaces

1 Ship Your First AI App 1 week

Get your hands dirty. Build something — anything — with AI assistance.

Set up

  • Claude Code — CLI-based AI assistance (what this whole site is built with) — code.claude.com/docs
  • Cursor — an IDE with AI built in — cursor.com

Pick a project

  • ·A personal dashboard that pulls data from an API you use
  • ·A CLI tool that automates something tedious in your workflow
  • ·A simple web app that solves a real problem you have

Rules

  • ·Must be functional (not a mockup)
  • ·Must use AI assistance throughout — keep your prompts
  • ·Must be deployed somewhere (Cloudflare, Vercel, Netlify, anywhere)

Checkpoints

  • ·You have a deployed application
  • ·You can show what you built
  • ·You can articulate what was hard and what was easy

Done when — You have a deployed application you can show

2 Specs, Databases & Workflows 2-4 weeks

Move from experimental prompting to documented, repeatable workflows.

Skills

  • Prompt patterns — zero-shot vs few-shot, chain-of-thought, system vs user prompts, picking the right model for the task
  • Context management — the right context without overwhelming the model; RAG basics; when to start fresh vs continue a thread
  • Plan-first development — write specs before code, have AI generate plans then execute them, document decisions as you go. This is exactly what specchain packages: npx create-specchain

Project: rebuild your Level 1 app — better

  • ·A written spec, before you start
  • ·Structured prompts, not just "build me X"
  • ·Documentation generated alongside the code
  • ·At least one integration (database, API, external service)

A stack that works

  • Frontend — Next.js or SvelteKit (pick one)
  • Database — Supabase (PostgreSQL + auth + realtime)
  • Styling — Tailwind CSS
  • Type safety — TypeScript, strict mode
  • Testing — Playwright for end-to-end

Checkpoints

  • ·You have a spec document for your rebuilt app
  • ·Your app has a database and authentication
  • ·You can explain your prompt workflow to someone else
  • ·You've hit at least 3 walls and figured out how to get past them

Done when — You have a spec document and your app has a database and authentication

3 Multiple Models & Tool Use 4-8 weeks

Move from single AI calls to multi-step workflows with tool use.

Skills

  • Multi-provider integration — when each model family fits, cost optimization, fallbacks, streaming (AI SDK: ai-sdk.dev; LangChain for complex chains; direct API for full control)
  • Agentic patterns — reflection, tool use, planning, multi-agent, and MCP for standardized tool definitions (modelcontextprotocol.io)
  • RAG systems — chunking strategies, vector embeddings and similarity search, hybrid search, context-window management (supabase.com/docs/guides/ai)

Worked patterns from real systems

  • Agentic product feed — AI analyzes a catalog, generates optimized descriptions, validates against brand guidelines, exports to channels
  • Search query orchestration — query → intent classification → multi-source retrieval → LLM reranking → structured response
  • Data quality agent — monitors incoming data, flags issues, suggests corrections, routes to human review. Ask BC ships this family of patterns in production against live stores

Project: build something that uses tools

  • ·Calls at least 2 external tools or APIs
  • ·Makes decisions based on retrieved data
  • ·Handles errors gracefully
  • ·Logs its reasoning

Checkpoints

  • ·Your app uses multiple AI providers or tools
  • ·You understand why different models fit different tasks
  • ·You can implement basic RAG — retrieve context, augment the prompt
  • ·You've built something that feels like an "agent," not just a "chatbot"

Done when — You've built something that feels like an "agent," not just a "chatbot"

4 Production Systems & Observability 8+ weeks

Build systems that are production-grade, not just demos.

What production-grade means

  • Reliability — it works consistently, not just sometimes
  • Observability — you know when it fails and why
  • Cost control — you are not burning money on API calls
  • Safety — it doesn't do things it shouldn't

Skills

  • LLM observability — tracing every request, cost per request, latency, quality drift (langfuse.com/docs)
  • Testing AI systems — evaluating non-deterministic outputs, regression testing for prompts, end-to-end tests with AI components
  • Governance and safety — prompt-injection prevention, output validation, audit trails, rate limiting and cost caps per user

Project: add observability & governance to your Level 3 app

  • ·Add tracing — every AI call logged
  • ·Add cost tracking — know your spend per user and per request
  • ·Add output validation — catch bad responses before they reach users
  • ·Add basic access control — who can use which features

Checkpoints

  • ·You can show metrics from your AI system
  • ·You know your cost per request
  • ·You have validation that catches bad AI outputs
  • ·You've thought about what your AI shouldn't do

Done when — You can show metrics, cost tracking, and output validation

Start at Level 0 npx create-specchain

Tool links and level references live in a dated data file checked by the link sensor on every build — this page can't silently point at a moved doc or a dead demo.