A Brief Introduction to Not Writing Code¶
A knowledge base and short course on AI-accelerated software engineering, from "what is a language model?" to running several coding agents in parallel and having them review each other's work. You're accountable for what ships.
You'll write a lot less code by hand. You still have to do the engineering. "Vibe coding" originally meant prompting an AI and never reading the code. This material uses the popular name for the disciplined version, where AI does most of the typing and humans decide, verify, and own the result.
This is the first full draft (November 2026). Facts are dated in each module's header. The field moves monthly, so check the Last checked date before you rely on a specific model name or price.
Start here¶
| You are | Read this | Then |
|---|---|---|
| A student or new to AI coding tools | Student path | The semester course and its labs |
| A working engineer | Engineer path | The two-week practice plan at the end of it |
| A lead or AI champion bringing a team along | Leads path | The workplace playbook |
Everyone should read 00: Orientation first. It's 20 minutes.
The modules¶
Every module has the same layers: Foundations (no background assumed), In practice (what working engineers do), and Going deeper (research and open debates), plus hands-on exercises and discussion questions. Read the parts you need.
| # | Module | What you'll learn |
|---|---|---|
| 00 | Orientation | What this is, what "vibe coding" means here, six ideas that outlast the hype |
| 01 | What is an LLM? | Tokens, transformers ("Attention Is All You Need"), pretraining, reasoning models, hallucination |
| 02 | Training, tuning, and self-tuning | How models are made, fine-tuning open weights with LoRA, distillation, models that update their own weights |
| 03 | The model landscape | The labs, current lineups, where Meta's Muse fits, the best model for each job |
| 04 | Open weight vs closed weight | Licenses, running models locally, where open models sit in the chain |
| 05 | How to tell if a model is good | Benchmarks and their flaws, benchmark-maxing, productivity studies, building your own eval |
| 06 | Harnesses, tools, and MCP | Chat vs IDE vs CLI vs cloud agents, MCP, skills, hooks, products like Claude Design |
| 07 | Context engineering and the buzzwords | Prompt → context → harness → loop → graph engineering, and the rise and fall of giant CLAUDE.md files |
| 08 | Agentic workflows | The workflow that works, Yegge's eight stages, how much autonomy is real today |
| 09 | Multiple agents | Agents building and reviewing together, and what the research says |
| 10 | AI code review | Review tools, cross-model review, how humans should review AI-written code |
| 11 | Security and prompt injection | Why injection is unsolved, the lethal trifecta, real incidents, sandboxing |
| 12 | Costs | Subscriptions vs API billing, why agents eat tokens, a worked cost example |
| 13 | Open-source ecosystem | Open harnesses, inference engines, frameworks, and the AI-slop problem for maintainers |
| 14 | Keeping up | Model ups and downs, which sources to trust, a sustainable routine |
| 15 | Learning, careers, and responsibility | Learning with AI without losing the skill, careers, privacy, licensing, ethics |
Also in this repo¶
- A 15-week course proposal with a syllabus, an instructor guide, and eight labs.
- Workplace material with an AI-first rollout playbook, honest answers to common objections, a ready-to-run internal session series, and a team policy template.
- Reference pages with a glossary of 300+ terms, an annotated reading list of 180+ research papers (start with the short list of eighteen), and a dated timeline from 2017 to now.
Conventions¶
- Bryan's take callouts are the author's opinions from daily use, clearly labelled. Everything else should be sourced. When credible people disagree, the module says so.
- Each module has a Last checked date at the top. Specific model names, prices, and benchmark numbers are kept mostly in module 03, module 12, and the timeline so they're easy to refresh.
- Papers are linked wherever a claim rests on research.
Contributing¶
CONTRIBUTING.md has the module template, voice, and sourcing rules. Anything wrong or stale is a bug, so please open an issue or a pull request with a source.
License¶
The written material and slides are CC BY 4.0: reuse and adapt them with credit. The scripts and build configuration are MIT. See LICENSE.
About¶
Written by Bryan Roscoe, a software engineer who works this way every day, with research and drafting help from coding agents. Module 00 explains how it was made.