GSD: Advanced Vibe Coding

Get Started
- Download the lesson folder: DOWNLOAD
- Open terminal in that folder
- Run
claudethen type/start-3-1
Prerequisites: You should be comfortable with basic Claude Code usage (vibecoding). If not, complete Module 2: Vibe Coding first.
GSD Framework: Spec-Driven Development
What This Is
An interactive lesson teaching the GSD (Get Shit Done) framework - a production-grade workflow for building complex projects with Claude Code. You’ll build a real Expense Splitter app while learning how to prevent context rot, use fresh subagent contexts, and create atomic, verifiable builds.
Source: GSD by Lex Christopherson
The Problem GSD Solves
Ever had Claude start strong then get worse as the session goes on? Like it’s rushing, cutting corners, forgetting things you told it earlier?
That’s context rot. It’s not your imagination - Claude’s quality actually drops as the context window fills up:
- 0-30% context: Peak quality. Thorough, comprehensive, remembers everything.
- 50%+: Starts rushing. “I’ll be more concise.” Cuts corners.
- 70%+: Hallucinations. Forgotten requirements. Drift.
Basic vibecoding works for simple stuff, but it doesn’t scale. Complex projects fall apart halfway through.
How GSD Solves It
GSD uses fresh subagent contexts. Instead of one long session that gradually degrades, GSD spawns fresh Claude instances for each task. Each subagent gets a clean 200,000 token context window.
Task 50 has the same quality as Task 1. No degradation.
Meanwhile, your main session stays at 30-40% context while agents do the heavy lifting.
The GSD Workflow
Four commands, one workflow:
| Step | Command | What It Does |
|---|---|---|
| 1 | /gsd:new-project | Interview → Research → Requirements → Roadmap |
| 2 | /gsd:plan-phase N | Break phase into atomic task plans |
| 3 | /gsd:execute-phase N | Fresh agents build, verify, commit |
| 4 | /gsd:verify-work N | Goal-backward QA walkthrough |
Each step builds on the previous. Requirements trace to phases. Phases trace to plans. Plans trace to commits. Everything is connected and verifiable.
Time Expectations
This is important. GSD does real work, which takes real time.
| Command | Time |
|---|---|
/gsd:new-project | 10-15 minutes |
/gsd:plan-phase | 5-10 minutes |
/gsd:execute-phase | 10-15 minutes |
/gsd:verify-work | 3-5 minutes |
Full lesson time: 45-60 minutes, about 30 minutes of which is GSD working while you wait.
That’s intentional. That’s how you build quality software. Grab a coffee during the long steps.
What You’ll Learn
- Context engineering - Fresh contexts beat one degrading session
- Aggressive atomicity - Small plans, each fits in ~50% context
- Plans are prompts - PLAN.md files ARE the executable instructions
- Wave-based parallelism - Independent work runs in parallel
- State persistence - Resume anytime with STATE.md
- Goal-backward verification - Check what must be TRUE, not what tasks were done
The Five Lessons
This module has 5 interactive lessons. You can run them in sequence, or jump to the section you want to learn about.
Lesson 3.1: What is GSD?
Command: /start-3-1
What You’ll Learn:
- The context rot problem and why it happens
- How fresh subagent contexts solve it
- Installing GSD (
npx get-shit-done-cc) - The full workflow sequence
- What to expect from
/gsd:new-project
Key Concept: GSD spawns fresh Claude instances for each task. Each gets a clean 200k context. Your main session stays clean while agents do heavy lifting.
Time: ~10 minutes (interactive, no waiting)
Lesson 3.2: Start a Project
Command: /start-3-2
What You’ll Do:
- Run
/gsd:new-projectfor the first time - Answer GSD’s interview questions
- Watch research agents investigate best practices
- Tour the
.planning/folder that gets created
Files Created:
PROJECT.md- Vision and constraintsREQUIREMENTS.md- Features with traceable IDsROADMAP.md- Phases with success criteriaSTATE.md- Persistent memory across sessions
Key Concept: GSD interviews you, spawns parallel research agents, and creates a complete spec. Your context stays clean while agents do the research.
Time: ~15 minutes (10-15 minutes is GSD working)
Lesson 3.3: Plan the Build
Command: /start-3-3
What You’ll Do:
- Run
/gsd:plan-phase 1 - Watch GSD create atomic task plans
- Understand the XML plan structure
- Learn wave-based parallelism
Key Concepts:
Aggressive atomicity: Each plan is 2-3 tasks, designed to fit in ~50% of a fresh context window. No single task is big enough to degrade quality.
Plans are prompts: The PLAN.md file isn’t a document that becomes a prompt. It IS the executable instruction. Subagents read it directly.
Waves: Independent tasks run in parallel. Dependent tasks wait. Wave 1 might run three plans simultaneously. Wave 2 waits for Wave 1, then runs.
Time: ~10 minutes (5-10 minutes is GSD working)
Lesson 3.4: Execute
Command: /start-3-4
What You’ll Do:
- Run
/gsd:execute-phase 1 - Watch fresh agents build your app
- See atomic commits in git history
- Run your app and see what got built
Key Concepts:
Fresh contexts: Each agent gets a full 200k context. No degradation from first task to last.
Atomic commits: Each task gets its own commit. Git bisect works. You can revert individual tasks without losing other work.
Phase 1 is foundation: You’ll have a working app, but it’s Phase 1 only. Full app needs Phases 2 and 3.
Time: ~15 minutes (10-15 minutes is agents building)
Lesson 3.5: Verify & Beyond
Command: /start-3-5
What You’ll Do:
- Run
/gsd:verify-work 1 - Walk through interactive QA testing
- Learn about
/gsd:quickfor small changes - Understand when to use GSD vs plain Claude Code
Key Concepts:
Goal-backward verification: Instead of “what tasks did we do?” GSD asks “what must be TRUE for this to work?” You test observable behaviors, not implementation details.
Quick mode: /gsd:quick "Add dark mode" - same fresh agent guarantees, skips heavy planning. Use for tweaks.
When to use GSD:
- ✅ Complex projects (multiple components, real logic)
- ✅ Multi-session work
- ✅ You care about clean git history
- ❌ Quick one-off tasks (just use Claude directly)
- ❌ Prototyping (basic vibecoding is fine)
Time: ~10 minutes
Other GSD Commands
Once you finish the module, here are more commands to explore:
| Command | What It Does |
|---|---|
/gsd:progress | Where am I? What’s next? |
/gsd:quick "task" | Ad-hoc task with GSD guarantees |
/gsd:discuss-phase N | Clarify preferences before planning |
/gsd:debug | Systematic debugging with agents |
/gsd:map-codebase | Add GSD to an existing project |
/gsd:pause-work | Clean handoff between sessions |
/gsd:resume-work | Pick up where you left off |
/gsd:help | See all commands |
By the End
You’ll have:
- A working Expense Splitter app (Phase 1 - you can continue to Phases 2 and 3)
- GSD installed and configured
- Understanding of fresh subagent contexts and why they matter
- Experience with the full GSD workflow
- Clean git history with atomic commits
- The ability to use GSD on your own projects
Download the lesson folder: DOWNLOAD