Source: GitHub · gsd.build · Created by Lex Christopherson
/gsd:new-project # Interview + 4 parallel research agents
/gsd:discuss-phase # Capture preferences for gray areas
/gsd:build # Atomic plans ? wave parallelism ? fresh agents
/gsd:verify # "Is it TRUE?" not "Was it done?"Core Idea
AI output degrades as context fills. Task 50 gets worse results than task 1. GSD spawns a fresh sub-agent per task, so every task gets a clean 200K token window with only what it needs.
What's In the Box
v1: ~50 markdown files (29 slash commands, 12 agents, 2 hooks). Runs on native Claude Code. A Node.js CLI handles deterministic logic so the LLM focuses on reasoning.
v2: Standalone TypeScript CLI that programmatically controls agent sessions. Adds crash recovery, stuck detection, cost tracking, and git worktree isolation.
How Each Command Works
new-project interviews you while spawning 4 parallel research agents (stack, features, architecture, pitfalls). Outputs: PROJECT.md, REQUIREMENTS.md, ROADMAP.md, STATE.md.
discuss-phase identifies gray areas and asks targeted questions. Answers go into CONTEXT.md. Every subsequent agent inherits them. Skipping this is the #1 source of misaligned output.
build creates atomic PLAN.md files (2-3 tasks each, ~50% context window). Independent plans run in parallel waves. Each plan is the prompt. The sub-agent reads it directly as its instruction:
<task type="auto">
<name>Create login endpoint</name>
<files>src/app/api/auth/login/route.ts</files>
<action>Use jose for JWT. Validate against users table.</action>
<verify>curl POST localhost:3000/api/auth/login returns 200</verify>
</task>Each plan = one atomic git commit, one clean agent session.
verify checks goals, not task completion. "Can a user sign up with email?" not "Does the signup route exist?" Failures generate new plans instead of retrying.
Strengths
| Strength | Detail |
|---|---|
| Fresh context per task | 200K windows. No degradation from first to last task. |
| Four commands | Complexity is in the system, not the interface. |
| Wave parallelism | Independent tasks run simultaneously. |
| Plan verification | Up to 3 iterations before execution. Checks scope, deps, conflicts. |
Tradeoffs
| Limitation | Detail |
|---|---|
| Token hungry | Multi-agent spawning burns through Pro plan limits fast. |
| Greenfield bias | Deep dependency chains in large codebases can exceed single-task comprehension. |
| Discuss phase is critical | Skip it and plans inherit bad assumptions. |
| Solo-oriented | Less multi-person coordination than BMAD. |
| No MCP in sub-agents | Claude Code limitation, not GSD's. |
Good Fit / Bad Fit
Use it: Solo/small team, greenfield, you want agents handling implementation while you make product decisions.
Skip it: Large brownfield codebases, tightly coupled systems, teams needing role separation and audit trails.