GSD Core

Workflow

Git. Ship. Done — a spec-driven framework that drives AI coding agents through a five-phase Discuss → Plan → Execute → Verify → Ship loop.

by open-gsdopen-gsd/gsd-core

Stars10.8k

Forks763

Watchers41

Open issues282

LicenseMIT

Last pushtoday

Install

install

$ npx @opengsd/gsd-core@latest

Update

update

$ npx @opengsd/gsd-core@latest

Re-running fetches the latest version.

When to use it

You want a disciplined, repeatable phase loop (Discuss, Plan, Execute, Verify, Ship) instead of ad-hoc prompting
Your agent sessions degrade from 'context rot' on long tasks and you want heavy work pushed into fresh-context subagents
You need structured artifacts (STATE.md, CONTEXT.md) so work survives across sessions and agent handoffs
You run more than one AI runtime (Claude Code, Codex, Cursor, Gemini CLI, Copilot, Windsurf…) and want one shared workflow across all of them

Commands & flags

/gsd-new-project: Scaffold a new GSD project and start the phase loop.
Discuss: Phase 1: lock scope and intent before any planning.
Plan: Phase 2: break the work into parallelizable waves.
Execute: Phase 3: run waves in fresh-context subagents (~200k each).
Verify: Phase 4: run built-in checks as a gate before shipping.
Ship: Phase 5: finalize and deliver the verified work.

How to use it

Install with the npx command, then start a project with the /gsd-new-project slash command and let GSD walk you through the phase loop.

npx @opengsd/gsd-core@latest

The installer detects your runtime and lays down the GSD skills/commands. Inside the agent, run /gsd-new-project to scaffold, then iterate through the five phases: Discuss (lock scope and intent) → Plan (break work into waves) → Execute (run waves in parallel fresh-context subagents) → Verify (built-in checks before shipping) → Ship. The main session stays lean while each wave gets its own ~200k-token context, and progress is persisted to structured artifacts like STATE.md and CONTEXT.md.

Why it earns its place

GSD Core turns context engineering into an enforced process: instead of one long, rotting conversation, it splits work into parallel waves that each run in a clean subagent context and report back through durable artifacts. The five-phase loop puts a verification gate before every ship, and the same workflow installs across nine+ AI runtimes, so the discipline travels with you rather than being locked to one tool. MIT-licensed, ~4k stars, by the open-gsd org.

Try it

claude

> /gsd-new-project a CLI that summarizes my unread email each morning