open-gsd/gsd-core

GSD Core is a meta-prompting and spec-driven development framework that runs AI coding agents through a repeatable five-step phase loop: discuss, plan, execute, verify, ship. It targets developers already using an agent such as Claude Code, Codex, Copilot, Cursor or Windsurf who find long sessions degrading in quality as context fills up. The project is MIT-licensed and written in JavaScript.

How it works

The core problem addressed is "context rot": output quality falling off as an agent's context window fills with accumulated research and history. GSD Core responds by moving heavy research, planning and execution into fresh-context subagents while keeping the main session lean. Executors are described as starting with a clean 200k-token context and running plans in parallel waves README.

Between sessions, structured artifacts such as STATE.md and CONTEXT.md carry decisions forward, so work is not lost when a session ends. The verify step walks through what was built and produces fix plans before a phase is declared complete, rather than relying on the agent's own claim of success README. The loop and reasoning are documented separately in the context-engineering and phase-loop explanation pages context engineering.

Installing and using it

The documented path is an npx @opengsd/gsd-core@latest installer, which asks for the target runtime and whether to install globally or locally README. The README explicitly says the installer is required for cross-runtime compatibility and that copying files from agents/ or commands/ directly is not supported. That matters if you are inclined to vendor the prompt files manually.

On a runtime without that installer, or without Node.js, there is a separate guide install on your runtime. After installation, /gsd-new-project starts a greenfield project and /gsd-onboard brings an existing repository into the workflow README. Guided walkthroughs exist for both the first project and existing-codebase onboarding first project tutorial.

What to check before adopting

The framework is designed around AI coding agents and assumes you already have one configured; it does not replace the agent. The exact runtime support list and any per-runtime caveats should be confirmed in the installation guide, since the README's list ends with "and more" without enumerating every supported tool.

The documentation also includes a dedicated recover-and-troubleshoot page, which is the place to look for what to do if the phase loop goes wrong recovery guide. Command, configuration and CLI details are kept in separate reference pages rather than the README commands.

HotRepo's take

One dated observation exists: 11,026 stars with +42 that day on 2026-10-30, the repository's only recorded day on the trending board. The supplied sample is too short to say anything about momentum or persistence. The more useful signal for a prospective user is the documented workflow and its single required installer path, not the star count.