A spec-driven loop to stop AI coders from losing the plot

What it does GSD Core is a meta-prompting framework for AI coding agents such as Claude Code, Codex, and Cursor. It drives development through a repeating milestone cycle—discuss decisions, plan, execute, verify, and ship—while farming heavy research and implementation out to fresh subagent contexts. Structured artifacts like STATE.md and CONTEXT.md persist between sessions so new agents resume with full situational awareness.

The interesting bit The project is fundamentally glue code: the agents/ and commands/ source files are native to Claude Code, and the installer translates frontmatter on the fly so the same skills load cleanly into OpenCode, Gemini CLI, Windsurf, and the rest of the supported runtimes. That cross-runtime polyfill is the unglamorous plumbing that makes the whole thing portable.

Key highlights

• Keeps the main agent context at 30–40% by offloading work to subagents with clean 200k-token windows.
• Survives session restarts via structured project memory files (PROJECT.md, ROADMAP.md, STATE.md, etc.).
• Insists on a verify step with dedicated debug agents before anything is declared done.
• Supports interactive and yolo execution modes, plus tiered model profiles (quality, balanced, budget).
• Idempotent installer adapts the same source skills for at least 15 distinct AI runtimes.

Caveats

• Codex CLI users must be on version 0.130.0 or later; earlier releases can show duplicate skill entries.
• Docker and no-Node.js setups (common on Windows with OpenCode) require manual environment-variable tweaks or per-runtime file conversion.
• Optional structural review depends on separately installing the fallow tool and matching its v2.70+ JSON schema.

Verdict Worth a look if you are trying to scale AI-assisted development past toy scripts and need repeatable, verifiable handoffs between sessions. Skip it if you just want quick chat-based autocomplete without the ceremony of structured phases and artifact maintenance.