What it does

GSD Core is a meta-prompting framework that installs as skills into AI coding agents. It drives a rigid five-step loop — discuss, plan, execute, verify, ship — with the explicit goal of fighting “context rot,” the quality degradation that sets in as a long session fills the model’s context window. Heavy research and execution work gets farmed out to subagents that each start with a clean 200k-token window, while your main session stays lean.

The interesting bit

The project treats the AI agent less like a pair programmer and more like a distributed system that needs orchestration. It maintains structured artifacts (STATE.md, CONTEXT.md, ROADMAP.md) that survive session boundaries, so a new chat doesn’t start from zero. The verify step is non-optional: a dedicated agent walks through what was built, diagnoses failures, and generates fix plans before anything gets shipped.

Key highlights

• Supports 15 runtimes including Claude Code, Codex, Gemini CLI, Cursor, Copilot, and Windsurf; an installer transforms source files per-runtime rather than maintaining separate ports
• Fresh-context execution: each subagent starts at 30–40% context usage, leaving headroom for the actual task
• Structured memory across sessions via markdown artifacts in a .planning/ directory
• Parallel plan execution with atomic commits; ships via PR creation
• Idempotent installer (npx @opengsd/gsd-core@latest) with profiles to limit installed surface area

Caveats

• Requires --dangerously-skip-permissions for Claude Code, which the README presents as “how it’s intended to run” — a notable security posture
• The agents/ and commands/ source files are Claude Code-format; manually copying them to other runtimes will break without the installer’s frontmatter conversion
• Codex users need CLI ≥0.130.0 due to a breaking change in skill discovery

Verdict

Worth a look if you’re already using Claude Code or similar agents for non-trivial projects and hitting the “it worked for the first three files then lost the plot” wall. Probably overkill if your AI usage is limited to one-off refactors or chat-based debugging.