Most teams blame the model when an AI coding agent gets dumber halfway through a job. The likelier culprit is the context window: as it fills with file dumps, tool logs, and earlier reasoning, output quality quietly decays — the failure mode known as context rot. GSD's wager is that the cure isn't a bigger model but a disciplined process that keeps the main session lean and exiles heavy work to throwaway subagents.
What Sets It Apart
- A repeating five-phase loop — discuss, plan, execute, verify, ship — turns ad-hoc "vibe coding" into a checklist the agent can't skip, so decisions get made before any code is written.
- Fresh-context subagents do the expensive reading and building in their own clean windows, then hand back compact results; the orchestrating session never bloats.
- Parallel execution waves let independent work run at once instead of crawling through one long serial transcript.
- Tool-agnostic by design — the same workflow drives Claude Code, Codex, Cursor, Gemini CLI, Copilot and more, so the process survives switching agents.
How the Loop Holds Up
The discipline is the product. Plans are researched in fresh context, execution happens in clean windows, and every wave ends in an explicit verify step before anything ships as a pull request. That structure is what keeps the system coherent across long, multi-step tasks where a single chat session would have drifted.
Great Fit / Look Elsewhere
Great fit if you push long, multi-file changes through an AI agent and have watched quality crater on big tasks, or if you want a repeatable spec-driven workflow that outlives any one model. Look elsewhere for quick one-off edits, where the phase ceremony is pure overhead. Note: this original repo is archived — active development now lives in the Open GSD project (opengsd.net), so treat it as the lineage's starting point rather than the maintained source.
