The Context Rot Problem, Sorted

Ever had a Claude Code session go long and watched the model just... forget what it was doing three steps ago? Context rot. The window fills up, quality tanks, you're repeating yourself, and the whole thing turns into a shambles. Everyone hits it eventually.

GSD (Get Shit Done) fixes this by splitting your work into phases: discuss, plan, execute, verify. Each phase keeps its own focused context instead of dumping everything into one massive window. The model knows what to pay attention to right now rather than trying to hold your entire session in its head.

48K stars on GitHub, which is mental for a prompting framework. 69 commands, 24 agents running researchers and verifiers in parallel, and it works across 12+ runtimes. Not just Claude Code. Cursor, Gemini, Copilot, Cline, OpenCode, the lot.

The bit I reckon matters most: you write what you want in a structured spec, GSD breaks it into atomic tasks, each task gets its own clean context, and you get proper git commits per task. No more "one massive diff that nobody wants to review" at the end of a session.

Built by TACHES. The GSD-2 fork takes it further with even more agents. If you've been faffing about with raw Claude Code prompts and wondering why long sessions go wonky, start here. Nick the bits that work for you, bin the rest.