GSD Fixes Context Rot in AI Coding Agents
GSD is an open-source workflow layer for tools like Claude Code and Cursor that breaks large coding projects into map, discuss, plan, execute, and verify phases to prevent context bloat, forgetting decisions, and unreliable outputs.
Combat Context Rot with Front-Loaded Codebase Understanding
AI coding agents fail on larger projects due to context rot: initial prompts yield brilliant responses, but bloat leads to shorter answers, forgotten decisions, random changes, and babysitting. GSD counters this as a spec-driven workflow layer installed atop Claude Code, Codex, Gemini CLI, OpenCode, Copilot, Cursor, or Antigravity. Start with /GSD:map-codebase (Claude Code/Gemini CLI) or $GSD-help (Codex) to spawn parallel agents analyzing architecture, conventions, stack, and pain points. This builds shared knowledge so agents avoid misguided changes. Follow with /GSD:new-project to generate persistent memory files: project.md, requirements.md, roadmap.md, state.md, and a planning research folder, extracting requirements into a structured roadmap.
/GSD:discuss-phase surfaces ambiguities early—like UI layouts, densities, interactions, empty states for frontends, or API response formats, flags, error handling for backends—preventing silent product decisions by the model.
Atomic Planning and Parallel Execution for Reliable Outputs
/GSD:plan-phase researches the phase, creates small atomic task plans fitting fresh context windows, and verifies them against requirements, chunking work to maintain focus without recalling entire conversations. /GSD:execute-phase groups tasks into dependency-based waves: independent tasks run in parallel (favoring vertical end-to-end slices over horizontal layers to minimize conflicts), producing atomic git commits per task for clean history and rollbacks. Use /GSD:next anytime to auto-advance to the logical next step, sustaining momentum.
User-Focused Verification and Practical Trade-offs
Most AI workflows halt at compilation or passing tests, but /GSD:verify-work extracts and tests user-facing deliverables—like login flows, onboarding, dashboard states—spawning debug agents for fixes if needed, ensuring working software. Install via npx get-shit-done-cc@latest (Mac/Windows/Linux), selecting runtimes globally or per-project; Codex uses skill folders in .codex dir.
GSD suits solo devs and indie hackers tackling medium/large features (tens of thousands GitHub stars, MIT licensed), adding structure without enterprise bloat. Downsides: overkill for small tasks (e.g., bug fixes); requires clear requirements upfront; model costs rise with parallel agents/expensive models; terminal-heavy with learning curve; recommends --dangerously-skip-permissions on trusted machines for speed, but use caution. Repo: github.com/gsd-build/get-shit-done.