Get Shit Done (GSD)

GSD is a meta-prompting and context engineering system for AI coding assistants. It solves context rot — the quality degradation that occurs as AI fills its context window during long development sessions.

Install: npx get-shit-done-cc@latest Repo: https://github.com/gsd-build/get-shit-done Compatible with: Claude Code, OpenCode, Gemini CLI, Cursor, Windsurf

Core Problem It Solves

Long-running AI coding sessions degrade in quality as context fills up. GSD combats this by:

  • Externalising project state to structured files (PROJECT.md, REQUIREMENTS.md, ROADMAP.md, STATE.md)
  • Running heavy work in fresh subagent contexts, keeping the main window at 30–40% capacity
  • Enforcing structured phases so AI never does more than one thing at a time

Six-Stage Workflow

StageCommandPurpose
Initialize/gsd-new-projectGather requirements, generate roadmap
Discuss/gsd-discuss-phaseCapture implementation decisions pre-plan
Plan/gsd-plan-phaseCreate verifiable task plans
Execute/gsd-execute-phaseParallel execution with atomic commits
Verify/gsd-verify-workManual acceptance testing and diagnosis
Ship/gsd-shipPRs, milestones, release management

Key Concepts

  • Context engineering: Structured files survive session boundaries; AI reads them on wake-up
  • Parallel execution waves: Independent tasks run simultaneously in fresh subagent contexts
  • Multi-agent orchestration: Distributes work across agents to prevent single-window overload
  • Atomic commits: Each plan step is committed independently for safe rollback
  • Profile-based installation: Core / Standard / Full feature sets; quality/balanced/budget model profiles

Notable Features

  • Package legitimacy checks during research/planning phases (supply chain awareness)
  • Optional “fallow structural review” pre-pass for code quality
  • Interactive and autonomous (--dangerously-skip-permissions) modes
  • Skills install to ~/.claude/skills/gsd-*/ for Claude Code

Assessment

Strong approach for solo devs and small teams doing multi-session AI-assisted development. The context engineering discipline (externalising state, capping window usage) is the key insight — it’s a systematic answer to a real and widely felt problem. The six-stage structure forces deliberate planning before execution, which reduces AI hallucination and drift.