GitHub - gsd-build/get-shit-done: A light-weight and powerful meta-prompting, context engineering and spec-driven development system for Claude Code by TÂCHES.

Source
claude-codecontext-engineeringspec-driven-developmentmeta-promptingai-codingsubagent-orchestrationvibecoding

Summary

GSD is a meta-prompting system for Claude Code that solves context window degradation by breaking work into atomic plans executed in fresh subagent contexts. It provides a full lifecycle workflow (discuss, plan, execute, verify, ship) with parallel wave execution, keeping the main context at 30-40% usage while orchestrating thousands of lines across parallel executors.

Key Insight

  • Context rot is the core problem being solved - as Claude fills its context window, output quality degrades. GSD keeps the orchestrator thin (30-40% context usage) while spawning fresh 200k-token subagent contexts for actual implementation work.
  • Wave-based parallel execution - plans are grouped into dependency waves. Independent plans run in parallel within a wave, dependent plans wait for earlier waves. Vertical slices (end-to-end features) parallelize better than horizontal layers (all models, then all APIs).
  • Atomic commits per task - each task gets its own git commit, making bisect possible and individual tasks revertable. Commit format: feat(phase-plan): description.
  • Multi-runtime support - works with Claude Code, OpenCode, Gemini CLI, Codex, Copilot, Cursor, and Antigravity via a single npx installer. Not locked to one AI coding tool.
  • Config-driven quality/cost tradeoff - model profiles let you assign Opus for planning and Sonnet for execution (balanced), or go full Opus (quality) or Sonnet+Haiku (budget).
  • Built-in security - path traversal prevention, prompt injection detection, and shell argument sanitization since v1.27. Important because planning artifacts (markdown) become LLM system prompts, creating indirect injection vectors.
  • The discuss phase is underrated - captures implementation preferences before planning. Planner reads locked decisions (“user wants card layout”) so it doesn’t guess. Skipping it gives reasonable defaults; using it gives your specific vision.