A spec-driven development workflow system for Antigravity — featuring a fully autonomous ⚡ Super Mode, 🛡️ Anti-Hallucination Q&A, and 🧠 Project Memory protocols.
The complexity is in the system, not in your workflow.
Behind the scenes: context engineering, structured XML plans, state management.
What you see: a few commands that just work.
Built on Get Shit Done, solving context rot, hallucination, and quality drops across models.
npx get-shit-done-antigravityWorks on Mac, Windows, and Linux.
"I stopped losing work to context rot. Every phase picks up exactly where the last one left off."
"Claude hit quota, Gemini took over, and the code quality didn't drop. That's when I knew this was different."
"It's not just a planning tool — it actually catches when the AI is making stuff up."
🧪 I'm actively testing this in production and made it open source so others can test, improve, and contribute. — Built by @KhanXBT
How It Works · Commands · Why It Works · What's New · User Guide
GSD by TÂCHES is the best spec-driven development system for AI coding — but it only runs on Claude Code, OpenCode, and Gemini CLI.
GSD for Antigravity takes the methodology and makes it better. Same structured approach. Same atomic commits. Plus original features the original GSD doesn't have — because Antigravity's multi-model architecture creates problems that single-model tools never face.
- Project Memory System — Long-term context distillation via
/gsd-commit-memory. - Automated Verification Suite — Production-grade CLI tests now part of the core repository.
- Post-Install Guidance — Replicated the original GSD "Done! Run /gsd-help to get started" experience in all installers.
- README Rendering — Moved showcase image to
assets/with relative paths for GitHub/npm compatibility.
This isn't just a port. ~40% is adapted from the original GSD, ~60% is our own innovation built for Antigravity's multi-model reality:
| Feature | Origin | Details |
|---|---|---|
| Spec-driven workflow (plan → execute → verify) | ✅ Ported from GSD | Adapted for Antigravity's tool system |
| Atomic git commits per task | ✅ Ported from GSD | Same approach |
| XML task plans with verify/done criteria | ✅ Ported from GSD | Same structure |
Context freshness (/clear → new conversation) |
⚡ Adapted from GSD | Original uses /clear, we recommend new conversations |
| Source verification hierarchy | ⚡ Adapted from GSD | Original uses Context7 MCP, we use read_url_content + search_web |
| Checkpoints (wait for user) | ⚡ Adapted from GSD | Original uses AskUserQuestion, we use notify_user |
| 🛡️ Anti-Hallucination Engine | 🆕 Our Invention | Structural safeguards + /gsd-no-halluc for verified Q&A |
| 🧠 Project Memory | 🆕 Our Invention | Permanent context preservation via /gsd-commit-memory |
| 🤖 Multi-Model Compatibility | 🆕 Our Invention | Per-model hallucination pattern guards (Claude, Gemini, GPT, OSS) |
| 🔄 Model Resilience | 🆕 Our Invention | Code quality stays consistent when Claude quota → Gemini switch |
| 📋 Code Patterns in Plans | 🆕 Our Invention | Plans reference existing files as style templates for any model |
| 🎯 Style Anchoring Protocol | 🆕 Our Invention | Read existing code before writing — match patterns exactly |
| ✅ Full Verification Suite | 🆕 Our Invention | Lint + types + tests + build after EVERY task, not just plan verify |
⚡ Super Mode (/gsd-super) |
🆕 Our Invention | Full autonomy: prompt → production with zero human input |
The original GSD doesn't need these features because it only runs on Claude. Antigravity rotates between models — creating problems that single-model tools never face. We solve them.
The first "walk away" build engine for AI coding agents.
Give it a prompt or PRD, answer a few questions, and go grab a coffee. /gsd-super runs the entire lifecycle autonomously.
- PRD Intelligence: Auto-expands one-liners or parses large PRDs into roadmaps.
- Context Discovery: Auto-branches (
gsd-super/feature) for total safety. - 5-Tier Verification:
Syntax → Types → Lint → Tests → Buildon every atomic task. - Visual Autonomous Testing: The AI verified the UI matches the spec via browser screenshots.
The end of "plausible fabrication" in AI coding.
Use /gsd-no-halluc [question] (or alias /no-halluc) for verified technical research.
- Mandatory Research Loop: Forbidden from answering without proof (
search_web/read_url_content). - Confidence Scoring: Tagged HIGH, MEDIUM, or LOW with direct citations.
- Truth-First Protocol: Instructed to say "I don't know" rather than guess.
Long-term architectural continuity across conversation resets.
Use /gsd-commit-memory "Lessons learned..." (or alias /gsd memo) to save major pivots and decisions.
Models often forget subtle design constraints after a conversation reset. Project Memory distills your STATE.md, ROADMAP.md, and manual insights into a permanent log inside .planning/memory/. Future AI workflows are instructed to read this first, ensuring your project doesn't drift.
Clone and install:
Mac / Linux:
git clone https://github.com/KhanXBT/get-shit-done-antigravity.git
cd get-shit-done-antigravity
bash scripts/install.sh # Global install (all projects)
bash scripts/install.sh --local # Local install (current project only)Windows (PowerShell):
git clone https://github.com/KhanXBT/get-shit-done-antigravity.git
cd get-shit-done-antigravity
.\scripts\install.ps1 # Global install (all projects)
.\scripts\install.ps1 -Local # Local install (current project only)Or copy manually:
cp -r workflows/*.md ~/.gemini/antigravity/.agent/workflows/Inside Antigravity, type:
/gsd-help
🚀 Want full autonomy? Use
/gsd-superto build an entire project from a single prompt.🛡️ Want verified answers? Use
/gsd-no-halluc [question]for general Q&A with strict verification.🧠 Want context to stick? Use
/gsd-commit-memoryto save key decisions to long-term memory.
/gsd-new-project
Initialization: questions → research → requirements → roadmap.
/gsd-discuss 1
Capture implicit preferences (Layout, API structure, Naming) before planning.
/gsd-plan 1
Detailed XML implementation plans with Style Anchoring.
/gsd-execute 1
Wave-based execution with atomic commits and 5-tier verification.
/gsd-verify 1
User Acceptance Testing with automated gap analysis for fix-plans.
/gsd-commit-memory "Major decision: switched to SQLite for simplicity."
Lock in the context permanently.
| Command | Profile | What it does |
|---|---|---|
/gsd-super |
Autonomous | Build an entire project from a single prompt (Walk-away mode) |
/gsd-no-halluc |
Verified | Technical Q&A with mandatory external research and citations |
/gsd-commit-memory |
Continuity | Distill and save major project decisions to long-term memory |
| Command | Task | What it does |
|---|---|---|
/gsd-new-project |
Setup | Questions → Research → Requirements → Roadmap |
/gsd-discuss [N] |
Context | Capture implementation preferences for a phase |
/gsd-plan [N] |
Blueprint | Technical planning and research for a phase |
/gsd-execute [N] |
Builder | Autonomous execution with atomic git commits |
/gsd-verify [N] |
Audit | Human-in-the-loop verification and bug tracking |
| Command | What it does |
|---|---|
/gsd-quick [desc] |
Small ad-hoc task with GSD quality guarantees |
/gsd-progress |
Pulse check: Current state, roadblocks, and next steps |
/gsd-help |
Comprehensive manual of all commands and methodologies |
Instead of dumping everything into one long conversation where quality degrades, GSD structures project knowledge across purpose-built files. Each workflow loads only what's needed — keeping context focused and AI output quality high.
Plans are structured XML that gives the AI precise, unambiguous instructions. Each task specifies exact files, actions, verification criteria, and done conditions.
Every task gets its own commit — traceable, revertable, meaningful. No more "WIP" or "fix stuff" commits.
STATE.md tracks current position, key decisions, blockers, and quick tasks. The AI always knows where you are, what's been decided, and what's next.
Add, insert, or remove phases without rebuilding everything. Each phase is self-contained with its own context, research, plans, and verification.
GSD-Antigravity/
├── workflows/ — 9 Antigravity workflow definitions
│ ├── gsd-new-project.md
│ ├── gsd-discuss.md
│ ├── gsd-plan.md
│ ├── gsd-execute.md
│ ├── gsd-verify.md
│ ├── gsd-quick.md
│ ├── gsd-no-halluc.md
│ ├── gsd-progress.md
│ └── gsd-help.md
├── agents/ — 11 specialized agent prompts
│ ├── gsd-planner.md
│ ├── gsd-executor.md
│ ├── gsd-verifier.md
│ ├── gsd-debugger.md
│ ├── gsd-phase-researcher.md
│ ├── gsd-project-researcher.md
│ ├── gsd-research-synthesizer.md
│ ├── gsd-plan-checker.md
│ ├── gsd-integration-checker.md
│ ├── gsd-roadmapper.md
│ └── gsd-codebase-mapper.md
├── templates/ — 22 file templates for planning docs
│ ├── project.md
│ ├── requirements.md
│ ├── roadmap.md
│ ├── state.md
│ ├── context.md
│ ├── research.md
│ ├── phase-prompt.md
│ ├── summary.md
│ ├── config.json
│ └── ...
├── references/ — 13 reference docs for patterns & conventions
│ ├── questioning.md
│ ├── git-integration.md
│ ├── verification-patterns.md
│ ├── planning-config.md
│ ├── ui-brand.md
│ └── ...
├── docs/ — User documentation
│ └── USER-GUIDE.md
├── scripts/ — Installation scripts
│ └── install.sh
├── README.md
├── CHANGELOG.md
├── LICENSE
└── package.json
.planning/
├── PROJECT.md — Project vision and context
├── REQUIREMENTS.md — Scoped v1/v2 requirements
├── ROADMAP.md — Phases and progress tracking
├── STATE.md — Current position, decisions, session info
├── config.json — Workflow settings
├── research/ — Domain research
└── phases/
├── 01-phase-name/
│ ├── 01-CONTEXT.md
│ ├── 01-RESEARCH.md
│ ├── 01-01-PLAN.md
│ ├── 01-01-SUMMARY.md
│ ├── 01-VERIFICATION.md
│ └── 01-UAT.md
└── 02-phase-name/
└── ...
The #1 reason AI coding assistants fail is hallucination — fabricated APIs, assumed test results, invented library syntax. GSD for Antigravity has structural anti-hallucination safeguards baked into every workflow:
| Protection | What it prevents |
|---|---|
| File-First Context | Always re-reads files — never relies on memory of them |
| Source Verification | Technical claims verified via docs before use in plans |
| Confidence Levels | Research tagged HIGH/MEDIUM/LOW by verification source |
| Verification Gates | Every task verified — output read, not assumed |
| No Auto-Pass | Tests never marked passed without user confirmation |
| Decision Attribution | Tracks USER-decided vs AI-suggested decisions |
| Context Freshness | New conversation recommended between workflow steps |
| Checkpoint Integrity | WAIT for user — never hallucinate completion |
These aren't suggestions — they're embedded instructions inside every workflow file that the AI must follow.
Unlike the original GSD (Claude-only), GSD for Antigravity works across all AI models because Antigravity rotates between them:
| Model | Common Hallucination Pattern | GSD Protection |
|---|---|---|
| Claude (Anthropic) | Polite hallucination — agrees with wrong assumptions | Decision capture integrity, no-lead questioning |
| Gemini (Google DeepMind) | Plausible synthesis — mixes real and fabricated info | Source verification hierarchy, confidence tagging |
| GPT (OpenAI) | Confident fabrication — invents APIs that don't exist | Library verification gates, URL checking |
| Open-source | Higher baseline hallucination rate | Structural guardrails, mandatory file re-reads |
Model-agnostic by design: uses structural safeguards (file reads, command verification, user gates) instead of relying on model-specific behavior.
See
references/anti-hallucination.mdfor the comprehensive guide.
"When Claude's model quota is exceeded in Antigravity and we switch to Gemini, the quality of code decreases — so what can we do so the code quality and workflow isn't affected with model changing?"
Every Antigravity user faces this. GSD solves it with plan-driven quality — the plan itself is the quality guarantee, not the model:
| Strategy | What It Does |
|---|---|
| Step-by-step plans | Instructions so detailed any model can follow them correctly |
| Style anchoring | Read existing code before writing — match patterns exactly |
| Code pattern refs | Every plan references existing files as style templates |
| Full verification suite | Lint + type-check + tests + build after EVERY task |
| Atomic task sizing | One function per task — less room for quality drift |
| Code review gates | Human review for complex logic, regardless of model |
The result: Whether Claude, Gemini, GPT, or an open-source model writes the code, every commit passes the same quality bar.
See
references/model-resilience.mdfor full strategies.
| Feature | Original (Claude Code) | Antigravity Port |
|---|---|---|
| Subagent spawning | Task() API with parallel agents |
Sequential execution within task boundaries |
| CLI tooling | gsd-tools.cjs for state management |
Direct file reads/writes via Antigravity tools |
| User interaction | AskUserQuestion structured API |
notify_user + natural conversation |
| Execution model | Parallel waves across agents | Sequential task execution, same structure |
| Model profiles | Quality/balanced/budget selection | Antigravity's default model |
| Slash commands | /gsd:command syntax |
/gsd-command workflow syntax |
- GSD by TÂCHES — the methodology, philosophy, agent prompts, templates, and workflow design
- Antigravity port — adapted to work natively within Google DeepMind's Antigravity AI coding assistant
MIT License. See LICENSE for details.
Antigravity is powerful. GSD makes it reliable.