Constraints That Cut Code.
4 gates. 90% less AI-generated bloat.
AI generates bloated code: error handlers for impossible states, interfaces for single implementations, configuration for fixed values.
Most of it is anticipatory—code written for scenarios that will never occur.
Same prompt. Same AI. Left: without gates (75 lines). Right: with gates (7 lines).
| Task | Without Gates | With Gates | Reduction |
|---|---|---|---|
| Env Validator | 75 lines | 7 lines | 91% |
| Health Checker | 166 lines | 14 lines | 92% |
| SMS Service | 249 lines | 28 lines | 89% |
| Total | 490 lines | 49 lines | 90% |
Same AI. Same prompts. Different instructions.
Sequential, blocking constraints. Each gate poses a question—if unanswered, the task halts.
| Gate | Blocking Question | Proceeds When |
|---|---|---|
| 1. Spike | Is the approach validated? | Uncertainty resolved via minimal experiment |
| 2. TDD | What failing test requires this code? | Test exists and fails |
| 3. YAGNI | Is this in the current requirements? | Explicit requirement confirmed |
| 4. Simple Design | Is this the simplest passing implementation? | No simpler alternative exists |
Gates evaluate in order. Blocked at Gate 1? Never reaches Gate 2.
The 90% reduction comes from elimination, not optimization.
| Gate | What It Blocks |
|---|---|
| Spike | Wrong approaches—blocks implementation before validation |
| TDD | Defensive code—blocks code without failing tests |
| YAGNI | Speculative features—blocks code without explicit requirements |
| Simple Design | Abstraction layers—blocks indirection without justification |
Most AI-generated bloat is anticipatory. XP Gates block anticipation at the source.
The 90% reduction is real—but not all eliminated code is waste. XP Gates enforce strict interpretation: if it's not in the requirements, it doesn't get built.
- Ambiguous prompts → minimal output
- Convention-based features get cut (docstrings, convenience returns)
- You must specify what you actually need
XP Gates don't read minds. They read requirements.
See docs/RESEARCH.md for detailed tradeoff analysis and examples.
Add the prompt file to your AI's system instructions. No tooling required.
curl -o CLAUDE.md https://raw.githubusercontent.com/en-yao/xp-gates/main/prompts/CLAUDE.mdcurl -o GEMINI.md https://raw.githubusercontent.com/en-yao/xp-gates/main/prompts/GEMINI.mdCopy prompts/CHATGPT.txt → Settings → Custom Instructions
- Add prompt file to system instructions
- Every task passes through all four gates
- Gates block silently—no code until all pass
| Format | Platform |
|---|---|
CLAUDE.md |
Claude Code |
GEMINI.md |
Gemini CLI |
CHATGPT.txt |
ChatGPT |
See prompts/README.md for platform differences and usage details.
The constraint set is fixed. Customization defeats the purpose.
XP Gates derive from Extreme Programming (Kent Beck, 1999):
| XP Practice | XP Gate |
|---|---|
| Spike Solutions | Spike Gate |
| Test-First Programming | TDD Gate |
| You Aren't Gonna Need It | YAGNI Gate |
| Simple Design | Simple Design Gate |
Key distinction: XP is a human discipline requiring judgment. XP Gates are machine-readable constraints—binary checks that block progress until satisfied.
XP assumes practitioners internalize principles. XP Gates assume nothing.
See examples/ for actual code comparisons:
| Directory | Description | Lines |
|---|---|---|
with-gates/ |
Code produced with XP Gates | 49 |
without-gates/ |
Code produced without gates | 490 |
Methodology and detailed test results: docs/RESEARCH.md
MIT