smol developer
**Smol Developer** is a **minimalist open-source AI coding agent created by Shawn Wang (swyx) that generates entire codebases by packing the full project context into a single LLM prompt** — intentionally kept under 200 lines of Python to demonstrate that agentic coding doesn't require complex frameworks like LangChain, proving that large context windows (100K+ tokens in GPT-4 and Claude) enable simple, effective AI software development without vector databases or RAG pipelines.
**What Is Smol Developer?**
- **Definition**: A tiny AI agent (~200 lines of Python) that takes a project description, generates a file manifest, and loops through each file — generating complete implementations with awareness of all other files in the project, exploiting large context windows to maintain coherence across the entire codebase.
- **Philosophy**: "The entire codebase is the prompt." Instead of complex retrieval systems (vector databases, chunking strategies), Smol Developer dumps the full list of file paths and shared dependencies directly into the LLM's context window. Simple but effective.
- **"Smol" Approach**: Intentionally minimal — no LangChain, no vector stores, no agent frameworks. Just Python, an LLM API call, and a loop. This proves that the "magic" of AI coding is in the LLM itself, not the orchestration framework.
**How Smol Developer Works**
| Step | Action | Implementation |
|------|--------|---------------|
| 1. **Describe** | User provides project description | Plain text spec |
| 2. **Plan** | LLM generates file manifest | List of filenames to create |
| 3. **Context Pack** | All file names + shared deps in prompt | Fits in 100K context window |
| 4. **Loop** | Generate each file, aware of others | Sequential file creation |
| 5. **Output** | Complete project directory | Ready to run |
**Key Insight**: With 100K+ context windows (Claude, GPT-4), you don't need RAG, vector databases, or complex retrieval. For projects under ~50 files, you can fit the entire context in a single prompt — making the "smol" approach surprisingly effective.
**Why Smol Developer Matters**
- **Framework Skepticism**: Demonstrated that complex AI agent frameworks (LangChain, AutoGPT) are often unnecessary — a simple loop with good prompting achieves comparable results with 200 lines instead of 20,000.
- **Context Window Advocacy**: Proved that **large context windows are the simplest path to coherent code generation** — no need for embedding-based retrieval when you can just include everything.
- **Educational Value**: The tiny codebase is readable in 10 minutes — ideal for understanding how AI coding agents work under the hood.
- **Inspired Practical Tools**: The "smol" philosophy influenced Aider, Continue, and other practical tools that prioritize simplicity over framework complexity.
**Smol Developer is the minimalist manifesto for AI coding agents** — proving in under 200 lines of Python that large context windows and simple prompting are sufficient for coherent multi-file code generation, challenging the assumption that agentic AI requires complex orchestration frameworks.