“The model lacks domain knowledge - do you fix the prompt, build RAG, or fine-tune?”
Decide between Prompting, RAG, or Fine-tuning for your AI feature
Recommended AI Agent Skills
Battle-tested prompt engineering patterns including few-shot demonstrations, role calibration, and schema constraints.
Prompt EngineeringLLMContext
Guidance for designing a RAG (Retrieval-Augmented Generation) pipeline - chunking, embeddings, hybrid search - and when RAG is actually the right call.
RAGRetrievalLLM
Advises when fine-tuning a model is actually warranted, which method to use (full fine-tune, LoRA, PEFT), and the ongoing maintenance cost.
Fine-tuningModel TrainingLLM
Action framework
- Ask: Is the issue missing knowledge or wrong format/style? New knowledge → Prompt/RAG; Format/tone → Few-shot or Fine-tuning.
- Ask: Does the data change frequently? If yes → RAG is required; fine-tuning will go stale immediately.
- Check document size: If under 10 pages, inject directly into the context window before engineering a RAG pipeline.
- Evaluate engineering cost: Prompting (1 day) → RAG (1-2 weeks) → Fine-tuning (1-2 months + continuous retraining).
- Always prototype cheapest-first: Optimize system prompt → Build RAG when exceeding context → Fine-tune only if RAG fails on specialized formatting.
Before
"Let's just collect all data and fine-tune a model for maximum accuracy."
After
"Docs change weekly → Use RAG with hybrid search. Injected 3 few-shot examples into the system prompt to hit 80% quality without expensive fine-tuning."