Start from the objective

Most people start a local AI project with a model or a shiny tool. The method that actually works starts one step earlier: with the objective. This is the difference between "I wired a RAG to a corpus" and "I solved a problem, and I can prove it."

The first question

Before any hardware or model choice, write down:

• What do I want to achieve? In one sentence, the outcome, not the mechanism. "Reduce the time to draft a client report" is an objective. "Run a local LLM" is not.
• Why? What is the cost of doing it the current way, or the value of doing it faster, cheaper, or more privately?
• How will I know it worked? A number, not a feeling. "Report drafts go from 2 hours to 30 minutes" or "answers cite the right document 90% of the time."

If you cannot answer those three, you are not ready to build. You are about to spend effort and money on a solution to a problem you have not stated.

Why this first

Every later decision flows from the objective:

• Local vs cloud depends on whether privacy, cost or control is the point.
• Model size depends on how much quality the task genuinely needs.
• RAG vs fine-tune vs prompt depends on whether the gap is factual or behavioural.
• The evaluation depends on the "how will I know it worked" line you wrote now.

Without an objective, these become taste. With one, they become testable choices.

The trap

The trap is starting from the tool, because the tool is exciting. "I want to build a RAG" is a tool statement. "I want my team to find answers in 200 past project documents" is an objective, and it tells you the RAG is probably the means, not the end.

Write the objective first, keep it in front of you, and let every choice below it be an answer to that objective. The next lesson turns the objective into a concrete design.