There's a reflex to treat fine-tuning as the "real" AI engineering and prompting as a shortcut. That's backwards. Both are tools, and most features should start with the cheaper one. The decision isn't about sophistication - it's about what the task actually needs.
Start with a prompt; reach for fine-tuning when consistency is the product
A well-written prompt - anchored to a rubric, constrained to a schema - handles a surprising amount. I reach for fine-tuning when the same task runs at volume and the output has to be consistent in a way prompting keeps drifting on. Grading is where I expected to need it, and did not: one institution's rubric, applied the same way every time. A prompt built from the rubric as structured JSON, a strict JSON reply with a score per criterion, and a professor approving every grade gave the consistency the work needed, with no training run at all.
Prompt when you can describe what you want. Fine-tune when you need the same judgment, the same way, ten thousand times.
And sometimes the answer is neither
The third option is the one people forget: don't train at all. For a well-solved perception task, a recognition API beats training your own model from scratch-you'd spend months to match what's already a call away. The skill is knowing which bucket you're in: a task you can describe (prompt), a task that needs consistent learned judgment (fine-tune), or a solved problem someone already trained (use the API).
My decision checklist:
- Can you describe the task in words? Start with a prompt and a schema.
- Is it high-volume and consistency-critical? Fine-tune on real examples.
- Is it a solved perception problem? Use an API; don't train from scratch.
- Whatever you pick, build the eval first-it's how you'll know the choice was right.