Choose few-shot prompting when you need flexibility and rapid iteration—you're trying different approaches to explain your background or answer interview questions; choose fine-tuning when you have a consistent body of your own writing and need the system to deeply understand your voice and values. Most people benefit from prompting first.
Most people using AI for reentry support will work with prompting—carefully crafting instructions to get the output they need. But a subset of advanced users might consider fine-tuning: updating an AI model's weights based on examples specific to your context. Understanding the difference and trade-offs is essential for making smart decisions about AI customization.
Prompting is crafting instructions within a single conversation. You tell Claude or ChatGPT exactly what you want: "Write a background explanation letter addressing a 5-year employment gap due to incarceration. Emphasize my growth and the skills I developed." The model generates output based on that prompt without permanent changes to itself.
Prompting's advantages: zero cost beyond standard API usage, immediate results, easy iteration (change your prompt, regenerate). For reentry support, you can maintain different prompts for different situations—one for explaining gaps to traditional employers, another for background-friendly companies, another for cover letters targeting diversity initiatives.
The limitation: prompting relies on the model's base training. If you're repeatedly giving the same context ("I completed a 2-year reentry program at XYZ nonprofit, worked in their job training track"), you need to restate it every time. The model doesn't learn your specific situation—it remains general.
Fine-tuning is training a model on your specific examples until it internalizes your context and style. You'd provide dozens of background explanations you've written, reference letters from mentors, interview responses—whatever demonstrates your voice and situation. The model then updates its weights to reflect those patterns.
After fine-tuning, you prompt the model without restating context: "Generate a background explanation for a tech role," and the model understands your reentry background, your values, your communication style without you restating it each time. The model has learned you.
Fine-tuning becomes worthwhile when: (1) you're generating many outputs over months and restating context repeatedly is inefficient; (2) you need consistent voice and values reflected across dozens of applications; (3) you're working with a team (reentry coordinator, mentor, employment specialist) and want a customized model reflecting shared standards.
The trade-offs: fine-tuning is expensive ($25–$100+ per training run for small models), requires 10–20 quality examples minimum, and adds complexity. Once fine-tuned, your model is less flexible—it's optimized for your specific background, which makes it harder to adapt to unexpected interview questions or novel scenarios.
Most reentry candidates should start with sophisticated prompting: develop a few detailed system prompts that include your background, values, and style preferences, then reuse them across conversations. This achieves 80% of fine-tuning's benefits without the cost or complexity.
For example, instead of fine-tuning, create a master prompt: "You are helping someone reentry into employment after a period of incarceration. They're skilled in [skills], hold [certifications], and value [values]. Generate materials emphasizing growth and concrete capabilities while being honest about their background." Paste this into Claude and reference it for each new task.
Fine-tuning updates model weights directly, making outputs stable but less exploratory. Prompting keeps the base model intact, allowing flexibility but requiring more detailed instructions each time. OpenAI and Anthropic now offer "custom models" (fine-tuned variants), but these are intermediate between basic prompting and full fine-tuning.
Try this: Develop a detailed system prompt capturing your reentry story, skills, and values (300–400 words). Use it consistently across 5–10 applications, iterating on the prompt based on results. Track whether hiring response improves as you refine your prompt. If after a month you're generating 10+ outputs weekly, revisit fine-tuning as a potential next step.
Peri can explain this concept, give practical examples, help you decide whether it applies to your situation, or recommend a journey if appropriate.
Explore related journeys or tell Peri what you're working through.