Prompt Engineering for Business AI Assistants

The AI assistant on this portfolio's Contact section answers questions about my background using only the resume content on the page — no web search, no general knowledge. That constraint is the whole design problem: how do you keep an LLM from confidently making things up when it doesn't know the answer?

Ground the model in a fixed context, not instructions alone

The naive approach is a system prompt that just says "only answer from what you know about Ashutosh." That's not grounding — it's a suggestion the model will happily ignore under ambiguity. The fix is to pass the actual source content as part of the system instruction, every request:

const systemPrompt = `You answer visitor questions about Ashutosh's background
using ONLY the resume information provided below. If asked something not
covered by this information, say you don't have that detail.
 
${resumeContext}`;

The resume context itself should be generated from the same structured data that renders the page — not hand-copied into a separate string. That second copy is where staleness bugs creep in: add a new project to the page, forget to update the chatbot's context, and now the assistant is confidently wrong about what you've built.

Keep the refusal path explicit

Models are more likely to hallucinate when the prompt doesn't give them an easy, sanctioned way to say "I don't know." Naming the exact fallback behavior removes the ambiguity:

If asked something not covered by this information, say you don't have
that detail and suggest they email ashunayak6393@gmail.com. Never invent
facts, employers, dates, or skills that aren't listed.

Cap history, not just message length

Multi-turn context matters for a conversational assistant, but unbounded history is both a cost problem and a drift problem — the model starts anchoring on earlier turns instead of the grounding context. Capping to the last N turns keeps the conversation coherent without letting it wander:

const MAX_HISTORY_TURNS = 8;
const priorContents = history.slice(-MAX_HISTORY_TURNS * 2);

Test the refusal, not just the happy path

Most prompt iteration time should go into questions the model shouldn't answer — "what's Ashutosh's salary expectation," "does he know Rust," anything outside the provided context. If the refusal path holds up under adversarial questions, the happy path usually takes care of itself.

A short checklist

• Ground the model in the actual source data, not a hand-maintained duplicate.
• Give the model an explicit, named way to say "I don't know."
• Cap conversation history to avoid cost blowup and context drift.
• Spend testing time on the refusal path, not the happy path.