The elvex Prompting & AI-Assistant Guide
PDF · 15 pages ↓ Learning to prompt well takes most people about 10 hours of experimentation. This guide is meant to shorten that. It follows a three-step learning curve: chat with LLMs, bring your own data to them, then turn a reliable workflow into an assistant your colleagues can use.
What’s inside
- How LLMs work and why they hallucinate. Models predict the next likely words. They make things up when they lack the information, and attaching the right documents fixes much of that.
- Which model to use (updated November 2026). Recommendations by job (hardest work, everyday writing and coding, cheap and fast tasks, long inputs) with current prices from models.dev and Intelligence Index scores from Artificial Analysis, plus a worked monthly cost example.
- Prompting basics. Iterate, give the model a role, balance specificity and ambiguity, use examples, and ask it to critique its own answer. One rule stands out: give more context and fewer tasks. Five prompts with one instruction each often beat one prompt with five.
- Five worked examples. Sample prompts and expected results for customer support, API documentation, tutoring, symptom triage and contract summaries.
- Grounding with your data. A plain explanation of retrieval augmented generation (RAG), when it works and when it doesn’t. It includes tips on file formats: avoid PDFs when another format exists, and keep spreadsheets to one simple table per sheet.
- Building an elvex assistant. The seven configuration steps, from rules and data sources to tools, permissions, Slack and model settings.
- Tool calling. How to name specific tools in a prompt (for example
create_draft,run_sql_on_datasource_file,search_google) to make shared assistants more reliable.
Key takeaway
Good prompting comes from understanding how the models work and refining by trial. Clear instructions, relevant context, examples and iteration matter more than any magic phrase.