Prompt Engineering Lesson 6: Long Documents and Summaries
Summarise and question long documents: summary types, asking for quotes and page references, chunking very long files, comparing versions and meeting-minutes prompts.
Read article →Summarise and question long documents: summary types, asking for quotes and page references, chunking very long files, comparing versions and meeting-minutes prompts.
Read article →Ask for answers you can use directly: markdown tables to paste into Excel, CSV, JSON with a fixed schema, and extracting fields from messy text like invoices and emails.
Read article →Get better answers on calculations and decisions: ask for steps, break big tasks into stages, have the model check its own work, and verify numbers yourself.
Read article →Show, don’t tell: give the model one to five examples to copy tone, structure or classification rules, and build reusable prompt templates with placeholders.
Read article →Write prompts that work first time with the role-task-context-format pattern, constraints, audience and length, plus before/after examples for office tasks.
Read article →What happens when you send a prompt: tokens, context windows, next-word prediction, why wording matters, why answers vary, and what the model does and doesn’t know.
Read article →LLM "hallucinations" explained simply: why a next-token predictor can sound confident and still be wrong, and six practical habits to catch fabricated facts, formulas and citations.
Read article →Install Ollama, run open models like Llama and Qwen locally, call them from Python and Excel-friendly scripts, and create a custom model with a Modelfile. Hardware guide included.
Read article →Combine everything into one always-on script: text snippets, app launchers, Excel helpers, clipboard cleaners, a tray menu and start-with-Windows setup.
Read article →Ten ready-to-use AutoHotkey v2 scripts: text expansion for emails and GSTINs, today’s date, always-on-top windows, clipboard cleanup, Excel shortcuts, window snapping and more — with setup steps.
Read article →How LoRA fine-tuning works, why it trains less than 1% of a model’s parameters, when fine-tuning beats prompting or RAG, and a minimal Hugging Face PEFT setup.
Read article →Step-by-step guide to retrieval-augmented generation in Python: load PDFs and Excel sheets, chunk them, create embeddings, search by meaning, and answer with a local LLM — with citations and evaluation.
Read article →