# Day 2 — AI Fundamentals, Prompt Engineering & Agents

- Core idea: context beats prompting. A mediocre prompt with the right source material is stronger than a polished prompt with no context.
- The model behaves like a capable new hire who has no access to the founder's files, history, inbox, or business until it is briefed.
- Context-engineering loop: gather real material, clean noise, load it into a persistent Project, then ask short questions.
- Useful source types: call transcripts, YouTube or podcast transcripts, Google Drive files, websites, dictation, and email threads.
- A Claude Project should contain three layers: short instructions, source knowledge, and two or three examples of the desired output.
- Minimum source mix for voice and positioning: one transcript, one careful document, and one public page.
- Instructions should state who the model is, who it writes for, voice rules, prohibitions, and the finish line for acceptable output.
- Chat is best for one-step answers or drafts. Agentic work is best when the task has dependent steps, touches tools or files, repeats, or needs checking.
- Claude Cowork handles multi-step tool work; Claude Code handles files and code with a higher ceiling and steeper setup.
- Safety rules: never paste passwords, API keys, bank details, or sensitive material into a public AI tool; verify specific claims before client use.
- Markdown is the portable structure shared by Claude, Obsidian, Notion, GitHub, and other tools.
- Homework: build a Project for Kasim Aslam from a transcript, document, and public page; test it against real source material.

## Relevance to the final project

- Build the Pareto Talent research brain from multiple source types before writing the offer, funnel, or ads.
- Keep facts traceable and flag unknowns instead of filling gaps with plausible claims.