AI & Prompt Basics

📘 Learn AI & Prompt Basics from zero

Start from zero. A Large Language Model (LLM) like ChatGPT is, at its core, a very sophisticated autocomplete. You give it text (the prompt), and it predicts the most likely next sub-word unit (a token), appends it, and repeats — token by token — until it has produced a full answer. It has not memorised facts the way a database has; it has learned statistical patterns of how words fit together from enormous amounts of text. That single mental model explains almost everything else.

Why it understands context. ChatGPT is built on the transformer architecture, whose attention mechanism lets it weigh every word in your prompt against every other word at once — so it reads the whole sentence, paragraph, and conversation, not one word in isolation. That is why it can stay on-topic and "remember" what you said a few turns ago: that earlier text is still inside its context window.

Analogy: think of the model as an extremely well-read intern on their first day. They write fluently and have seen millions of resumes — but they know nothing about YOU until you brief them. A weak brief ("write my bio") yields a generic bio. A strong brief ("write a 3-sentence bio for a backend engineer applying to fintech; emphasize a payments project that handled 10k requests/sec; confident, not boastful") yields something usable. Prompt engineering is just learning to brief that intern well, then iterating.

Worked example. Weak: Write my LinkedIn summary. → filler. Strong adds context the model can't know: "Write a 50-word, first-person LinkedIn summary for a new-grad software engineer. Facts: built a React expense tracker used by 200 students; strong in JavaScript and SQL. Tone: confident, no clichés like 'passionate'." The output is now specific and on-voice — because you supplied the missing context.

Key insight: the model supplies fluency; you supply the facts, judgment, and verification. Output quality is a direct function of the context and constraints you put into the prompt.

✨ Added by the guide to build intuition — not from the source course.

🎯 Guided practice

1. Easy — rewrite a flat resume bullet.

   Start: Worked on the website. Goal: a quantified, action-led bullet.

   Step 1 — name the pattern. Strong bullets follow action verb + what you did + measurable impact. The model can format this, but only you know the numbers, so feed them.

   Step 2 — write the prompt with context. Rewrite this resume bullet as "action verb + task + quantified impact". Facts: I rebuilt the checkout page in React; page load dropped from 4s to 1.5s; conversions rose ~12%. One line, no clichés.

   Step 3 — expected output: "Rebuilt the checkout flow in React, cutting page-load time from 4s to 1.5s and lifting conversions ~12%."

   Step 4 — verify. Check every number is real, then edit to your voice. The reusable move is: give facts, then ask for a specific format.

2. Medium — tailor and iterate using context (the lingo in action).

   Goal: a tailored "About Me" plus a behavioral-question set for a specific job, refined over a few turns.

   Step 1 — set the context once. The model carries context across the conversation, so open with it: "I'm a backend engineer, 2 years, Java + Postgres, targeting fintech roles focused on payment reliability and on-call. Use this for everything that follows." You won't need to repeat it every turn.

   Step 2 — first task. "Write a 60-word first-person About Me for that profile. Keep temperature low — I want it consistent and on-message, not flowery." (Low temperature = less random, steadier wording — right for professional copy.)

   Step 3 — iterate, don't restart. The book's whole point is iterative content improvement: "Tighten to 45 words and lead with the payments work." Refining beats re-prompting from scratch.

   Step 4 — generate the question set, grounded in the JD. Paste 3-4 lines of the real job description, then: "Now give me 5 behavioral questions targeting those exact responsibilities, each with S/T/A/R prompts as bullets." Pasting the JD is the context that makes them job-specific. (Adding one example answer first — few-shot prompting — locks the format even tighter; you'll meet it formally in later lessons.)

   Step 5 — mind the window, then verify. If the thread gets long, the earliest turns can drop out of the context window, so re-paste your project facts before asking for a STAR draft: "Draft a STAR answer for question 3 using: [paste facts]." Then edit into your own voice — the model drafts, you own the truth.

✨ Added by the guide — work these before the full problem set.