Prompt Fundamentals

📘 Learn Prompt Fundamentals from zero

Start from zero: a prompt is the instruction you hand an AI model. Within a single call the model sees only the words in front of it — it cannot infer your goal, your stack, or your standards. So the quality of the output is capped by the quality of the prompt.

Analogy: treat the model as a fast, capable freelancer who starts the instant you message and never asks a clarifying question. Say "make me a logo" and you get a logo — probably not yours. Say "a minimalist logo for a fintech startup named Ledger, navy and white, no text, conveying trust" and you get something usable. Same freelancer; the brief is the whole difference.

The canonical scaffold is RTCF: role (who the model acts as), task (the precise action), context (facts it can't infer), and format (the shape of the output) — plus a fifth lever, constraints (length limits, guardrails, what to refuse).

Worked example. Weak: "Help with my resume." Strong: "You are a senior backend hiring manager. I have 3 years in Java/Spring. Rewrite this bullet — 'worked on the payments API' — as one STAR-format line, under 25 words, leading with an action verb and ending with a quantified result." The strong version names a role that can actually judge the work (a hiring manager assesses bullet quality; a recruiter screens for keywords), supplies context, gives one crisp task, and pins the format and constraints. The output goes from filler to interview-ready.

Key insight: the model mirrors your specificity — every ambiguity you leave becomes an assumption it makes for you. Close the ambiguities and you control the result.

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

🎯 Guided practice

1. Easy — turn a vague ask into a structured prompt.

   Goal: get a strong GitHub README from "write a README for my project."

   Step 1 — add a role: "You are an open-source maintainer reviewing a project for newcomer clarity."

   Step 2 — add context the model can't infer: "Project: TaskFlow, a CLI task manager in Python; stores tasks in local JSON; single-user."

   Step 3 — pin the task and format: "Write a README with these sections, in order: one-line description, Install, Usage example, Features (bulleted)."

   Pattern learned: Role + Task + Context + Format. Each element you add removes one assumption the model would otherwise make for you.

2. Medium — design a reusable, parameterized template for portfolio bullets.

   Goal: a prompt you can rerun for any project, not a one-off.

   Step 1 — identify the variables that change per run (target role, the raw note) and mark them with explicit placeholders.

   Step 2 — write the template: "You are a senior hiring manager for a {ROLE} position. Convert this raw note into one STAR-format resume bullet, under 25 words, leading with an action verb and ending with a quantified result. Note: {RAW_NOTE}."

   Step 3 — add a guardrail (constraint): "If the note contains no metric, ask me for one instead of inventing a number." This blocks fabricated stats — a fireable problem on a real resume, and a known LLM failure mode (hallucination).

   Step 4 — test on two different inputs and confirm both obey the word limit and format. If one drifts, tighten the constraint wording or add a one-shot example of a correct bullet.

   Pattern learned: promote a good one-off into a parameterized template with explicit constraints — the prompt analog of writing a reusable function instead of copy-pasting code.

3. Harder — use a few-shot prompt when the shape is hard to describe.

   Goal: generate consistent project-pitch lines for a portfolio when adjectives alone keep missing the mark.

   Step 1 — recognize the case: you've described the format in words ("punchy, one sentence, impact-first") and the model still drifts. The fix is to show, not tell.

   Step 2 — supply 2-3 examples of input-to-output, then the real input: "Examples — Note: built a URL shortener → Pitch: 'Shipped a URL shortener handling 10k redirects/day at p99 under 20ms.' Note: made a chat app → Pitch: 'Built a real-time chat app for 500 concurrent users over WebSockets.' Now do this one — Note: {RAW_NOTE}."

   Step 3 — keep examples representative: the model pattern-matches on them, so a sloppy example teaches sloppy output. Two clean examples beat a paragraph of instructions.

   Pattern learned: zero-shot (instruction only) is the default; switch to few-shot (instruction + worked examples) when the desired shape is easier to demonstrate than to specify. This is a distinct prompt type, not just a longer prompt.

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