Advanced Prompting

📘 Learn Advanced Prompting from zero

Start from zero. A large language model is a next-token predictor: given the text in front of it, it samples a probable continuation. It has no persistent memory of you across separate chats — its entire working "memory" is the context window, the run of text in the current thread. Add a correction to that thread and the model now conditions on it; start a fresh chat and that signal is gone. So the quality of what you get is almost entirely set by the quality of what is in the window. "Advanced prompting" is the discipline of shaping that text so the probable continuation is the one you actually want.

Analogy: think of the model as a brilliant, literal-minded contractor who just walked onto your job site with zero briefing. Say "build me something nice" and you get a generic shed. Hand over a blueprint, the lot dimensions, the budget, and a photo of the style you like, and you get the house you imagined. The three lessons map directly: Scenario-Based Prompting is writing the blueprint (role, situation, constraints). Iterative Feedback Loops is the walkthrough where you point at the wall and say "move this two feet left" instead of demolishing and rebuilding. Troubleshooting Pitfalls is diagnosing why a wall came out wrong — vague spec, contradictory orders, or a brief so long the contractor lost the plot.

Worked example. Zero-shot: "Write a resume bullet about my project." Output: Worked on a project that improved system performance. Now apply a scenario: "You are an Amazon bar-raiser screening for SDE-II. Rewrite this as one bullet using action verb + quantified metric + scope." Output: Cut p99 checkout latency 38% by adding a Redis read-through cache, serving 12M requests/day. Same model, same facts — the scaffolding changed which continuation was most probable.

Key insight: you are not "asking" the model, you are conditioning its output distribution on the context you supply. Specificity in equals signal out.

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

🎯 Guided practice

1. Easy — turn a weak bullet strong (Scenario-Based). Starting point: Helped with the migration. Step 1: identify what context is missing — role/audience, the metric, and the scope. Step 2: name the scenario: "You are a FAANG hiring manager screening for a senior data engineer." Step 3: impose a structure — action verb + quantified impact + technical scope. Step 4: supply the raw facts (migrated 4TB Postgres to Snowflake, 6-week deadline, zero downtime). Result: Led a zero-downtime migration of 4TB from Postgres to Snowflake in 6 weeks, cutting analytics query cost 45%. Pattern learned: role + structure + facts beats a bare request every time.
2. Medium — recover a stalled STAR answer (Iterative Loop + Troubleshooting). You prompt for a "Tell me about a conflict" story and get a vague, hero-narrator paragraph with no result. Step 1: diagnose the pitfall — the prompt never specified the STAR frame, so the model defaulted to a generic essay (ambiguity pitfall). Step 2: do not restart the chat — keep the thread so the facts you already gave stay in the context window. Step 3: give a delta correction, one constraint at a time: "Restructure as Situation / Task / Action / Result; target ~90 seconds spoken." Step 4: read the new output and apply a second delta only if needed: "The Result is qualitative — add one concrete metric." Step 5: stop when it converges; resist endless tweaking (diminishing returns). Pattern learned: iterate with small, single-purpose deltas inside one context window, and trace a bad answer to its specific cause — ambiguity, conflict, or overload — before re-prompting.
3. Hard — fix a prompt that's failing for two reasons at once (all three lessons). Setup: you paste your whole resume plus a JD and ask "tailor my resume to this role," and the model returns generic claims, invents a tool you never used (hallucination), and ignores the seniority signal. Step 1: isolate the pitfalls. Two are stacked: an over-stuffed prompt (it can't tell which resume lines matter) plus an open invitation to fabricate (no instruction to stay grounded). Step 2: re-scope, don't re-ask louder. Narrow the task — feed only the 4 bullets relevant to this JD, not the whole resume — and add an explicit grounding constraint: "Use only facts I provide; if a JD requirement isn't in my history, flag the gap instead of inventing experience." Step 3: set the scenario: "You are a senior recruiter calibrating for an L5 backend role; rank my bullets by relevance to this JD, then rewrite the top 3." Step 4: run the loop with targeted deltas: first "the rankings look right, now rewrite #1 and #2 with metrics," then "tone is too junior — lead with scope and ownership, not tasks." Step 5: verify the fabrication is gone — confirm every claim traces to a fact you supplied. Pattern learned: when a prompt fails, separate the causes (overload vs. hallucination vs. wrong frame), fix each with the right tool — trim scope, add a grounding guardrail, set the scenario — then converge with small deltas. Loudness never fixes a structural prompt defect.

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