Published on

Prompt Engineering Guides: Google vs. OpenAI – What You Actually Need to Know

Authors
  • avatar
    Name
    Ajinkya Kunjir
    Twitter

Large‑language‑model documentation can feel like a fire‑hose. Below is a distilled guide for everyday Gen‑AI users who just want to write better prompts fast—without wading through 60‑page PDFs.

1 · Why These Guides Exist

GuidePrimary Goal
Google Prompt Engineering GuideTeach Gemini / Vertex AI users how to pair model settings (temperature, top‑K, top‑P) with advanced prompting patterns (CoT, ReAct, ToT) for production apps.
OpenAI Cookbook › GPT‑4 PromptingProvide a minimal playbook for getting reliable answers out of GPT‑4—emphasising clarity, structure, and iterative testing.

2 · Shared Golden Rules

  1. Say exactly what you want. Specific instructions > vague requests.
  2. Show a pattern. One‑shot/ few‑shot examples boost accuracy.
  3. Iterate. Tweak wording and model settings; re‑run; compare.
  4. Structure outputs. Ask for JSON/CSV when you need machine‑readable data.
  5. Control randomness. Lower temperature ≈ factual & repeatable; higher ≈ creative.

3 · Quick Cheat‑Sheet: What You Need to Know

A) Google Gemini / Vertex AI Users

Do thisWhy
Pick the right model first (gemini‑pro text vs gemini‑pro‑vision multimodal).Different capabilities + cost.
Set sensible defaults – temperature 0.2, topP 0.95, topK 30, maxTokens only as big as you need.Balances coherence & creativity; saves money.
System / Role / Context promptsGoogle treats the entire prompt string as one chunk → you embed role or context inside the text.
Use #Context: blocks (or a JSON schema snippet) to anchor the answer.Reduces hallucinations; enforces format.
Chain‑of‑Thought for reasoning – add "Let's think step by step." after the question (temperature 0).Forces explicit reasoning → better answers for maths, logic.
Advanced patterns (ReAct, Step‑back, Tree‑of‑Thought).When basic prompting stalls, switch to these.

Gemini Example – JSON extraction

System: Extract movie sentiment as JSON.
Review: "A breathtaking masterpiece with minor pacing issues."
Schema: {"sentiment": "POSITIVE|NEUTRAL|NEGATIVE", "notes": string}
JSON Response:
{
"sentiment": "POSITIVE",
"notes": "Visually stunning but minor pacing issues"
}

Quick Decoder

SettingWhat it controlsRecommended default
TemperatureGoverns randomness in token selection.
0 → deterministic (always choose the top‑probability token).
1 → highly creative / unpredictable.		0.2 keeps answers mostly factual yet allows mild variety.
topP (nucleus sampling)Keeps only the smallest probability mass ≤ p. topP 1 uses the full vocabulary; lower values remove fringe tokens.0.95 removes extreme outliers while preserving 95 % of useful vocabulary.
topKConsiders only the K most‑probable tokens before applying temperature / topP.30 = “use the 30 best candidates, ignore the rest” — coherent yet flexible.

Mental Model

  1. topK caps the maximum candidate pool.
  2. topP trims ultra‑low‑probability outliers.
  3. Temperature spices the final pick: lower for facts, higher for creativity.

Rule‑of‑thumb presets • Factual queries: temperature 0–0.3, topP 0.9–0.95, topK 20–40 • Creative writing: temperature 0.7–0.9, topP 0.98, topK 40–100

Use these knobs together to balance coherence vs originality without blowing your token budget.

B) OpenAI ChatGPT / GPT‑4 Users

Do thisWhy
Use the 3‑message pattern → system, user, optional assistant examples.Mirrors how the API fine‑tunes behaviour.
Give the model a role first – system: "You are a precise financial analyst."Sets tone + knowledge scope.
Be explicit about style / length – e.g. “Answer in two bullet points.”GPT follows formatting instructions well.
Temperature guide: 0–0.3 deterministic · 0.4–0.7 balanced · >0.7 creative.Quick mental model.
JSON mode (response_format={"type":"json_object"}) or function‑calling when you need structured data.Guarantees valid JSON; no post‑parsing headaches.
Few‑shot > Many‑shot – keep within context window (~128k GPT‑4o) but don’t bloat prompt.Saves tokens; keeps core examples visible.

GPT‑4 Example – Few‑shot classification

{
  "system": "You are a support triage bot. Answer with 'LOW', 'MEDIUM' or 'HIGH'.",
  "user": "Classify: 'My screen occasionally flickers when waking from sleep.'",
  "assistant": "LOW",
  "user": "Classify: 'Laptop battery swollen and leaking.'"
}

Expected answer: HIGH

QA Agent Output Screenshot

4 · Key Differences at a Glance

AspectGoogle Gemini GuideOpenAI Cookbook
Depth of theoryHeavy: sampling maths, multi‑step reasoning patterns.Light: theory trimmed; focus on recipes.
Tooling focusLangChain, Vertex AI APIs, external search & function calls.Native chat completions + JSON / function calling.
Prompt scopeOne big string (you manage roles inside).Structured message array (system/user/assistant).
Advanced patterns coveredCoT, ToT, Self‑consistency, ReAct, APE.Mainly CoT + few‑shot + JSON format.

5 · TL;DR for Busy People

  1. State the task, the format, and any examples—up‑front.
  2. Clamp temperature ≤ 0.3 for factual answers; raise it for creative tasks.
  3. Ask for JSON if a machine will consume the output.
  4. Iterate: tweak words or model knobs, rerun, compare.
  5. Use Chain‑of‑Thought (Let's think step by step) when the first answer feels wrong.

Stick to those five rules and you’ll harvest 80 % of the value from both GPT‑4 and Gemini—no 70‑page whitepapers required.

Further reading: Google Prompt Engineering Guide · OpenAI Cookbook

Discuss on Twitter •