Prompt Techniques

A Beginner's Guide to Unleash the Power of GPT with Better Prompts

Zehan

2023.06.21

You may see these charts recently

And you may have heard that ChatGPT can be these roles

Writer(stories, jokes, reports)

Translator(natural, programming, and even visual languages)

Summarizer(articles, papers, codes)

Scheduler(meetings, event reminders)

Assistant(reply emails, reviews, or social media comments)

...(I have discovered a truly remarkable proof of this theorem which this margin is too small to contain)

LLM can even Design Chip!

8-bit shift register (by GPT-4)

Accumulator-based datapath (by GPT-4)

Processor Synthesis Layout

Use LLM to design hardware

LLM conversation flowchart

How they make all of these?

LLM

Prompts

Popular LLMs

New Bing (GPT4)

ChatGPT (3.5 & 4)

Google Bard

Anthropic Claude

Open LLMs ...

For most tasks, New Bing + ChatGPT-3.5 are sufficient.

New Bing

ChatGPT-3.5

Google Bard

Anthropic Claude

Open LLMs ...

New Bing

Backed by GPT4(While ChatGPT-4 has a limit of 25 messages per 3 hours)

Has access to Web

Recognize images(New!)

Even registration is not necessary(Sign-in would unlock the limits of messages per session from 5 to 30)

All you need is a valid IP(Most tricky part)

ChatGPT-3.5

Abundant third-party plugins and extensions

Active community and ecosystem

Competent for most tasks

Most importantly, Fast and Cheap

What is Prompt

Question (what you want to know)

Instruction (what you want to do)

Example (what you want to learn)

What is Prompt Engineering

The methodologies and techniques

to build and organize prompts

Tutorials

Learn PromptingContains everything you should learn

Prompt Engineering GuideFor researchers and practitioners. By DAIR.AI.

OpenAI CookbookExample codes for common tasks with OpenAI API

GPT best practicesStrategies and tactics for better results from GPTs. By OpenAI

FlowGPTFind & Use The Best Prompts

Prompts Cheat SheetGroup prompts by principles

Three Principles

More Informative

Less Ambiguous

Keep Iterating

More Informative

Detailed ...

Roles, Scenarios, Contexts, Instructions, Questions

Additional ...

Examples, Shots, Choices, Options, Paths

Less Ambiguous

Formatted ...

Inputs, Outputs

Limited ...

Length, Content, Style, Tools

Do ONE thing and do it WELL

Prioritize/Emphasize/Focus/Repeat

Keep Iterating

Chain/Tree of Thoughts

Break to Subtasks / Steps

Search before Reason, Plan before Action

Create Multiple Paths

Combine to Summary / Best Solution

Automation

Self Ask/Search

Self Check/Criticize

Popular GitHub ReposApplication Level

LangChainBuild LLM applications with workflow

LlamaIndexData framework for LLM applications

GPT4AllEcosystem of open-source on-edge LLMs

AutoGPTAttempt to make GPT fully autonomous

GPT-EngineerUse prompts to generate entire codebase

Let's make our hands dirty

Mission:
Build an arXiv paper abstract extractor

May the Prompt be with you