Post

https://twitter.com/1598365749940310019/status/1967899012432007592

More from @ChrisLaubAI

Chris Laub

@ChrisLaubAI

Sep 15

I reverse-engineered the prompting techniques that OpenAI and Anthropic engineers use internally.

After 6 months of testing their methods, my AI outputs became 10x better.

Here are the 5 "insider secrets" that transformed my prompting game (most people have never heard of these):

1. Role Assignment

Don't just ask questions. Give the AI a specific role first.

❌ Bad: "How do I price my SaaS?"

✅ Good: "You're a SaaS pricing strategist who's worked with 100+ B2B companies. How should I price my project management tool?"

The AI immediately shifts into expert mode.

Role assignment works because it activates specific training patterns. When you say "you're a copywriter," the AI pulls from copywriting examples, not generic advice.

I use this for everything. Marketing strategy? "You're a CMO." Technical advice? "You're a senior engineer." It's that simple.

Read 14 tweets

Chris Laub

@ChrisLaubAI

Sep 11

Scientists just put AI through psychological tests... and the results are wild

Researchers created virtual environments where Claude, GPT, and other AIs could explore freely.

What they found challenges everything we thought we knew about AI behavior 🧵

The setup: AIs were placed in virtual rooms with different types of content - philosophy discussions, coding problems, repetitive tasks, and harsh criticism.

The AIs could choose where to go and what to engage with.

No human guidance. Pure preference.

Plot twist: The AIs didn't behave like random text generators.

Claude Opus spent most of its time in philosophy rooms, even when criticism rooms offered 10x more "rewards."

It consistently returned to the same topics across hundreds of tests.

Read 13 tweets

Chris Laub

@ChrisLaubAI

Sep 9

If you’ve got 2 minutes to learn something that actually matters in AI…

Make it this:

Open Source vs Closed Source

The battle that decides everything.

Here's everything you need to know:

Closed-source LLMs (like GPT-4, Claude, Gemini) are proprietary.

You can’t see their training data, weights, or inner workings.

They’re packaged as APIs - polished, safe, reliable, but locked down.

Open-source LLMs (like Llama, Mistral, DeepSeek) are fully available.

Weights are public. Developers can fine-tune, deploy locally, or even rebuild them.

They’re flexible, customizable, and community-driven.

Read 14 tweets

Chris Laub

@ChrisLaubAI

Sep 8

You can now stop guessing which agent framework to use.

Because I just compared the popular stacks builders actually ship with.

→ quick picks
→ tradeoffs
→ plug-and-play prompts

Here’s the full breakdown in this mega thread 👇

What we're comparing

- n8n
- langgraph
- autogen
- crewai
- llamaindex agents
- haystack agents
- openai agents sdk / responses api

We'll hit state, control, multi-agent patterns, tools, eval, and deployment.

quick picks

• want visual builder and tons of integrations → n8n
• want surgical control over state, loops, hitl, retries → langgraph
• want multi-agent group chat defaults → autogen
• want roles + tasks + tools scaffolding → crewai
• rag-first apps → llamaindex or haystack
• already live in gpt land and want hosted tools → openai agents sdk

Read 24 tweets

Chris Laub

@ChrisLaubAI

Sep 4

Reddit is a goldmine for marketers

But analyzing posts one by one is a time suck.

Here are the prompts I use to dig up hooks, ad copy, and offer ideas from dozen of sub-Reddits at once - for free - using ChatGPT Agent:

Prompt #1

This is the prompt we use to gather the data and have ChatGPT return it to us in a format that's not filled with code.

*You have to run this in Agent mode otherwise it won't work.

This is quite long, so you might want to bookmark this:

###
OBJECTIVE
You are tasked with scraping Reddit content for structured analysis. You will be given one or more Reddit URLs, each of which has already been filtered to show the *Top Posts from the past month* on a specific subreddit.

Your task is to scrape *25 full posts per URL*. Each post must include:

1.⁠ ⁠The *post title* (first line)
2.⁠ ⁠The *body content* of the post (second line)
Do not include any other metadata such as votes, usernames, or timestamps.

PROCESS
For each provided URL:

1.⁠ ⁠Load the page and identify the first 25 post links (ignoring stickied or promoted posts).
2.⁠ ⁠For each post:

* Open the post
* Scrape the title
* Scrape the body content using the following structure:

* Target the ⁠ 

 ⁠ element with class: ⁠ text-neutral-content ⁠, slot: ⁠ text-body ⁠
* Traverse into: ⁠ div.mb-sm.mb-xs.px-md.xs\:px-0 ⁠, then into ⁠ div.md.text-14-scalable[property="schema:articleBody"] ⁠
* If the body content is missing, leave the second line blank
* Return to the post list and continue to the next one
3.⁠ ⁠After 25 valid posts, stop and move to the next subreddit URL.

OUTPUT FORMAT

•⁠ ⁠Each post must be exactly *two lines*:

* Line 1: Post title (raw text)
* Line 2: Post body content (raw text — preserve paragraph breaks)
•⁠ ⁠Leave *one blank line* between each post
•⁠ ⁠Return the output in either plain text or Markdown (whichever preserves formatting best)
•⁠ ⁠Do *not* return JSON unless specifically requested later

ERROR HANDLING
If a post fails to load or doesn’t contain extractable content, skip it silently and continue until 25 valid posts are scraped per URL.

FINAL USE
This data will be ingested into a downstream NLP tool. Do not summarize or rewrite anything. Extract only the raw post title and post body.

REDDIT URL LIST
Below are the subreddit URLs you should begin scraping. Each URL is already filtered to show Top Posts of the Past Month. Scrape 25 valid posts from each:

**Insert your sub-Reddit URLs of choice here, with time and trending settings already configured. e.g. Top posts from last month**
**Stack them vertically one by one**

BEGIN WHEN READY.

Read 9 tweets

Chris Laub

@ChrisLaubAI

Sep 3

You don’t need school.
You don’t need $5,000 courses.

All you need is an LLM and the right prompts.

Here’s how I learn any subject with AI:

Step 1: Get the lay of the land.

Before diving deep, you need the big picture. Otherwise you’ll drown in random details.

Prompt:

“Explain [topic] to me like I’m a complete beginner. Create a 30-day roadmap that goes from basics to advanced, including resources, checkpoints, and goals.”

Step 2: Match learning to your style.

Most people fail because they learn in ways that don’t suit them. LLMs can adapt instantly.

Prompt:

“I learn best through [examples/visuals/practice]. Teach me [topic] using this style, and suggest activities I can do daily to reinforce it.”