🎯 Core Goals
- Consolidate the practical takeaways from the basics section.
- Walk away with things you can apply to your own LLM usage today.
LLMs predict the next word based on patterns — they don’t think or understand. That single fact explains most of their strengths and weaknesses. The sections below are your practical cheat sheet.
🧠 8 Things to Remember
LLMs predict, they don’t think Treat every response as a first draft, not a final answer. Good at generating options, bad at knowing what’s true.
They’re great at finding similar words and concepts Want alternatives, related ideas, or rephrasing? This is where LLMs shine — they excel at “things like X.”
Context matters more than you’d expect The LLM picks up on everything in your conversation. Ambiguous requests get ambiguous answers; specific requests get specific answers. You don’t need to spell everything out — but the more context you give, the better the output.
They can’t reliably count, spell, or do math LLMs see text in chunks (tokens), not individual letters. Don’t ask them to count r’s in “strawberry” or do precise arithmetic. Use a calculator or spreadsheet for that.
They can track relationships across long text LLMs are good at connecting ideas across a document — finding themes, resolving pronouns, summarizing. Give them long text and ask them to find patterns.
They have no memory between conversations Every new chat starts from zero. Repeat important context, and don’t expect the LLM to “remember” what you discussed last week in a different thread.
Hidden instructions shape behavior Custom GPTs, Claude Projects, and app integrations work by pre-loading instructions. The LLM’s personality and expertise come from context, not from “knowing” things.
Hallucination cuts both ways The same mechanism that lets an LLM write a creative story also lets it invent a fake legal case. For creative work, that’s a feature. For factual claims, that’s a danger. Always verify facts that matter.
🚨 8 Red Flags
“The LLM understands what I’m asking.” — It produces convincing-sounding responses, but it has no actual comprehension. The chat interface creates an illusion of understanding.
“Just ask it to count the letters.” — It sees chunks, not letters. Never rely on an LLM for precise counting, arithmetic, or character-level tasks.
“I told it to forget the wrong thing.” — Once false information enters the conversation, it’s re-read every turn. Start a fresh conversation instead of trying to correct it in-place.
“More context = better results.” — Long conversations degrade in quality (context rot). Start fresh for important work, especially after many back-and-forth turns.
“This chat thread is my permanent workspace.” — Long threads accumulate errors through context poisoning and truncation. Use separate threads for separate tasks.
“LLMs are too unreliable to use seriously.” — With proper safeguards (verification, human review, not trusting blindly), LLMs power real production systems daily. The risk is using them without safeguards.
“Enterprise plans keep my data perfectly safe.” — Data still travels over the internet and is processed on external infrastructure. Treat LLMs like a trusted but external contractor.
“I’ll just use one long conversation for everything.” — Context fills up, old messages get truncated or summarized, and quality drops. Shorter, focused conversations produce better results.
💡 Pro Tip
Use the LLM to improve your own prompting. Ask it to create a cheatsheet, checklist, or set of DOs and DON’Ts tailored to your specific use case. Example: “I use LLMs to draft client emails. Create a checklist of what I should always include in my prompt, and common mistakes to avoid.”
The LLM has seen millions of prompts and their outcomes. Let it help you write better ones — then iterate on that cheatsheet as you learn what works for you.
📐 Quick Reference
Trust Spectrum — Ask yourself: “What’s the worst thing if this is wrong?” → Summaries/brainstorms: use freely → Email drafts/outlines: quick scan → Facts/code: cross-check → Medical/legal/financial: use a human expert. (Learn more)
The 6W Prompt Framework — Better prompts answer more Ws: Who is the audience? What do you want? When/Where is the context? Why does this matter? How should it be formatted? (Learn more)
Few-Shot > Zero-Shot — Showing 2-3 examples in your prompt dramatically improves output quality compared to just describing what you want. (Learn more)
Chain-of-Thought — For complex reasoning, ask the LLM to think step by step. It produces more accurate results on multi-part problems. (Learn more)
Privacy Heuristic — “Would I put this in an email to someone I don’t fully trust?” If no, don’t paste it into an LLM. (Learn more)
Context Window — Every turn resends the entire conversation. Turn 10 costs more than turn 1. Start fresh for important tasks. (Learn more)
🗺️ Where to Go Deeper
| Topic | Explore |
|---|---|
| Core intuition — autocomplete at scale | Keyboard Autocomplete vs LLM Completion |
| How attention works | ELI5 Attention |
| Tokenization — why they can’t count | What is Tokenization? |
| How chat actually works behind the scenes | The Sandwich |
| Hallucination | Hallucination |
| Context rot and degradation | Context Rot |
| Jailbreaking and security | Jailbreaking & Security |
| Trust spectrum — when to verify | Know When to Trust It |
| Chain-of-thought reasoning | Chain-of-Thought & Reasoning Tokens |
| Privacy — what never to paste | Privacy |