Stop asking.
Start engineering.

Prompt engineering — the real skill

Prompt engineering is clear communication with a pattern-matching machine. There are no magic phrases. The better you describe what you want, the better the output, and a vague prompt like "write about dogs" gets a vague encyclopedia entry while a specific one gets a checklist you can actually use.

The CRIC Framework: every good prompt has four components. You don't always need all four, but the more you include, the better the result. Hover a card:

Example: without vs with CRIC.

Advanced techniques

Few-shot prompting: show the model examples of what you want. The model learns the pattern from your examples and applies it, which is often more effective than describing the format in words.

# Convert sentences to JSON:
Input: 'John, 25, engineer' → Output: {'name': 'John', 'age': 25, 'job': 'engineer'}
Input: 'Sarah, 30, designer' → Output: {'name': 'Sarah', 'age': 30, 'job': 'designer'}
Now convert: 'Mike, 28, developer'

Chain of thought: ask the model to think step by step. Adding "think step by step" dramatically improves accuracy on reasoning tasks, because the model shows its work and that visibility cuts down on errors.

"Think step by step: If a train leaves at 3 PM
traveling 60 mph, and another leaves at 4 PM traveling
80 mph in the same direction, when does the second
train catch the first?"

Negative instructions: tell the model what NOT to do. Surprisingly effective at shaping output.

"Explain machine learning.
Do NOT use jargon without defining it.
Do NOT use analogies to the human brain.
Do NOT start with 'Machine learning is a subset of AI.'"

Tool calling — from talking to doing

By default, an LLM can only generate text. It can't open the internet, run code, or read files. You ask it for Bitcoin's current price and you get an answer from training data that's months out of date. Tool calling fixes that: it lets the model ask an external system to check something, then fold the result into its answer.

Example flow:

The model never touched the internet. It requested an action, the system performed it, and the model used the result.

Common tools:

• Web search (access current information)
• Code execution (run and test code)
• File operations (read, write, edit files)
• Database queries (access structured data)
• API calls (interact with external services)
• Calculations (precise math, not token prediction)

How to leverage this:

• When AI has access to tools, ask it to USE them rather than rely on training data
• "Search the web for current BTC price" beats "What's the BTC price?"
• "Run this code and show me the output" beats "Will this code work?"
• Tools make AI reliable because they provide real data, not predictions

Memory & context management

The biggest limitation you'll hit with AI is memory. You can spend 30 minutes teaching it your project, your preferences, your codebase, and the next day it answers "Who are you? What project?" Every conversation starts from zero. That's infuriating, and it's not a bug that will get fixed. It's how LLMs work by design.

Strategies to manage this:

• Save context externally. Don't rely on the AI to remember. Save important facts in files, notes, or documentation. Re-inject them when starting a new session.
• Use system prompts for persistent context. If the platform allows it (ChatGPT Custom Instructions, Claude Projects, etc.), put your core context in the system prompt. It gets injected every conversation automatically.
• Summarize before switching. At the end of a long session, ask the AI: "Summarize everything we discussed and decided in a format I can paste into a new session."
• Structure your conversations. One topic per conversation. Don't mix unrelated tasks. This keeps the context window focused on what matters.
• Build a "context file." Maintain a document with your key information (project details, preferences, decisions made). Paste it at the start of new sessions.

RAG — give AI your own data

RAG (Retrieval Augmented Generation) solves a fundamental problem: AI only knows what it was trained on. It doesn't know your files, your data, or your specific situation.

The concept: instead of asking AI to answer from memory alone.

Why this matters:

• AI becomes useful for YOUR specific data, not just general knowledge
• No need to fine-tune (expensive). Just retrieve relevant docs.
• Data stays under your control (it's injected at query time, not trained on)
• You can update your data anytime without retraining

Multi-model strategy

Don't be loyal to one model. Be strategic. Different models have different strengths. Using the right model for the right task saves money, improves quality, and speeds up your workflow.

Practical model selection:

Cost optimization:

• Use cheap models for bulk/simple tasks
• Reserve expensive models for complex reasoning
• Start with a cheap model, escalate to expensive if needed
• Batch similar tasks to one model session

What can go wrong at this level

Power-user habits have their own failure modes. Here are the four most common ones.

What you should know after Level 2

You now have the skills of a power user. Tap each as it clicks: