Tips & Lessons
Everything we wish someone had told us before we started — and everything we learned the hard way after.
Workshop Tips
Stop asking it questions. Start giving it jobs. The shift from “What is X?” to “Build me X” is where the magic happens.
Vague prompts get vague results. Tell AI exactly what you want: the format, the tone, the length, the audience. Constraints make output better.
Give AI an example of what good looks like. One example is worth a hundred words of instruction.
A single prompt is a question. A system prompt + tools + goals is an agent. Build systems, not one-off interactions.
Your first version will be rough. That’s the point. Get something working, then iterate. Perfection is the enemy of learning.
The more relevant context you give AI, the better it performs. Background, constraints, examples, audience — feed it all.
Don’t just accept the first output. Ask AI to review, critique, and improve what it just made. Self-correction is a superpower.
If you do the same thing more than twice, make an agent do it. Scheduling, formatting, summarizing — these are agent jobs now.
AI explains its reasoning. Read it. You’ll learn more from how it thinks than from what it produces.
Don’t ask one agent to do 10 things. Give each agent a clear, focused role. Then orchestrate them like a team.
AI handles small, clear tasks better than big, vague ones. Chain reasoning: step 1 feeds step 2 feeds step 3.
Your job is to describe what you want and review what you get. AI writes the code. You make the decisions.
What 4 Years of AI Taught Me
When it doesn’t have real data, it invents convincing answers. Always verify before you trust.
Give it a reference and it adds features, changes layouts, “fixes” things. Be explicit: “match exactly, do not improve.”
First-pass comparisons always skip differences. Budget at least 2 rounds of review for anything you ship.
Feeding AI more information doesn’t help unless you explain what it means and how to use it. Context engineering > prompt engineering.
A typo in a config, a missing import, one wrong character — tiny errors cascade into system-wide breakdowns. Sweat the details.
When something breaks, the instinct is to restart and try again. Resist it. Find the root cause first or you’ll trigger the same failure on loop.
If you can’t show it working, it’s not done. Proof of working is the only definition of complete.
When you find one problem, look for more. Multiple issues hide behind each other — fixing just the first one wastes time.
Shorter, focused context windows let you direct AI with more precision. More isn’t better — relevant is better.
Create a process to capture what went wrong and what you learned. If you don’t write it down, you’ll repeat it.
A defined identity helps AI understand how to execute within constraints. Role, tone, boundaries — these shape better output.
Build learning loops into your agents. If they’re just repeating the same thing, they’re scheduled scripts with extra steps.
