A hands-on lesson · The Literal Machine
Tell the robot how to make a PB&J
Write instructions for the robot in plain English, press Run, and watch it do exactly what you said — not what you meant. Real AI won't be this literal: it quietly fills the gaps you leave, and sometimes guesses wrong. This robot just makes those gaps impossible to miss.
On the counter
- A loaf of sliced bread, inside a closed, tied plastic bag
- A jar of peanut butter, lid on
- A jar of jelly, lid on
- A clean butter knife
- A plate
Your goal: get the robot to make one peanut butter & jelly sandwich — it needs both. The robot does only and exactly what you tell it, in order.
Robot execution log
Stuck? A few ways to tighten your instructions
- Number your steps. An ordered list forces you to decompose the task — and makes a missing step obvious, to you and to the machine.
- Define your terms first. Name things up front — "call the two slices A and B" — then refer back to them. Naming kills ambiguity and lets you reuse a step instead of re-describing it.
- Meet every precondition. The robot can't take bread from a bag it never opened. If a step depends on an earlier one, put the prerequisite first.
- Don't rely on it filling the gaps. A person brings a lifetime of context and quietly fills in what you left out. AI fills gaps too — but probabilistically, so an unstated detail is a gamble, not a guarantee. It might get it right; you can't count on it.
- Describe what "done" looks like. Spell out the finished result, not just the steps. A clear definition of success lets it — and you — tell a correct outcome from a plausible-looking wrong one.
Edit your instructions above and run it again.
I'm ready — reveal the lesson
If the robot just embarrassed you, good — that's the point. You didn't fail; your specification did. A real AI would likely have hidden that same gap by quietly guessing — which is worse, because you'd never see it coming. Here's the gap you just felt, and the five skills that close it:
Specification
"Put peanut butter on the bread" felt complete in your head and was wildly ambiguous on the counter. Saying precisely what good looks like is the highest-leverage skill with any AI.
Decomposition
A sandwich is really a handful of small, ordered, checkable actions — open the bag, then take the bread. Things break exactly where a step is skipped.
Verification
"Run the robot" is your acceptance test — output versus intent, gap exposed in seconds. Skip it and AI output just looks done and breaks later; build it in and the same output becomes something you can trust enough to ship.
Extraction & reuse
Notice "spread {filling} on {slice}" is the same move twice. Name it once and reuse it instead of rewriting it. This is the difference between a pile of one-offs that collapses as it grows and software that scales — abstraction is how senior engineers keep it healthy.
Taste & judgment
Several runs "technically completed" and still produced a buttered countertop. Knowing the difference between done and good is the final skill — and now the job.
This is where I'd pause the demo and teach the tactical version: how to write a spec, decompose a task, wrap AI output in checks, and factor out the reusable parts. Verification and reuse, especially, are what separate vibe-coded work that dazzles then breaks from agentically-built work that lasts. Swap the sandwich for a launch brief, a data pull, or a hiring rubric and the lesson is identical — which is why it works for anyone doing AI-native work, not just people who write code.
Now watch it happen for real
A dad asks his kids to write instructions for making a sandwich, then follows them with glorious literalness. Same lesson, no robot required.