The biggest bug in any agent is "I assumed."
Here's what almost every beginner does when they design their first agent: they sit down at the computer, imagine what their grandma might want, and start coding. They build the agent they think grandma needs. Three weeks later, they give it to her, and she never uses it. Why? Because they designed for a grandma who only exists in their head.
"I bet grandma would love an AI pill reminder."
You imagine the problem. You design the solution. You never actually watch her take her pills, never ask her about her morning, never learn the real thing.
"I watched grandma's morning. Her actual problem was remembering names."
You watched her for a day. You asked what bugged her. The answer surprised you — her pills were fine, but she couldn't remember her new doctor's name and it was making her anxious.
The biggest advantage you have as a young builder is that you know real people — your family, your friends, your teachers. You know them better than any Silicon Valley founder knows their "user." But only if you actually pay attention. Designing for a person you've only imagined is how adults waste millions of dollars. Don't make the same mistake with your weekend.
Pick a real person you can actually watch.
In Module 01 you picked a kind of person to design an agent for. Now make it concrete — one specific human you have access to, someone you can observe, ask questions, and give the agent to afterward. If that's still the same person, great. If you want to switch, switch now.
Pick the person you can actually talk to in the next week. Not a hypothetical. If you pick "myself, honestly" — that's a legitimate choice. Designing an agent for yourself is valuable if you're brave enough to observe your own weak spots instead of imagining perfect ones.
Five questions. Ask them in this order.
These five questions are designed to bypass the thing people say they want and get to the thing they actually struggle with. Notice: not a single question asks "what feature would you like?" That question is worthless. Real design information lives in these five instead.
"Walk me through your last Tuesday, from waking up to bedtime."
Why Specific days are concrete. "Tuesday" forces them to remember actual events instead of summarizing. You'll hear what their real day looks like — not the version they present to strangers.
"What was the most annoying thing that happened this week?"
Why Pain is the best signal. The thing that annoyed them is almost always the thing worth fixing. Don't ask "what problems do you have?" — people will lie politely. Ask about annoyance; they tell the truth.
"When do you ask someone for help, and what do you ask for?"
Why The moments they ask other humans for help are exactly the moments an agent could help instead. Every "can you read this for me?" or "can you remind me?" is an agent-shaped hole.
"What's something you keep forgetting, even though you want to remember it?"
Why Memory gaps are the simplest thing an agent can fix. And they tell you what matters: the thing they want to remember is a thing they care about. Care + forgetting = perfect agent territory.
"If you had one extra hour today, what would you actually spend it on?"
Why The answer tells you what they value. A good agent doesn't free up time for random stuff — it frees up time for the thing the person already wishes they were doing. That's the agent's real purpose.
Notice what's missing: "What features would you like?" "Would you use an AI assistant?" "Do you want me to build something for you?" Those questions get polite-but-useless answers. People don't know what they'd use until it exists. But they always know what annoyed them this week.
Here's Maya's actual interview notebook.
Maya is 12. Her grandma is 74 and lives alone two towns over. Maya used the five questions from the last step to interview her over the phone. Here's what she wrote down — and what she noticed. The yellow notes in her handwriting are her own observations between the answers.
Interview with Grandma · three weeks ago
Saturday afternoon phone call, about 40 minutes
Walk me through your last Tuesday, from waking up to bedtime.
"I woke up at 6:30. Made tea. Fed the cat. Watched the news. Around 10 I tried to call the pharmacy but I forgot why I was calling by the time they picked up so I hung up. Had lunch. Read my book. Then I realized I forgot to take my blood pressure medicine in the morning so I took it at 2pm which is not when you're supposed to. Then Mrs. Chen came over and we played cards."
What was the most annoying thing that happened this week?
"Oh, I had to call the insurance company about a bill. They put me on hold, and when the person came back she said a bunch of words I didn't understand — deductible and co-pay and something else — and I said yes to whatever she said just to get off the phone. Now I don't know if I agreed to something I shouldn't have."
When do you ask someone for help, and what do you ask for?
"I ask your mom to read things for me when the letters are too small. I ask my friend Mrs. Lin to explain the doctor's paperwork. I ask you, sometimes, to figure out my phone when the little icon goes away. That's about it. I don't like asking for help, honestly."
What's something you keep forgetting, even though you want to remember it?
"The names of my new neighbors. There's a young couple that moved in next door six months ago and I've forgotten their names three times and I'm too embarrassed to ask again. I made up a nickname for them just to talk about them with your mom."
If you had one extra hour today, what would you actually spend it on?
"Probably the garden. The tomatoes need staking and I've been meaning to get to them for two weeks. Or I'd read. I'd really rather read than do most things, but there's always something that needs doing first."
Look at what Maya found. She didn't ask "do you want an agent?" — she asked about a Tuesday and found out grandma forgets things three times a day. She didn't ask "what feature?" — she asked about the annoying thing and learned grandma is saying yes to insurance calls she doesn't understand. The agent Maya should build is nothing like what she would have built by guessing.
Turn observations into an agent spec.
Fill in the form below. Use real observations (not made-up ones) — if you haven't interviewed your person yet, at least think of specific moments you've actually seen. The design spec builds itself on the right as you go.
Untitled Agent
↳ for someone specific
(what did you actually see?)
(one sentence)
(the tiny thing it actually does)
- (when should it stop and ask?)
Notice what's missing from this spec: every feature you could add. The good design has one job, one move, one pause. If your spec has more than that, it's not an agent yet — it's an app wearing an agent costume.
Feature-driven or observation-driven?
The hardest taste choice in agent design. Both versions below sound reasonable. Only one was built by someone who actually watched a real person.
Round 1. Two "helpful assistant for elderly people" agents. Which was actually designed for a real person?
Round 2. Your younger sibling (age 8) struggles with homework. Which agent design came from watching them?
Every time you design an agent, ask yourself: "Can I point to the specific moment, in a specific real person's life, that this agent was built to solve?" If the answer is yes — like Maya pointing to grandma forgetting why she was calling the pharmacy — the design is real. If the answer is vague — "elderly users" or "kids who struggle with homework" — you're imagining, not designing.
You just learned empathy-first design.
Which means you're now better at designing agents than most of the people shipping AI products for money. They forgot to watch.
What you just learned
- The biggest bug in any agent is "I assumed." Assumption-driven design builds agents nobody uses.
- Your advantage is you know real people. Use it.
- Five interview questions get you to the real problem — and none of them ask "what feature would you like?"
- Pain is the best signal. The most annoying thing this week is almost always the thing worth fixing.
- A good agent design has one job, one move, one pause. More than that means you're drifting into app territory.
- Ask yourself: "Can I point to the specific moment this agent was built to solve?" If yes, it's real. If no, you're imagining.
In Module 03, you'll learn why agents that work in silence are scary — and how to make agents that show their work so the person using them can actually trust them. Transparency as the foundation of trust.
★ Before you call it done
Three questions. Same three. Every time.
These are the same three questions for every module in Kindling. They are how you check whether AI did the part it should and you did the part only you could. Tap each one to mark it true.
★ ★ ★