Ask AI for an answer, you get an answer. Send it on a treasure hunt, and it draws a map. The difference is the structure of your question. Today we walk slowly through three steps: territory, criteria, map.
When you hand off code or writing or data to AI, you know this feeling. "Fix this bug" — AI fixes it fast, but you're not sure it was actually the real problem. Maybe a bigger issue is hiding, and only the surface got patched. I had this anxiety for a long time. Then I changed how I asked, and the anxiety left. Today's story. Don't ask AI for an answer — send it on a treasure hunt. Let me go slowly.
This essay covers three things: the difference between answer-requests and hunt-requests, the 3-step structure of a hunt-request, and how to use it in practice.
Start with the two styles. They look the same but the results differ completely.
Answer requests sound like this. "Fix this bug." "Make this writing better." "Summarize this data." AI gives an answer. Fast and clear. But it only moves within the range you set. You said "fix the bug"; it fixes the bug. It doesn't touch the bigger problem nearby. Because you didn't ask for "exploration."
Hunt requests are different. "Find three risks I'd have missed in this code." Now AI doesn't fix. It explores. Scans the whole codebase — security, performance, maintainability — and draws a map of three findings. Which of those three is actually important — you decide.
The difference: in an answer request, AI is the protagonist. In a hunt request, you are. AI digs, you steer. That's the best form of collaboration.
Don't ask for an answer. Have it draw a map.
One-line summary.
An analogy. You're moving to a new city and asking a real estate agent for help.
Mode 1 — answer request. "Recommend one house in this budget." The agent recommends one. A fine house. Right before signing, you realize post-move: "Oh, there's an elementary school nearby — had I known, I'd have picked a different area." Regret. Because the agent didn't know your hidden criteria.
Mode 2 — hunt request. "Pull 5 candidates in my budget. Summarize pros and cons for each, specifically across 5 axes: transit, school district, noise, sunlight, HOA fees." Now the agent draws a map. Looking at it, you realize: "Ah, school district matters most to me." Then you choose.
Which approach finds the better house? Obviously #2. Not because the agent is more skilled — because you structured the question well.
Same with AI. AI doesn't need to get more skilled. Your question needs to get better.
So how do you structure a hunt request? Remember three steps.
Step 1 — set the territory. Where to dig. "Across this entire codebase," "within this file's functions," "in the middle section of this essay." Without a territory, AI digs in the wrong place. A too-wide map is useless.
Step 2 — set the discovery criteria. Define what counts as treasure. "Things that affect performance." "Things that will be hard to maintain in 10 years." "Things hurting user experience." Without criteria, AI looks for "generally good things" — which may not be what you need.
Step 3 — request the map. Ask for the relationships between findings. "Which is most urgent?" "How do they relate?" "What's the priority?" This meta-information is the map. Only with a map can you decide.
Here's a prompt I actually use.
"Across this 500-line file (territory), find 3 performance issues I probably missed (criteria). For each, give severity (high/med/low) and related potential issues (map)."
One run of that prompt produces, not one answer, but a map for decision-making. I look at the map and decide.
Numbers I tracked over a year using this method.
Hidden problem detection rate: answer-request average 2 → hunt-request average 7. 3.5× difference.
Rework frequency: 12 times/month → 3 times/month. 4× drop.
Decision speed: answer-only gave me no confidence — I double-checked repeatedly. With a map, I decide immediately. Average decision time down 40%.
All three metrics improved not because AI got better. Because my questions did.
One trap. "Hunt" doesn't mean fully open questions. "Find anything in this code" isn't a hunt — it's grazing. AI digs in useless places.
Good hunt questions are open but bounded. "Three performance issues" is a good example. Territory (performance), criteria (issue level), map (prioritized 3). Not "anything" — "three along this axis." Give it an axis and a number.
I have a checklist when writing prompts — territory / criteria / count — and make sure all three are present. Just that check lifts hunt-question quality dramatically.
The philosophy in one line. Humans direct. AI digs. That's the division.
Many people try to hand AI the direction too. "AI, figure it out." Result: average. AI learned from the world's average. Your unique direction, your current context, what matters to you — AI doesn't know. Only you do.
Conversely, if humans do all the digging too, nothing moves fast. A task that takes 3 hours to direct and 30 hours to dig would be 33 hours solo. With humans giving 3 hours of direction and AI doing the 30 hours of digging in 1 hour, it's 4 hours total.
This division is what real collaboration looks like. Each doing what they do best.
An advanced tip. You can combo hunt and answer requests. Order matters.
First request a hunt. Get the map. Look at the map and decide: "Let's fix #1 of the 3." Then send an answer request to execute. "From the map, fix problem #1 using this approach." AI now gives a clean answer.
The power of the 2-step combo is that discovery and execution are separated. Discovery phase = hunt. Execution phase = answer. Each phase runs in its best shape. I use this combo for major decisions. Combined time is under 30 minutes. Output quality is in a different league.
Don't ask AI for answers. Send it on a treasure hunt. Structure the question in 3 steps — territory, criteria, map — and AI becomes your exploration partner. Hidden problem detection 3× higher, rework 4× lower, decision speed 40% faster.
Tonight, try one thing. Instead of writing "do X," write "find 3 Y's in X that I probably missed." One line of prompt difference — entirely different output.
Don't ask for an answer. Ask for a map.
Remember — territory, criteria, map.