00:00:00 You've been there.
00:00:01 You open a new project, drop in a prompt, and the AI just works.
00:00:07 It scaffolds the auth, it sets up the routes, And it seems to know what you mean before you even finish typing.
00:00:13 And for the first 30 minutes, you're thinking, this changes everything.
00:00:18 Then the cracks start.
00:00:20 An hour in, It adds a function it already wrote.
00:00:24 You paste back an error and the fix breaks something else.
00:00:28 You ask it to touch one thing and it rewrites three files that you didn't ask about.
00:00:34 The context gets long, the responses get slower, and the answers get less precise.
00:00:40 Two hours in and you're not sure if the AI is helping you or just generating confident sounding code that you are too tired to check.
00:00:49 So you start a new chat and try to catch it up.
00:00:53 And now you just explaining things you already explained.
00:00:57 If that sounds familiar, here's what's actually happening.
00:01:01 Two root causes, and neither one improves with a better prompt.
00:01:05 Root cause number one.
00:01:07 The model gives you the average, but not your version.
00:01:12 A coding tool is trained to predict the next token across a massive amount of public code.
00:01:18 So when you ask it for something, it produces the most common way that that thing is written across everything it has read.
00:01:25 That's correct in general, But generic by default.
00:01:30 Your project is not general.
00:01:32 You have a specific API client, specific conventions, decisions you made last week.
00:01:38 The model has seen none of it.
00:01:41 Your code was never in its training and it's not in front of the model right now unless you put it there.
00:01:48 So you ask it to call your API.
00:01:51 and it writes a plain fetch with its own error handling because that's what calling an API looks like across the public internet.
00:01:59 It didn't ignore your client, it just had no idea your clients existed.
00:02:03 You think you're working with a senior engineer who knows your code base, but you are actually working a very capable contractor who just walked in the
00:02:12 door and everything they know about your project is what you've told them in this one conversation.
00:02:17 So yeah.
00:02:18 No real context.
00:02:20 That's the first root cause for all the issues.
00:02:24 The second root is that it plans toward what you said, not what meant.
00:02:29 You might expect me to say that the agent has no plan.
00:02:33 That used to be the story and it's now wrong.
00:02:36 A modern agent does plan.
00:02:39 It reads across your files, maps an approach, runs the tests, and even corrects itself when they fail.
00:02:45 That loop genuinely works.
00:02:48 But the real problem is what the plane aims at.
00:02:51 The agent plans towards the goal that you handed it, And it fills everything you left unsaid with the most likely default.
00:03:00 So you say, build the cart feature, and it makes 10 decisions that you never got to see.
00:03:06 State shape, API surface, error handling, it didn't ask.
00:03:11 It just picked specific things to do it with.
00:03:14 And because the code runs and the tests it wrote pass, nothing looks wrong.
00:03:19 Until those decisions pile up into a 300-line diff touching files that you weren't thinking about.
00:03:27 You can't review it, you can test it cleanly, and when something breaks, You CAN'T find it.
00:03:33 Because everything changed at once.
00:03:35 It didn't fail to plan.
00:03:37 it just planned toward an underspecified goal and filled the rest with reasonable guesses.
00:03:43 It's own guesses, not your decisions.
00:03:46 So the second root cause for most of the issues is that you didn't give it the right plan.
00:03:52 Now, when you put those two together, you get the cycle that most developers are stuck in.
00:03:57 Big prompt, big output, can't verify it, something breaks, even bigger prompt to fix it.
00:04:04 And now you really can verify.
00:04:07 It's just chaos.
00:04:08 And the instinct is always the same.
00:04:11 Keep prompting.
00:04:12 More detail, a different model, new chat.
00:04:16 But a longer, stricter message can't teach the model a codebase it never read.
00:04:22 It definitely can make a vague goal very specific by force.
00:04:26 You're trying to solve an engineering problem with a sentence.
00:04:30 The problem was never the prompt.
00:04:33 It's the system or rather the lack of one.
00:04:37 And that's what this course fixes.
00:04:39 Not with tricks or with prompting hacks, but with actual workflow.
00:04:43 One that engineers context on purpose, plans before it builds and keeps you in control instead of just hoping.
00:04:51 So from here on, you also have a diagnostic.
00:04:55 Whenever something goes wrong in the course or in your own work, ask one question.
00:05:00 Did the agent not have the context or did it not own the decisions that were yours to make?
00:05:07 Almost always, it's one of those two.
00:05:10 So in the next lesson, we'll talk about the mental shift that makes all of this possible because the technique only lands if the mindset shifts first.
