YT YouTube2Text

YouTube2Text → Transcripts → Is Learning to Build Software Still Worth It with AI? (My Honest Take)

Is Learning to Build Software Still Worth It with AI? (My Honest Take) — Transcript

by Cole Medin · 3,455 words · 488 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00reverse engineer anything. This is the
0:02latest viral GitHub repository and it is
0:05genuinely fascinating. It's a really
0:06good example of what we're able to
0:08accomplish with agents when we have
0:10proper harness engineering. But I'm
0:12seeing this all over the place right
0:14now. People are completely blowing this
0:16out of proportion, saying if agents are
0:18able to do something this difficult, it
0:20means that coding and engineering is
0:21entirely solved and we don't even have
0:23to apply our minds to it anymore. Or at
0:24least that's coming very soon. But there
0:27is a very specific reason why agents are
0:29so good at this task in particular. And
0:32so we'll talk about this today and how
0:34that kind of plays into the larger
0:35conversation I want to have here, which
0:37is what is even worth learning as AI
0:40starts to take over more and more
0:42complicated tasks like this. And I get
0:44it. With all the hype around AI and the
0:46doom and gloom in the engineering space,
0:48like this tweet that went viral last
0:49month, it might genuinely feel like to
0:52you that it's less important for us to
0:54invest in the skills of building as AI
0:57gets more capable and we're just simply
0:59becoming meat proxies where agents are
1:01getting good enough where we're just
1:02going to delegate all of the critical
1:03thinking to them. But this is not the
1:06way to go. I firmly believe that there
1:08are some skills that actually become
1:10more important for us to invest time
1:12into as agents get more capable because
1:14not everything is reverse engineering.
1:16There are certain tasks that agents
1:18completely knock out of the park, but
1:19that doesn't mean we can just
1:20extrapolated to it solving everything.
1:23So, keep this repository in mind cuz
1:24we'll come back to it in a second. But
1:26first, I want to hit the nail on the
1:28head here. The big question that I've
1:29been teeing up is, is anything still
1:32worth learning with AI? And spoiler,
1:34yes, very much. There are skills that we
1:36want to be investing in. And the biggest
1:38reason is large language models can't
1:40read your mind. No matter how powerful
1:42they are, you still have to have the
1:44larger harness around the agent to
1:46properly communicate your intent. That
1:49is the shift that we're seeing in the
1:50engineering space. So agents do
1:52genuinely write all the code. I haven't
1:54written a single line of code myself in
1:56over a year. So I delegate the coding,
1:59but I'm very much a part of the planning
2:01up front and the validation at the end.
2:03Even if I have a very elaborate harness
2:05with a super good large language model
2:07and I trust it to write very clean code,
2:10I still need to make sure that I'm
2:12getting what I actually want. So if the
2:14agent makes a mistake, it's actually
2:16more my fault that I didn't properly
2:18communicate my intent. And so I need to
2:20have the harness up front where it's
2:22grilling me with questions, getting on
2:24the same page, validating my intent at
2:26the end as well. This is never going to
2:28go away. And the skill of communicating
2:31our intent and building the harness
2:32around the agent to allow us to do so is
2:35only ever going to get more important as
2:37we task our agents with accomplishing
2:39harder things. But the problem is
2:41exactly because agents are so good at
2:44the middle part. Everyone is tempted to
2:46become a meat proxy. I mean, I've dealt
2:48with this temptation myself. Wherever I
2:50might receive a request like a Jira
2:51ticket or something from a teammate, I'm
2:53just going to dump that right into my
2:55coding agent without any planning, any
2:56critical thinking of my own. And then
2:58whatever I get back out from the agent,
3:00I just spew that back into Slack or as a
3:02pull request, whatever that might be.
3:04And so in this situation, the person
3:05giving the original request may as well
3:07have just gone to the coding agent
3:08themselves because you've added no value
3:11in going through this process. But the
3:13code is still on you, right? Like as an
3:15engineer, you still have to take
3:16responsibility for the output even if
3:18you didn't apply any of your own
3:20critical thinking. So you better make
3:21sure that you've communicated your
3:23intent properly because if you really
3:25are just a meat proxy, I can guarantee
3:27for most tasks, the agent is not going
3:30to do it as you actually need because it
3:32doesn't understand the nuances of how
3:34you want to fix the bug or the feature
3:35you want to build, whatever that might
3:37be. The point is, you absolutely do not
3:39want to be that person where if you get
3:40asked a question in Slack, you just dump
3:42it to your coding agent and then you
3:44reply, "Claude said," and then here's a
3:46giant AI response. You make the other
3:47person review. This is the ultimate form
3:50of being a meat proxy. Okay, cool. I get
3:52it. But if it's so important for us to
3:54communicate our intent properly and
3:57build that larger harness to make that
3:59possible, like I'll talk about more in a
4:00little bit, how can we have examples
4:02like this? I mean, with reverse engineer
4:04anything, if you take a look at the
4:05readme here with their quick start, it
4:07is a very simple request to reverse
4:10engineer any kind of application. And
4:12with this MCP, I mean, the reason this
4:14GitHub repo blew up is because it is so
4:16incredibly easy to build anything, to
4:18reverse engineer anything with it. But
4:21like I said at the start of the video,
4:22there is a very specific reason why this
4:25works. Think about it. The reason this
4:27works so incredibly well is because the
4:29application that you are reverse
4:30engineering literally is the intent. You
4:33need nothing else to communicate
4:35properly what the final product should
4:37be because the final product is already
4:39there. And so the agent is constantly
4:41able to validate its own work 100%
4:43reliably because it's able to see what
4:45it's building right now. Does that align
4:47with the features or the behavior of the
4:49application that it's reverse
4:51engineering? That is why this process
4:53works so incredibly well. And of course
4:55also added in is the fact that there's a
4:57very well-designed harness here. So of
4:59course that is important as well. But
5:00even with a harness generally you have
5:02to do a lot more of that intent
5:04engineering. But here you don't need to
5:06because you already have it. That's why
5:08I think people are so wrong when they
5:10say that this repository proves that
5:12soon agents will be able to build
5:13literally anything. It's not a fair
5:15example because it has such a good
5:17reference. Now, a lot of the things that
5:19we build, there are good references we
5:21can give the agent and you do want to do
5:23that as much as you possibly can. But
5:26for most of what you build, it's simply
5:28going to start with an idea in your head
5:30or someone who's assigning you the task.
5:32And in that case, it is still up to you
5:34to communicate the intent with your
5:36agent. if you just give it a vague idea,
5:38if you're just the meat proxy, you're
5:40going to produce good code because LLMs
5:42write good code, but it misses the point
5:44of what you actually want to build. And
5:46so most of the time when you aren't, you
5:47know, reverse engineering an app and you
5:49don't have that clear example, you
5:51essentially have to make the reference,
5:53right? Like this is where as builders,
5:55we still very much have to apply our own
5:57critical thinking because we have the
5:59agent gorillas with questions. We're
6:00answering these and actually putting a
6:02lot of thought into them as we should.
6:04We're building up that reference with
6:05documentation like a PRDA, a plan
6:08document, whatever that might be. The
6:10point is we build the reference for the
6:11agent to go off of so it actually
6:13understands what we want in the end. And
6:15so with all that being said, the most
6:17important skill to invest your time into
6:19right now as AI gets more and more
6:21capable is building the harnesses around
6:23communicating your intent. So giving it
6:26references and allowing it to discover
6:28references when possible. Otherwise,
6:30going through the process of essentially
6:32building your own. A lot of people call
6:34this intent engineering. Whatever it is,
6:36the point is you have to be able to
6:37recognize, apply your own critical
6:39thinking to know when you can give it
6:41examples and when you have to
6:42essentially craft a replacement. Right?
6:45I've always said for years, examples are
6:47the most powerful thing for large
6:49language models. And it doesn't matter
6:51how powerful the LLM is, it still needs
6:53that to really know what you actually
6:55want. The sponsor of today's video is
6:57Firecrawl. They call themselves the
6:59context API for AI agents. I covered
7:02their incredible search a few weeks ago,
7:04but this time I want to show you
7:05Alexandria, which they just launched. It
7:07is a knowledge library, a massive
7:09catalog of over 100 data providers your
7:12agents can pull from directly for things
7:14like benchmarks, company data, job
7:16boards, podcasts, and Firecrawl's own
7:19indexes for research papers and
7:21developer docs. And I'm able to access
7:23Alexandria directly in my coding agents
7:25like Claude Code through the Firecrawl
7:28MCP server. So here I asked it to use
7:30Alexandria to answer the question which
7:32openweight model gives me the most
7:34intelligence per dollar for coding
7:35agents right now which is a genuinely
7:37interesting question for me. And so
7:39first it calls into Alexandria to figure
7:41out the right data provider and then for
7:43free it can determine the cost of
7:45searching through that data provider and
7:47then after that just a single call to
7:48pull the full leaderboard and get me the
7:51answer. And now that I found this model
7:53I want to hook it into my cloud code.
7:54And so Firecrawl also offers their
7:56developer index which allows me to
7:58quickly search through a bunch of GitHub
7:59issues, pull requests, readmes,
8:01documentation. It finds the setup very
8:03quickly here to incorporate this model
8:05with Cloud Code and even find some
8:07gotchas from the GitHub issue list for
8:09the Cloud Code GitHub repo. Firecrawl is
8:12free to get started with a,000 credits
8:13per month and browsing everything in
8:15Alexandria doesn't cost anything. I'll
8:17have a link to them in the description.
8:19Now, not everything is reverse
8:21engineering where we have the perfect
8:23reference. But for a lot of what you
8:24build, you can give a reference or even
8:26craft one for at least part of your
8:29intent. And so I want to give you an
8:30example of that from my own work. And
8:32then I want to end the video by covering
8:34the test to help you identify as LLM get
8:37more capable. Is a skill still worth
8:40learning? Because yeah, there's a lot
8:41more than just intent. I don't want to
8:42just leave you with that, even though
8:44that is the main example for this video.
8:46Cool. So now let me show you an example
8:48from my own work. So I've been spending
8:50a lot of time refining my AI software
8:52factory recently and building a lot of
8:54different applications with it just to
8:56test it out. And so I've covered my
8:57software factory a lot on my channel
8:59already. I'll link to a video right
9:01here. But the main idea is that you give
9:03it a PRD where you've done a lot of
9:05planning and of course communicating
9:07your intent. And then the factory is
9:09responsible for autonomously splitting
9:11it into tasks and then knocking them out
9:13each one at a time to build out your
9:15full application. So very autonomous,
9:17but you're still very much a part of the
9:19process building that initial PRD. And
9:22so I've been experimenting with this a
9:23lot throughout the year, and it's
9:24getting really, really good now. And so
9:27this is an example of a game that I
9:28built with it. And so I let the factory
9:30go for maybe like a dozen laps to build
9:32out the initial proof of concept for
9:34this. And it's a it's a great game. And
9:36so genuinely, like this isn't like
9:38something I'd go and publish on Steam
9:39right now, but it's a good starting
9:41point. I could definitely build on top
9:42of this. But the problem is the real
9:45game. I didn't actually want to be 2D
9:47like this. I really wanted a 3D game.
9:50But the problem is I'm not a graphic
9:52designer or an artist in any way. I
9:54actually had trouble figuring out how I
9:56wanted to communicate that this needed
9:58to be 3D to the agent when I was first
10:01building the PRD. Without a reference,
10:03it was really hard for me to get the
10:05agent to actually do a good job making
10:07something 3D. And so I ended up having
10:09to settle for this this 2D version of
10:12the game. So, it works well, but it's
10:14not what I actually wanted because I
10:16failed to communicate my intent. And the
10:18reason I know it's not the agent's fault
10:20is because of what I'm about to show
10:22you. And so, going along the same lines
10:24as reverse engineer Anything, I figured
10:26instead of trying to meticulously
10:29explain the 3D graphics and exactly how
10:31I wanted it to look, what if I gave it a
10:33reference to work towards and set up the
10:35workflow in PRD to actually have it
10:38build towards that and then compare
10:39things side by side to validate. And
10:41that is exactly what I did with this
10:43game. I've always really loved the
10:45graphics of this Legends of Zelda game.
10:47It's super unique cuz it's not fully 3D
10:49like, you know, Super Mario Galaxy or
10:51like of course a firstp person shooter,
10:53but it's also not like fully top down
10:55like Pokemon cuz that honestly is more
10:57like just top down 2D. I don't know. I
10:59just I love these graphics. And so I
11:01figured what if I give this to the
11:02agent? I plan with it up front like
11:04here's the game I want to emulate the
11:05graphics of maybe just a little less
11:07toyish and then I want you to constantly
11:09compare the output of the game to this
11:12and just align on the graphics over
11:14time. And so this worked really really
11:16well. So here is a very initial proof of
11:19concept for the game that I built with
11:21the factory using the Legend of Zelda
11:23game as the graphics reference. This is
11:25the first time I've actually attempted a
11:273D version of this game probably like
11:29five or six times already. This is the
11:31first time it actually looks close to
11:32what I wanted because of this whole, you
11:34know, harness around validating against
11:36the graphics of the other game that I
11:38specifically chose. And there's a lot of
11:40planning and designing that went into
11:41that. So, I definitely applied my mind.
11:44A lot of critical thinking went into
11:46getting here, even though I'm using a
11:48software factory with a very autonomous
11:50version of AI coding. But, I love this
11:52and I just got started building this.
11:53Literally, I've had the factory only
11:55running for like a day or two on this
11:56application, but it's already starting
11:58to look really, really good. And just
12:00for fun here, I'll show you what a
12:02battle looks like. Kind of like what we
12:03had with the 2D version of the game.
12:06Definitely a lot of things I have to
12:07tune up like this weird circle appearing
12:09once in a while. U but yeah, like the
12:11combat is actually pretty smooth. Yeah,
12:13the game is just fantastic. And the
12:15important thing here is with every
12:16failed version of the 3D game, I'm using
12:18the most powerful LLM, Opus 5.5, with a
12:21very elaborate harness, my software
12:23factory. yet I still need to spend a lot
12:25of time communicating my intent because
12:27without that your agent will write good
12:29code that misses the point. That's why
12:31it is the most important skill to learn
12:32and it doesn't matter how much more
12:34powerful LLMs get. And so that segus
12:37really nicely into the test. We've
12:39talked a lot about communicating our
12:41intent as the most important skill, but
12:43there's a lot of other important skills
12:44as well. The question to ask yourself is
12:47if the next model is twice as good. Does
12:50this particular skill we're analyzing
12:52matter more or less? And I would say if
12:55it matters the same amount, it's, you
12:56know, obviously still worth putting your
12:58time into learning it. And there are a
13:00lot of things that matter less. Now, as
13:02LLM get more capable, we have to deal
13:04with prompt engineering less. Like those
13:06little tricks that used to help you get
13:08better output just don't matter anymore.
13:10We don't have to roll our own loops.
13:11LLMs do a lot of reasoning and walking
13:14themselves through a workflow
13:15automatically. And we have things built
13:17in harnesses like /goal. Giant rule
13:20files. Even especially in the last few
13:21months, we've learned that really long
13:23rule files trying to communicate all of
13:25our constraints and conventions can
13:27sometimes just bloat the agent more than
13:28help. LLM don't need to be taught
13:31general software engineering principles,
13:33for example, like they used to be maybe
13:35a couple of years ago. And it's crazy
13:37because even a year or two ago, these
13:39things used to matter a lot. like prompt
13:40engineering and crafting good rule
13:42files, but it's just not where you
13:43should be investing your time. Where you
13:45want to invest your time is everything
13:47at the bottleneck with the planning that
13:49you do before writing the code and the
13:51validation after, right? especially
13:53communicating your intent, but even just
13:55like the harness engineering as a whole,
13:57building your skills and sub agents and
14:00hooks and bringing that together into
14:02the more comprehensive system for your
14:04agent. Like what I'm working on with the
14:05software factory, what the person
14:07engineered for reverse engineer
14:09Anything, right? For every single
14:11powerful system, there's someone who's
14:12put a lot of critical thinking into this
14:14system to set the agent up for success.
14:17Because as LLMs get more and more
14:19powerful, we're just going to give it
14:20harder and harder tasks. And so we still
14:22need our conventions and workflows built
14:25in. And so yes, we're not going to focus
14:27as much on the rule file and dedicating
14:30hours and hours and hours to crafting
14:31out like we did before, but going up to
14:33the higher level. This system is still
14:35very critical. Like I always tell
14:37software engineers, we are no longer
14:38engineers of the codebase. We have
14:40delegated that to the agent. Instead, we
14:43are engineers of the harness for the
14:45agent, setting it up for success and
14:47using every mistake as an opportunity to
14:50make our agent better. And so, for some
14:52engineers, they don't like this. You
14:54might like coding, but you don't like
14:56building these harnesses. I get it. It
14:57doesn't jive with everyone. If that's
14:58you, I am truly sorry. Like I understand
15:00that your role might not be as
15:02fulfilling, but at least for me, and I
15:04hope this is for you as well, building
15:06these harnesses is very satisfying and
15:08fulfilling because you're getting the
15:10job done in the codebase just in a
15:12different way. And it's so cool to see
15:14your harness get better and better over
15:15time. That data flywheel making it so
15:17you just operate more and more
15:19efficiently, assuming you're applying
15:20your critical thinking to build it out,
15:22and you're always making sure that you
15:24communicate your intent and you're a
15:25part of that process. And so with that,
15:28that's everything I had to cover in this
15:29video. I hope the test was helpful, just
15:31helping you think about what are the
15:33skills still worth investing your time
15:34into because there are definitely still
15:36a lot even with LM's getting so capable,
15:39especially with this biggest example
15:40here of we always need to communicate
15:42our intent and have a part of the
15:44harness for that. And so I hope that you
15:46found this useful. And if you're looking
15:47forward to more things around agentic
15:50engineering and what I've covered here
15:51with intent engineering, I would really
15:53appreciate a like and a subscribe. And
15:54with that, I will see you in the next

About this transcript

This page contains the full transcript of Is Learning to Build Software Still Worth It with AI? (My Honest Take) by Cole Medin, generated from the public captions YouTube serves with the video. The transcript has 3,455 words across 488 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.