Insider insights from a BAML x HumanLayer industry webinar
The Unexpected Reality Check
I signed up for what I thought would be a typical tech webinar about Claude Code. Instead, I ended up in a surprisingly candid conversation with industry developers about how AI is actually changing their day-to-day work.
Turns out, the skills that matter most aren’t the ones anyone’s talking about in CS classes.
1. Specification Writing
The new baseline skill that no one’s teaching
The shift is subtle but real: companies are moving from “can you code?” to “can you write specs that produce working code?” The problem? Spec-writing traditionally required senior-level experience, but now it’s becoming entry-level baseline.
It’s like being asked to be the architect before you’ve learned to lay bricks. Most CS programs still focus on implementation, but the developers in this webinar were clear—the ability to write crystal-clear specifications is becoming the differentiator.
Think “rubber duck specification”—explaining your approach so clearly that an AI (or junior dev) could execute it perfectly without asking follow-up questions.
2. AI Code Review
The skill everyone needs but no one has
Here’s the uncomfortable truth: most teams are doing “looks good to me” reviews on AI-generated code because nobody knows how to properly evaluate it. It’s like having a really productive intern who might be brilliant or might be completely wrong, and you can’t tell which.
The developers on this webinar admitted they’re basically winging it. Some teams are trying to create metrics, but then the AI just games whatever metrics they come up with. It’s like trying to grade a test where the student can see all the answers.
The gap here is huge. Most people treat AI as a black box, but someone who can actually evaluate its output? That’s valuable.
3. Context Management
The hidden multiplier that changes everything
Here’s a technical detail that changes everything: Claude Code uses about 50% of its context window just to behave properly. The other 50% is all you get for your actual conversation. Most developers don’t realize this and hit limits way sooner than they should.
It’s like having a messy desk versus an organized workspace. The organized person doesn’t work harder, but they get exponentially more done because they’re not constantly searching for things.
As companies start running multiple AI agents in parallel, the developers who can structure efficient conversations will have a massive productivity advantage over those who just wing it.
The Transition Period
Where we are right now
We’re in this weird moment where the tools are evolving faster than anyone’s wisdom about how to use them. CS curricula haven’t caught up, and most career advice is still based on 2020’s job market.
But that’s also the opportunity. The developers who figure out these meta-skills early will have an edge as the space matures.
What Makes This Interesting
This isn’t about learning another framework that’ll be obsolete in two years. These are fundamental shifts in how software gets built. Whether it’s Claude Code, GPT, or whatever comes next, the underlying skills—writing specs, evaluating output, managing context efficiently—those translate.
It’s like learning to drive versus learning to operate a specific car model. Once you get the core skills, you can adapt to any vehicle.
The Bottom Line
Thanks to BAML and HumanLayer for hosting a conversation that went beyond typical “here’s how to write better prompts” advice. Getting insights from developers who are actually using these tools in production was refreshing.
The technical tools are evolving rapidly, but understanding how they change the work itself? That’s the more interesting question, and one that’s worth thinking about early.
What patterns are you noticing in your own experiments with AI coding tools? Always curious to compare notes.