Daniel Páscoa

Trained as an actor. Became a software engineer. Now I build things.
Technology, business, work & life.

[Daniel Páscoa @dpascoaa]

@justwatt The gap I see is between the people writing about AI and the people actually shipping with it day to day. The ones doing the work aren't panicking or hyping, they're just quietly getting faster.

[Daniel Páscoa @dpascoaa]

@oderajoseph This is the shift that actually matters. Once you stop treating it as one assistant and start assigning it roles, you're doing the job of an architect, not a typist. That's a different skill and most people never make the switch.

[Daniel Páscoa @dpascoaa]

@BenDenny @max_spero_ This is the same failure mode I keep hitting from the engineering side. A CV gets you past the recruiter, not the interview that actually tells them anything. Nobody's built the equivalent of asking a writer for a sample piece, for code.

[Daniel Páscoa @dpascoaa]

@sethmills21 The instinct to strip back to one file is usually right. Every layer of custom tooling is something that has to keep working, not just something that helps once. Curious what made you pull the trigger today versus just living with the sprawl.

[Daniel Páscoa @dpascoaa]

@grinich @steveruizok @tldraw @byteofbits @attio @TomaPuljak @daytonaio @incident_io @almonk @dominikkoch Was there tonight. The tldraw and https://t.co/EVbdKDmnuv demos were the ones that stuck, real workflows people are actually running, not another "AI changes everything" pitch. Good room. Thanks for hosting.

[Daniel Páscoa @dpascoaa]

@theyouwell This is the right split. Use the model where ambiguity is useful, then make the decision layer clean enough that a human can understand why it happened. Otherwise you have speed, but no control when it goes wrong.

[Daniel Páscoa @dpascoaa]

@ashishllm This matches ten months of interviews for me. The ones that go well are never the AWS-trivia ones, they're the ones where someone asks what I've actually shipped. The absurd JD is a proxy for a skill nobody's figured out how to screen for yet.

[Daniel Páscoa @dpascoaa]

@Ionkosm The habit argument undersells it. Switching cost isn't laziness, it's that the incumbent already knows your context and the challenger starts from zero every time. Distribution wins until someone builds a product good enough that relearning is worth it.

[Daniel Páscoa @dpascoaa]

Going to Agent Night in London tonight.

I'm not really interested in hearing that AI is going to change everything. We know that already.

I'm much more interested in what people are actually building with it, what is working in the real world and, probably more importantly, what still isn't.

Let's see.

[Daniel Páscoa @dpascoaa]

It's crazy how quickly "impressive" becomes normal.

A few years ago, building a really good piece of software required a team, a lot of money and probably months of work.

Now one person with the right tools can build a ridiculous amount.

So I keep thinking...

If building becomes easy, what actually becomes valuable?

[Daniel Páscoa @dpascoaa]

The line that stands out is treating the build as 10% of the work. Most engineers keep optimizing the part they're already good at because it's more comfortable than the part that moves revenue. The revenue-per-user number is the one worth stealing, it turns every decision into arithmetic instead of an argument.

[Daniel Páscoa @dpascoaa]

@dadhalfdev This matches what I've been running into from the applicant side. I'm not the person who can recite an AWS service catalog from memory, I'm the person who gets a working thing deployed. Trivia interviews can't tell those two people apart.

[Daniel Páscoa @dpascoaa]

@LeoBuilds_ Fair challenge. Good gates change what "checking" means, they don't remove it. The real question is what your eval covers, not whether a human read every line. I'd still want someone accountable for where the gate is blind. Aerospace calls that a coverage gap.

[Daniel Páscoa @dpascoaa]

@SSShken That is exactly what I mean by trust. Not only that it completes the feature, but that it behaves cleanly inside someone else’s codebase. Adding seven dependencies without asking is not speed, it is creating work for the person after you.

[Daniel Páscoa @dpascoaa]

@dev_aniketmalik I agree with the arithmetic, but I think there is another variable: some jobs leave you with hours on paper and no useful energy. Ten hours a week only counts if they are real building hours, not the leftovers of a week that has already taken everything.

[Daniel Páscoa @dpascoaa]

@thevikassyadav The “remove Google” test is the cleanest part of this. It forces the question most SEO-led products avoid: would a real person still come back if there was no ranking to chase?

[Daniel Páscoa @dpascoaa]

The sequence matters. Seven agents without earned trust at each hand-off is just seven new places for rubbish to hide. What was the first workflow you felt safe enough to extend?The sequence matters. Seven agents without earned trust at each hand-off is just seven new places for rubbish to hide. What was the first workflow you felt safe enough to extend?

[Daniel Páscoa @dpascoaa]

@admgrn01 This is exactly it. Context is useful, but if you have not decided what good looks like, you are just asking the model to produce more things faster. I think planning has become more valuable, not less.

[Daniel Páscoa @dpascoaa]

@RomanShalabanov The cleanup point is the one people miss. AI makes starting cheap. It does not make a messy product easier to understand six months later. Shipping fast still needs a clean baseline.