How to do valuable work in 2026
Some thoughts on how to do valuable work in 2026 (as of Oct 1). I originally wrote this for my engineering-heavy team at Peanut. It was extracted from a series of internal threads, as well as a 1:1 check-in with one of our engineers. I thought I'd share it more widely.
AI writes code better than any of us now. Opus 5.5 is a better coder than me, and you. Knowing what some React hook does, or how a specific API works, used to be specialized knowledge. We all took years to master it, pained through a STEM education or self-learned. Now it's pretty much worthless: it can be bought with a $200 subscription. If your value is "I know javascript and write good frontend code", that's ngmi. AI is savant-level at directed work. It can do a 200k LoC refactor better than I ever could. It can ship basically any well defined spec with zero issues.
Same model as my peak post: value is the length of the arrow. Slide the year and watch what happens when writing code stops being scarce. Drag to rotate.
So what still remains valuable? The other parts of the job. Marrying messy, complicated requirements to the right product features. Thinking holistically. Being an excellent communicator. And having initiative.
Here are the areas I think each of us should be excellent at:
Writing code is the visible tip. The rest of the job sits under the water. Hover a layer.
Context.
The messy, scattered stuff AI can't see. Your brain holds many short and long-term things: Tech debt we talked about 3 months ago, a user who complained on Crisp about the exact thing you're touching, a file or project two people are doing in parallel without knowing it. You also have product and engineering taste you only get from years of shipping. In contrast, AI might've read the entire internet, but has no peanut-specific context; only the 1M token session context we give it (and that's only as good as our documentation and our instructions) Much of our knowledge is still somewhat ephemeral: lives in calls, in a coworker's standup that day, another Claude session or just various half-formulated thoughts. It is your job to manage this and keep the bigger picture in mind.
The agent only sees what is written down. We keep piping in more (support tickets, meeting notes), but it's partial and lossy. Drag its context window around, then toggle what you know.
Don't act like a 2016 Facebook engineer who goes through their tickets and thinks that's being effective at the job. It is not. Any AI session can crush a semi-well defined ticket nowadays, almost always one shot, especially with our strong CI/CD pipeline. What is valuable today is the bigger picture. How does it relate to other tickets? Is there a smell? Am I fixing a localized instance of a larger pattern? If so, raise it! Change scope! Nobody wrote assembly in 2016, and in 2026 you shouldn't just write TypeScript either.
Seven one-off fixes. Seven problems, or three? Press "find the smell".
Product feel & user empathy.
AI is godlike with code, and crushes most text-based tasks. But it doesn't really know what it's like to hold a phone on a bus, fluidly tap through the app, or fail a QR payment at a dinner w/ friends and feel your stomach drop. It has much poorer training data for this type of thing. Ultimately, it has poor empathy for users and poor product "taste". You have, and should improve, both. Watch PostHog recordings, read support tickets, use the app every day, look at how the best apps do it. Then design for that feeling, not just the ticket.
Slide the spinner time. The logs say 200 OK every time. Your stomach disagrees.
Initiative.
AI is very direct. It works inside the box we give it, and it doesn't escalate or switch context well. We, however, can go up and down the whole stack: Crisp, Discord, PostHog, a user interview, a competitor app, even a conference. When you smell something, chase it to the root and escalate if needed. Break your rails. Silent failure, silent disagreement and checkbox work are the worst outcomes. "I disagree, let's talk" is always fine. And be clear: "I own this, done by Friday", not "we should maybe look at this".
AI works inside the box you give it. You can go wherever the problem leads.
Communication.
Nobodies work is just PRs anymore. As part of delivering value at a higher level, it also means being excellent on communication. This is not just being a fast responder in discord, or not letting stuff slip and itemizing properly. It's not just writing a clear standup on time, or having agenda items for a meeting ready, or taking meeting notes. These things are the basics. Good communication goes beyond, and it means sharing specs, mockups, thoughts and ideas, presented well and easy to digest.It means elevating yourself to not just be a code monkey. Claude is the code monkey. We don't need that. We need you to notice a smell, take a step back, reevaluate, and present this larger level picture and idea with your peers and lead, and implement it once aligned.
Same task, two first messages. A vague one costs everyone else time.
The world is moving scary fast, and it seems like we're only accelerating. I'll probably have to write another update in January. Keep your wits about you. We're a small set of high agency people; I trust you can all evolve and adapt to these new circumstances (scary, but also exciting! You can literally get so much shit done now).
Next update: January.
Hugo
PS: Rewatching this is worth it (again):
PS2: Show more graphs, because graphs are fun
Zoom out. The line you are fixing is one speck in someone's week.
The market value of knowing a framework really well. Hover to read a year.
Every step up the stack turns the step below into a commodity.
Velocity is not progress. Switch the team and watch both lines.