As AI takes on more and more of the work people used to do, every job function — not just in Silicon Valley but everywhere — is wrestling with hard questions about careers. A senior engineer at Meta has completely changed how they work by leaning on Claude, and in the process has felt firsthand how the productivity gap now stretches 10x to 20x depending on how well someone wields AI. Here are twelve practical career tips from that Meta engineer, covering deeper understanding of agents, shipping fast, executing across levels and teams, and staying focused on the real problem.

1. Master AI

Simply knowing how to use AI is no longer a differentiator. Now that everyone uses AI as a basic tool, the real gap in ability shows up in how deeply you understand AI and how skillfully you handle it.

This gap is clearly visible inside Meta. People who handle the tools well produce overwhelming impact — 10x, 20x and beyond — while those who do not remain stuck at their old way of working and old level of productivity.

"At first the difference honestly wasn't that big. Everyone was fumbling about equally. But looking at what's happening at the company these days, the people who are good at it are producing 10x, 20x impact without breaking a sweat."

2. Become an agentic engineer

Just because we're in an era where one click seems to do the job doesn't mean engineering knowledge has become unnecessary. On the contrary, the gap between what an engineer who truly understands and controls agents produces and what everyone else produces is night and day.

To stay competitive, you have to dig deep into the core elements of agentic engineering:

• Prompt and context engineering: carefully designing the background knowledge and context you hand to the agent
• Harnesses and hooks: building the safeguards that control agent misbehavior and steer it toward the behavior you want
• Token efficiency: optimization techniques that avoid wasting cost and compute
• Evals and feedback loops: structures that objectively measure agent performance and improve it continuously

3. Ask AI (breaking the learning bottleneck with visualization)

When absorbing a new codebase or a large body of knowledge, how well you use AI as a learning tool makes a huge difference in how fast you learn. Reading long, complex text straight from an AI tends to overload your brain.

To break that learning bottleneck, asking for the output to be visualized is a highly effective technique.

"I found that reading the long text these agents spit out actually made me understand less. I kept experiencing how easily my brain overloaded. So these days I use this visualization skill a lot at work — I have the agent visualize its output."

When information comes back restructured as charts, infographics, interactive HTML pages, or flow diagrams, you can grasp the overall structure at a glance and quickly spot gaps that are easy to miss. 💡

4. Learn in your bones where AI fails

Beyond fixing settings each time an agent makes a mistake, what matters is developing a felt sense for what models are fundamentally good at and the patterns in which they fail. That intuition is an asset only available to people who don't ration their tokens and instead run into things themselves, in many different ways.

When a new model ships, prompts and behavior patterns that used to work well can change completely.

"There's a well-known line from Boris Cherny, who built Claude Code: when the model changes, he deletes all the prompts and runs it from scratch. When the model changes, the way it works changes too. So you actually keep the harness lighter."

Every time a new model arrives, you have to drop your existing assumptions and flexibly re-explore where the failure points are.

5. Taste is the moat

The more advanced AI becomes, the more human competitive advantage comes down to taste. Taste here doesn't mean some grand artistic sensibility — it means the ability to read the customer's hidden intent and feelings, the things not written down in any document.

• Customer-centered attention to detail: catching the subtle needs of colleagues for an internal tool, or of end users for a commercial product
• Communication and networking: the context beyond the written document comes from frequent conversation and collaboration with colleagues and customers

Even when an agent can execute the implementation perfectly, the final touch that decides the details that fit a user's taste exactly is still a human's job. 🎯

6. Ship before you're 100% ready

Delaying a launch until every feature and every design detail is perfect blocks fast execution. The tools that actually succeeded inside Meta were, without exception, released quickly in an unfinished state.

The important distinction here is between 70% polish and 70% accuracy.

"Aiming for 70% polish and aiming for 70% accuracy are two different things. For instance, buttons that aren't quite aligned, an error message that isn't pretty, code that's a bit messy — I think you can ship that."

If the accuracy of the core functionality is verified, accept the minor blemishes in the UI and the messy code, ship immediately, and finish the product with feedback from real users.

7. Work regardless of your level

Roles used to be strictly divided among designers, PMs, and developers, but in the AI era those boundaries of level and role are collapsing. Waiting indefinitely for help from another function because you lack expertise in that area just creates a bottleneck.

"Say I have work to do and I'm blocked on design. You can't go to a designer any more and ask them to unblock you. I have to be able to do the design. Not by doing it myself — you have to build the habit of doing that work through an agent."

When you hit a moment mid-task where you need help from another discipline, the essential posture is to break through it yourself with an agent rather than waiting on a colleague. 🚀

8. Don't be bound by your team

The era of filing a ticket with the infrastructure or platform team and passively waiting for them to prioritize it is over.

In one case during a hackathon, a project nearly stalled because the needed infrastructure support did not exist — so the engineer used agents to implement and deploy the other team's infrastructure support code directly.

"I built that support myself and even deployed it. And the funny part is that during the process the lead of that team reviewed my PR. It set a good precedent. 'My area' no longer exists. If an agent can do it, then it's my area."

You should be able to use agents as leverage to cross your team's boundary and take the initiative on problems that belong to the whole company.

9. Aim to be in the top 1% (use Claude as a coach)

To deliver the best results in your organization, you have to set your target as high as the top 1% and move strategically. One useful approach is to use Claude as a personal coaching system.

1. Inject full context: teach Claude not only your own work but your teammates' work, and beyond that the roadmap and open issues of the entire department.
2. Extract high-impact tasks: ask it to produce "the 10 core problems I should solve right now to become a top-1% engineer delivering 10x, 20x impact in this department."
3. Validate with leadership: take the resulting candidate list and discuss it directly with your tech lead and manager, then focus on the highest-impact areas.

10. Solve the problem (the essence of AX)

Skill with AI tools and the ability to actually solve problems are two separate domains. Meta also encouraged heavy AI use early on, but now the focus is on which problems you actually solved and what impact you created.

"It's a bit like handing a carpenter a chainsaw instead of a handsaw. And what matters here is that being good at using AI and being good at solving problems are completely separate domains. If you use AI to solve the company's problems — that, I think, is AX (AI transformation)."

Adopting the technology cannot be the goal in itself; the starting point and the destination must always be the problem you are trying to solve.

11. Meet good people

People who are exceptionally good with AI have only just started to appear, which means the gap between the very top and ordinary users is smaller today than it will ever be again. So actively engage with the people around you who use AI exceptionally well.

• Use 1-on-1 meetings: meet outstanding colleagues and find out how they weave AI into their work and what domain knowledge they hold.
• Turn meeting notes into assets: feed the auto-generated AI notes from the meeting into Claude and extract action items from your colleague's workflow that you can apply to your own work right away.
• Analyze code and artifacts: if conversation is difficult, absorb their know-how by analyzing the PRs they submit and the documents they share internally.

12. Protect your work-life balance

The one who directs the agent and makes the final call is still a person. No matter how powerful the AI at your side, if the mind and body of the person steering it break down, you cannot draw out 100% of the tool's performance.

"If you ask what matters most in using AI well, I think it's the mental state and physical health of the person using it. You have to be healthy both mentally and physically to command an agent well."

The more your environment makes it easy to lose sleep and exercise to overwork and side projects, the more managing your condition and keeping your mental health strong becomes the final key variable determining your agent's performance. 🌿

Closing

As new technologies and tools pour out, FOMO — the anxiety of falling behind — keeps intensifying. But the era of agents has only just begun, and no one is late.

Rather than rushing to keep up with every piece of news, the surest way to grow is to start with small attempts to solve the real problems right in front of you today, one at a time, with an agent.