If you could learn only one thing in the AI era, learn strategic thinking, says a Meta senior engineer: "Give goals, not tasks"
Nov 6, 2026Original 실밸개발자 · 클로드 코드 문서 · METR ↗
In 3 lines
- With execution near free, delegating well comes from strategic thinking more than from prompt technique
- Giving agents goals like KPIs instead of tasks turned 'done' reports into 'this moves the metric'
- Feeling is not measuring: in METR's trial developers were 19% slower yet believed they were 20% faster
"If I had to pick just one thing to learn in the AI era, it would be strategic thinking." That is the opening line of a video posted Nov. 4 by the YouTube channel 실밸개발자 (Silicon Valley Developer). Over 20 minutes it covers what is left for people once execution costs near zero, and what changes when you give an agent goals and KPIs instead of tasks. Each claim is paired with context from the Claude Code docs and METR research.
When execution costs near zero, two jobs are left for people
Silicon Valley developer channel 실밸개발자 argues that once work can be delegated to AI agents, the cost of execution drops toward zero. Human time then goes to two things: giving the work well, and verifying what comes back so it turns into real value. In the speaker's words, an output from AI does not become value just because it exists. The video focuses on the first of the two, how to delegate well.
The gap comes from strategic thinking, not prompt technique
The speaker once said prompt engineering was over, and he says his view has shifted a little. Results from a vague instruction and results from first planning a strategy with the AI began to look visibly different. He defines strategic thinking as weighing and proposing the options and routes toward a goal. That applies to fixing one bug as much as to building a revenue line.
The official docs say the same: tell the AI what it can't see, and give it a check
Claude Code's official best-practices page frames its advice as telling Claude what it can't see and giving it a check to run. Its point is that when a deploy process lives in someone's head, or nothing defines what counts as done, the AI tends to build something other than what you had in mind. That lines up with the video's advice to give a goal. Deciding the goal and the finish line comes before polishing the wording of an instruction.
Anthropic · Claude Code docs ↗Work got easier, yet people feel more tired: the 'strategic load' effect
He says the time once spent executing is now filled with daily strategy and deciding what to hand the AI. A day that demands more thinking naturally brings more cognitive load and more fatigue. As AI advances, he expects people to move up from the execution layer to the strategy layer, and to split into those who can do that and those who can't. He sees this as a personal outlook on a structure where roughly 70–80% of a team used to be executors.
Feeling is not measuring: why the METR study makes the case for verification
The video stresses that verifying output matters. METR's July 2025 experiment supports the point. With AI allowed, 16 experienced open-source developers took 19% longer across 246 issues. Yet they had expected a 24% speedup beforehand and still believed AI had made them 20% faster afterward. METR notes the result reflects early-2025 tools and has since published updated data.
METR · metr.org ↗Now a boss with an AI workforce: what his promotion talk revealed
The speaker says we have become something like a boss managing AI-agent employees. Preparing for promotion, he heard from his manager that moving toward senior and staff levels means strategic thinking is a must, and realized he had never thought about strategy. He had long believed hard skills made you an expert, but now sees soft skills gaining weight. He adds that he could be wrong.
From task-first to goal-first: a first prompt that hands over KPIs and metrics
Just as good companies align around one vision, he says an agent should be given a large goal instead of a string of tasks. A senior engineer told him that work able to move the metric already fills all your time. So when a project starts, his first prompt shares the organization's KPIs and the metric he wants to improve, then asks to build a strategy to lift the number by 10–20%. The habit began with a question: with three, four or five times as many PRs now possible, how many create value?
Put the strategy in a 'second brain' so every task remembers the goal
When work is split across sessions, an agent forgets the strategy. So he stores the goal and strategy in agent memory, a so-called second brain, so any task starts with the goal in mind. The Claude Code docs flag the same failure: having to rediscover the same fact at the start of every session is a common cause of poor results. A goal stated once is fragile; a goal kept in a file lasts.
Anthropic · Claude Code docs ↗Three things changed: quality, output, productivity, and domain knowledge as the precondition
After switching to goal-first, he says the agent reports 'this change should move the KPI' instead of 'done.' It pushes back when a metric could worsen and offers options he hadn't thought of. What changed, he says, is the choice of work rather than the amount. He adds that good strategy rests on domain knowledge, so learning unfamiliar fields alongside AI is his next challenge.
