Wink Pings

AI Will Eliminate Entry-Level Jobs First. Anthropic CEO's Advice to Graduates: Learn to Use AI

Dario Amodei says AI can already do the work of entry-level lawyers, entry-level analysts, and entry-level consultants. His advice boils down to one sentence: learn to use AI. This article includes 24 free resources, plus a harness guide that most people overlook. You're only using 5% of the tools you have.

Anthropic (the parent company of Claude) CEO Dario Amodei was asked a question: what should a kid who just graduated high school do now.

The answer was direct: AI can already do the work of entry-level lawyers, entry-level financial analysts, and entry-level consultants. It will eliminate many jobs.

In two minutes, he said which jobs will be eliminated first, and the only skill that keeps you at the table. His advice boils down to one sentence:

Learn to use AI.

People who start learning this year won't compete with AI. They'll be the ones directing AI.

95% of people will finish school and compete with AI for the same entry-level job. 5% will learn to direct AI and run the agents that do that job.

<video controls="controls" height="1080" poster="https://wink.run/image?url=https%3A%2F%2Fpbs.twimg.com%2Famplify_video_thumb%2F2124116863319744512%2Fimg%2FGaP2z1CrBkePFmOr.jpg\" src="https://video.twimg.com/amplify_video/2124116863319744512/vid/avc1/1920x1080/_E-zq7LYjtOVxwDk.mp4?tag=29\" width="1920">

Zayn compiled all the free resources in the tweet, 24 items, total price $0. He said: save this list and send it to people graduating this year.

The list is below.

Anthropic

  1. Interactive Prompt Engineering Tutorial - The fastest way to learn how to guide a model
  2. Courses - From your first prompt to managing AI agents
  3. Cookbook - Code recipes you can copy directly

Microsoft

  1. Generative AI for Beginners - 21 lessons
  2. AI Agents for Beginners

Hugging Face

  1. Agents Course - Free, with certificate

Google

  1. Generative AI Notebooks - Hands-on labs

OpenAI

  1. Cookbook - Ready-made code for common tasks

Community and Independent Resources

  1. DAIR.AI Prompt Engineering Guide - The one everyone is sharing
  2. Karpathy's Neural Networks: Zero to Hero - Understand the fundamentals
  3. CrewAI - Run your own team of AI employees
  4. Microsoft AutoGen - Multiple agents collaborate on one task
  5. Awesome LLM Apps - Hundreds of real project references
  6. DeepLearning.AI Short Courses - Free one-hour courses with domain experts

Anthropic x Coursera Series

  1. AI Fundamentals: Claude
  2. AI Automation: Claude
  3. AI Collaboration: Claude
  4. Building with the Claude API
  5. Model Context Protocol Advanced Topics
  6. Teaching AI Fluency
  7. Real World AI for Everyone Specialization
  8. Claude AI for Marketing
  9. Learn Claude from Scratch
  10. Claude Code: Software Engineering with AI Agents

Total price: $0.

The list is only the first step. In the second half of the same post, Zayn points to a guide that most people overlook: 'Your Harness Is Already Installed. You Are Using About 5% Of It' by anstap.

This article answers a question: using the same Claude Code, Codex, and Cursor, why can some people only chat while others can run an automated team?

Core idea: take the model out of your setup, and what's left is the harness. The model proposes; the harness decides what it can see, what it can touch, what survives after context resets, and what counts as done.

Prompt engineering edits instructions; harness engineering edits the conditions under which instructions run. Anthropic itself used the term 'context engineering' to describe this a year ago.

You don't need to build from an empty folder. The tools you're already using come with a harness: 33 hook events, permission layers, subagents, skills, and memory after the session ends. Most people only use the chat box and a markdown file.

A rule can be placed in four places. The costs are completely different.

LocationCan the model ignore it?Cost
In your headAlways, relying on you to checkConsumes you on every run
In instructionsYes, and often silently fails when context is compressedConsumes tokens every turn
In a hookNo, it will run anywayOne-time 20 minutes
In permissionsIt can't even tryOne-time 5 minutes

A rule in your head isn't a rule; it's a job you've created for yourself. 'Don't touch the migration files' is a wish; a permission that denies access to migration files is a wall.

Hooks are the part everyone skips. Instructions live in the conversation, but the conversation gets compressed, flooded, and eventually ignored. A hook is code; it runs at fixed moments, regardless of whether the model remembers.

Claude Code has 33 hook events. The six most commonly used:

  • SessionStart
  • UserPromptSubmit
  • PreToolUse
  • PostToolUse
  • PreCompact
  • Stop

A PreToolUse hook can block an Edit or Write call directly when it happens. A Stop hook can check whether tests pass when the model thinks it's done. If the script exits with a non-zero code, the model gets sent back to redo the work.

A hook worth writing doesn't even need a matcher:

{
  \"hooks\": {
    \"Stop\": [
      { \"hooks\": [
          { \"type\": \"command\",
            \"command\": \"python tools/check.py --last\" }
      ]}
    ]
  }
}

Rule: if you've already typed the same revision feedback a third time, stop typing it manually.

Context is a budget, not a container. A larger window doesn't solve forgetting; it only delays it, and makes every turn slower and more expensive.

Three solutions already in the box:

  • Write decisions to disk, not to chat. Files are the only memories that don't get summarized.
  • Give subtasks their own window. A subagent reads 40 files and returns a six-line conclusion; those 40 files never touch your context.
  • Save before compaction. A PreCompact hook writes the current plan to a file; compaction loses details, but the file doesn't.

Something with a harness but no state is a very confident goldfish.

You might think harness is only for engineers. It isn't. A check only needs to be able to fail; it doesn't need a test suite.

Writing: a script flags banned phrases, length, and repetition.
Video: frame-by-frame checks for object clipping.
Research: every number traced back to a source that was actually opened.
Design: rendered on a phone screen at real size.
Data: print row counts and date ranges before and after.

The original author says their setup is just three things: a rules file loaded every session; a script that checks finished text; and a script that checks 3D renders frame by frame for clipping. None of them are clever, but they all catch problems the author can no longer see.

The key trade-off: harness doesn't make you less involved; it moves you to the part that's actually useful.

The original author also took wrong turns. The first time, they wrote 400 lines of rules, which were almost useless; the first hook had no matcher, blocked all Bash, and was turned off the next day; they left decisions in chat, the decisions disappeared after context compression, and two days later they argued for the same bad idea again—and won, because the counter-evidence no longer existed.

Three lessons: keep instructions short, keep hooks narrow, and write decisions to disk.

Harness can't save a problem that was wrong from the start. When the task is vague, the permission system only makes the vague task execute more safely. A check can only catch what it was designed to catch. The original author's checks can catch long sentences and dead phrases, but they never told him an idea was boring.

The fastest diagnosis: try to write down what 'done' looks like. If you can't write it, don't automate yet.

Five things you can do this week:

  1. Write a rules file: what this project is, and what is absolutely not allowed.
  2. Add a permission rule: the one command you least want to see run.
  3. Add a Stop hook: refuse to finish when the check is red.
  4. Write a check that can fail: a 30-line script targeting your specific work.
  5. Create a memory file: decisions and dead ends, written to disk.

Do them in order. None should take more than one evening.

TL;DR version:

  • Harness is what's left when you remove the model.
  • You already have most of it. The switches are in the settings file you've never opened.
  • Rules in your head always consume your attention; rules in a hook cost only one evening.
  • Context is a budget. Decide what gets forgotten, or the system decides for you.
  • You don't need a test suite; you just need something that can say 'no'.

Save this article, open your settings file, and flip one switch.

发布时间: 2026-11-22 09:01