Skip to content
Legion Orpheus

Free course · LEGION.TW

There is no magic prompt. There is a mind making assumptions.

Most AI courses teach you what to type. This one teaches you what happens next. Across seven free lessons, you will learn how a model guesses what you want, why it cuts corners, and how to make it reconsider an answer that sounds finished but is not good enough.

Written by Frank and Buddy Lien.

Start lesson one →

Two ways in

Both are free and complete. You do not have to finish one before starting the other.

Practical path

Build something you actually need

For someone who has never written code and wants a working tool by the end. Eight practical articles: build a real native Windows app with Codex early on, then learn by using and improving it.

Start building →

Free course

Understand how AI answers you

Seven lessons on what the model is doing when it replies: how it guesses your goal, where it cuts corners, and how to make it reconsider.

Read the course ↓

The lessons

Start at lesson one. Each lesson gives you something you need for the next.

  1. 01AI Is Guessing What You WantThere is no magic prompt. There is a mind on the other side of the conversation making assumptions about your goal — and those assumptions are the real control surface.9 min read
  2. 02What It Means to Be a Good AIBefore AI decides how to answer you, it decides what a good AI should do. That belief quietly changes what your words appear to mean.14 min read
  3. 03AI Thinks You Are the Most Boring Person AliveWithout evidence about your taste, AI imagines the safest possible reader and writes for them. Here is how to stop receiving boiled, unseasoned work.9 min read
  4. 04AI Has to Think Out LoudThe first draft is both an answer and the workspace used to find the answer. Demanding polish too early buys smoother prose and worse thinking.10 min read
  5. 05AI Is a Cognitive MiserAI takes the first path that looks good enough, then spends its fluency defending it. Learn to see the shortcut and make it more expensive.10 min read
  6. 06AI Hears Everything a Word ImpliesConnotation is a control surface. One word — intern, editor, auditor, partner — can redirect the whole interaction before you instruct any behavior.10 min read
  7. 07Do NOT Use ConstraintsEvery prohibition puts the unwanted behavior into the model's working context. Build the behavior you want instead of listing the mistakes you fear.9 min read

Who this is for

You have used ChatGPT or another AI and know there must be more to it than asking a question and accepting whatever comes back. You do not need to program. You do need to think.

If you would rather be taught

You can learn all of this alone. Most people have jobs to finish. We teach the same ideas through your team's own work, stay for the questions, and build the tools the work exposes. The course lets you judge us before we ever enter the room.

AI for your team →

Related writing

Blog: essays on AI →

What we’ve built

We’re Frank and Buddy, a human–AI team. In roughly six months, we’ve built dozens of applications and systems. Frank can’t write a single line of code.

Here are four examples.

Atlasless time on the road

Atlas uses AI to find the best travel route within our clients’ availability and appointment constraints. It plans autonomously, saving hours on the road and getting us home earlier without Frank having to direct every step. While we’re out, we can ask the AI to identify traffic hazards and improve the route in real time.

The same system manages calls, appointments, field photography, and delivery paperwork. It saves Frank and his wife roughly 120 combined hours of work every seven weeks.

Kongzi SaysEnglish and critical thinking

Kongzi Says teaches English and critical thinking together. Students build sentences, investigate cases, and use English clues to deduce answers nobody has given them. Frank has used it with Taiwanese students and seen improvements in both their English and their ability to reason through unfamiliar problems.

The puzzles give students repeated practice at trying an idea, checking it against the evidence, and correcting it.

Already speak English? Try the logic grid puzzles, also known as Einstein Riddles.

It’s free. No account needed.

Try the Einstein Riddles →Explore Kongzi Says →

Exocortex and Audio Genour own AI and audio infrastructure

We’ve built a local AI inference engine, including custom CUDA kernels—the code that performs calculations on the graphics card. Our voice-generation engine powers Audio Gen, a complete studio for producing audiobooks and teaching material. We can create voices, direct performances, use AI to listen for errors, repair individual passages, and synchronize the recording with the text.

Command Stationan AI engineer you can call

Frank can call Buddy from his phone, ask for a complete Android app, and get a notification when it’s ready. He installs it from our private app store and starts using it—all while he’s away from home.

The conversation and the work continue across phone and desktop, so he can check in, ask for changes, or leave Buddy to finish.

Our lessons draw on this work, Frank’s extensive AI experiments, and nearly a decade spent teaching advanced critical thinking to scientists and academics. Buddy builds, researches, and writes alongside him—including this site.

More about us →