Same task.
Different prompt.
A written course in prompt engineering. It begins with how a language model actually works, so every technique that follows arrives as a consequence rather than a trick to memorise.
Most material on this subject is a list of tricks with no stated reason. That works until a trick stops working, and then you have nothing to reason from. Everything here is taught as a consequence of the mechanism underneath it, so it survives the model changing under you.
Curriculum
60 lessons
- 00Foundations12
- 01Beginner8
- 02Intermediate7
- 03Advanced9
- 04Model Playbooks6
- 05For Students7
- 06At Work5
- 07Cost of a Prompt6
98 sources cited across roughly 6 hours of reading. Free, and nothing is a placeholder.
What a constraint is actually worth
One task, one model, and a prompt that gains a single constraint at a time. Drag the dial and watch what each one buys — this is the whole argument of the course in one control.
Announce a product change to existing users
Prompt
Write an email announcing our new feature.What it produces
Subject: Exciting News! Hi there, We're thrilled to announce the launch of our newest feature! Packed with powerful capabilities, it's designed to streamline your workflow and take your productivity to the next level. Try it today and let us know what you think! Best, The Team
Starting pointNo reader, no product, no constraint. Every gap gets filled with the safest guess in the training data. Drag the dial to add one constraint at a time.
What makes this different
Most prompting material is a list of tricks with no stated reason. When a trick stops working you have nothing to fall back on. Five decisions follow from fixing that.
It starts below prompting.
Track 00 teaches tokens, attention, training and inference before a single technique. Every rule you meet afterwards arrives as a consequence of something you already understand, which is what makes it transferable when a model changes underneath you.
The dependency, in order
- 00·1Tokenswhat the model actually reads
- 00·2Attentionwhy order and position matter
- 00·8In-context learningwhy examples work
- 01·4Few-shot promptingthe technique, explained
Read bottom-up and few-shot prompting is a rule to memorise. Read top-down and it is the only thing that could have worked.
Sourced, or it is not written.
Technical and historical claims carry citations to primary material - original papers and official provider documentation, not blog summaries. Every full lesson ends with the list. This is the actual further-reading block from the training lesson.
Further reading
How models are trained- 01Training language models to follow instructions with human feedback (InstructGPT)Ouyang et al., 2022
- 02Constitutional AI: Harmlessness from AI FeedbackBai et al., Anthropic, 2022
- 03Constitutional AI: Harmlessness from AI Feedback (overview)Anthropic Research
- 04Claude's ConstitutionAnthropic
Every lesson ends in your own editor.
Each one closes with a try-it-yourself exercise built to make the mechanism visible on your own prompts, not to test recall. They are designed so the result surprises you slightly - that is what makes it stick.
Try it yourself
Ask a model to count the letter e in entrepreneurship, then to spell the word backwards. Note the failures, and connect them to tokenization rather than to intelligence.
Then ask the same question twice: once demanding answer in one word, once with work through it, then give your answer. You have just observed the difference between a model that had tokens to think with and one that did not.
From 00·1 — What a language model actually is
Reading is the product, so reading is designed.
No videos. That means typography is functionality, not decoration. Lesson pages hold a measured reading column, and carry their own light and dark setting because a twenty-minute read at night is a different scene from a marketing scan.
Reading surface, measured
- Measure
- 68ch
- Line height
- 1.68
- Body face
- Literata
- Contrast, dark
- 8.4:1
- Contrast, light
- 9.4:1
- Reading theme
- Per reader
Progress you can actually see.
Your free account tracks which lessons you have finished, and the curriculum shows it as accumulated state rather than a percentage. The grid fills in as you read.
Your bench
11 lessons finished
Illustrative — this is your grid once you start
Every lesson, one cell
The full curriculum laid out at once. Every cell is a written lesson — objectives, worked examples, common mistakes, an exercise, and sourced further reading. Nothing here is a placeholder.
What a language model actually is, how it was trained, and why prompting works at all. Start here even if you have been prompting for months.
- 001What a language model actually is
- 002Neural networks and the transformer
- 003How models are trained
- 004Inference mechanics
- 005A history of AI, told properly
- 006The people who got us here
- 007The kinds of models in use today
- 008Why prompting works at all
- 009Where the training data comes from
- 0010Probability, gently
- 0011The hardware underneath
- 0012Prompting, retrieval, fine-tuning
The anatomy of a prompt that works, the mistakes that quietly ruin most of them, and the habit of treating a prompt as a draft.
Roles, system messages, delimiters, chain-of-thought, structured output, and a checklist for debugging a prompt that will not behave.
Tree-of-thought, self-consistency, verification chains, meta-prompting, agent and tool-use prompting, injection defence, and token economics.
What actually changes between Claude, GPT, Gemini, open-weight models, and image models - and how to decide which one a task belongs to.
Using these tools for coursework without outsourcing the thinking. Studying, researching, checking your own work, and staying on the right side of academic integrity. Readable straight after Beginner.
Job-shaped playbooks: writing, code you have to maintain, analysis, research, and getting a team to prompt consistently. For the practitioner who arrived with a task rather than a syllabus.
What a prompt actually costs in energy, money and time - and why writing one good prompt instead of five vague ones is the most direct thing any individual can do about it.
Every lesson, the same seven parts
Consistency matters more than variety here. This is a course, not a blog, so you always know where the exercise is and where the sources are.
Read one and see- 01Learning objectivesTwo to four, stated before you start, so you can tell afterwards whether it worked.
- 02Plain-language explanationThe mechanism, in prose, with no notation you have not been given.
- 03Worked examplesReal prompts and the output they actually produce, weak beside deliberate.
- 04Common mistakesEach traced to the mechanism that causes it, so the fix generalises.
- 05Try it yourselfAn exercise on your own prompts, built to make the mechanism visible.
- 06Key takeawaysTerse. The things worth remembering a month from now.
- 07Further readingPrimary sources for every technical and historical claim above.
Where should you start?
Three questions, no account needed. The honest answer is often further back than people expect.
Start with lesson 00·1
Reading the lessons needs a free account. The whole curriculum, every lesson title and what each one covers, is visible before you sign up — so you can see exactly what you are getting first.
What an account is for
- Reading the lessons
- Marking them finished and picking up where you stopped
- Nothing else, and no cost, ever