Cancel Preloader

What You'll Learn

Diagnose why a prompt underperforms — and tell a prompting problem from a retrieval
tool
or model problem before you edit anything,Build an evaluation set that turns prompt changes from guesses into measurements
in about a day of work,Structure prompts with boundaries so untrusted document text can never be read as your instructions,Control length
format
tone and document size deliberately — and stop using the effort setting to do a job it cannot do,Write tool descriptions that get the right tool called
and stop paying for the wrong one,Fix agent scope creep
stalled overnight runs
test-gaming
and confident claims about unread code,Run an effort sweep and route by task type instead of inheriting one global setting for years,Migrate a prompt system to a new model with an eight-audit checklist instead of three weeks of archaeology,Keep conversation history append-only so you dont silently break caching and reasoning blocks,Recognise which of your instructions are compensations for a model that no longer exists — and delete them with evidence

Requirements

You call an LLM API already
in any language — this course is about the prompts
not the SDK,Comfortable reading JSON and using a terminal,No machine-learning background required
and no prior prompt-engineering theory assumed,Access to any current frontier model to run the labs (the techniques are not vendor-locked
though examples reference Claude models)

Description

This course contains the use of artificial intelligence.

Most prompting courses teach you to add things. This one will mostly teach you to take things out.

The craft inverted. The older discipline was compensation: models under-delivered, so you pushed. You wrote instructions in capital letters to make a tool fire. You said always double-check your work, because otherwise it wouldn't. You forced a JSON shape by starting the reply yourself. Every one of those existed to drag a reluctant model up to a standard.

Current models arrive at that standard on their own. They check their work, they reach for tools, they explore. So those same instructions now stack on top of behaviour that is already there — and you get an agent that verifies three times, spawns subagents for a task it could have finished itself, and writes nine paragraphs when you wanted two. In this course, four of the five largest measured effects come from deleting a line, not adding one.

You will work at one company for the whole course

Abstract prompting advice is close to useless, so there are no toy examples here. You join Vantel, a 400-person B2B invoicing company, on day one, and you stay there for all 34 lectures. Vantel runs three AI surfaces that between them cover essentially everything prompting has to handle: Ledger Assist (in-product chat), Invoice Intake (a document-extraction pipeline), and Cobalt (an internal coding agent that runs unattended for hours). You will meet the engineers, read their prompts, and watch their measurements.

Every claim in this course has a number attached

One emphatic tool instruction, written two model generations ago, deleted: 38% fewer tool calls, no change in task success.
One clause of reasoning added to a bank-details rule: a fraud control went from holding 12 of 41 probes to 39 of 41 — and became less likely to over-refuse.
Eleven format prohibitions replaced by one sentence describing the target.
A six-line bug fix that came back as 430 lines across 11 files, and the five specific clauses that cut diff sizes 70%.
An overnight agent that asked permission at 19:11 and waited 13 hours for an answer.

What makes this different from a blog post

Prompting advice ages badly, and most of it is published without measurement. This course is built the other way round: you build an evaluation set in Section 1, and every technique afterwards is something you can prove or disprove on your own system. Section 6 then teaches the thing that survives the next release — how to read a model behaviour guide, how to run an effort sweep, and the eight-audit migration checklist that took one team from three weeks to two days.

Cinematic, not slides

Zero bullet-point lectures. Every lecture is a motion-driven scene — animated architecture diagrams, live prompt-versus-response demos, terminal walkthroughs of real eval runs, and before/after comparisons with the metrics moving on screen.

Who this is for

Engineers, ML engineers and technical PMs who already call an LLM API and are tired of their prompt file becoming four hundred lines nobody dares touch. No ML background needed.

This is an independent, unofficial course. It is not affiliated with, endorsed by, or sponsored by Anthropic. Claude is a trademark of Anthropic, PBC. Always check the official documentation for current model behavior.

Who this course is for:

Engineers whose prompt file has grown past 200 lines and nobody remembers why half of it is there,ML engineers responsible for extraction or classification pipelines that need to be measurably correct,Anyone building agents that run unattended and occasionally do too much
too little
or the wrong thing,Technical product managers who need to judge whether an AI feature is actually ready to ship,Teams who have just been told to migrate to a new model and dont have a checklist