AI for Developers masterclass

What a coding assistant can and cannot see, how to brief it, where AI breaks software, how to treat model output, and what never goes into a prompt.

1. What the assistant sees, and what it does not · 12 minThe learner can say what context a coding assistant has in a given session and what it is missing.
2. Give it the right context and ask for a plan · 14 minThe learner can choose the files, errors and constraints a change needs and write a brief that asks for a plan first.
3. Where AI breaks software · 15 minThe learner can name the usual ways AI-written code fails and spot them in a short function.
4. Treat model output as untrusted input · 15 minThe learner can handle model output in a product the way they handle user input: validate it, encode it and never execute it blindly.
5. What never goes into a prompt · 12 minThe learner can keep keys, credentials, customer data and restricted code out of prompts and knows where to find their company's rules.

Small changes, explaining code, reviewing AI diffs, tests first, debugging from evidence, true documentation, safe migrations and faster pull request reading.

1. One small, reviewable change at a time · 13 minThe learner can split a feature into small changes that can each be read, tested and undone.
2. Explain unfamiliar code, then check the explanation · 12 minThe learner can use an assistant to explain unfamiliar code and verify the explanation against the code itself.
3. Review an AI diff the way you would a colleague's · 15 minThe learner can review assistant-written code against a short list that catches the usual mistakes, including invented packages.
4. Write the test first, then the code · 15 minThe learner can have an assistant write tests from a specification, check those tests, and then have it write code to pass them.
5. Debug from logs and stack traces · 15 minThe learner can give an assistant the right evidence for a bug and test each hypothesis it offers instead of accepting the first fix.
6. Write documentation that is true · 13 minThe learner can generate documentation with an assistant and check every claim in it against the code.
7. Refactor and migrate safely · 15 minThe learner can plan a refactor or migration with an assistant as a series of steps that each keep the tests green.
8. Read a pull request faster, without approving on its word · 12 minThe learner can use an assistant to summarise a pull request and point out risks while keeping the approval decision their own.

The types of AI developer tool, choosing a model on quality, speed and cost, running models yourself, and keeping the bill under control with live prices.

1. The tool types and what each is for · 13 minThe learner can describe the main types of AI developer tool and match each to the work it suits.
2. Choose a model for the job · 15 minThe learner can choose a model for a task on quality, speed, context length and price, using live prices rather than remembered ones.
3. Running models yourself · 15 minThe learner can judge when a local or self-hosted model makes sense and estimate what the hardware costs from live GPU prices.
4. Control the cost of a feature · 15 minThe learner can estimate a feature's token bill and reduce it with caching, batching and a smaller model where it is good enough.

Context that fits the task, prompt templates, retrieval over your own code and docs, structured output, evaluation at volume and prompt injection.

1. Context that fits the task · 13 minThe learner can choose what to include in context as a code base grows, using summaries and maps of the repository.
2. Prompt templates for code work · 13 minThe learner can write reusable prompt templates for reviews, tests and commit messages and improve them from results.
3. Your own code and docs as context · 15 minThe learner can explain how retrieval over a code base and internal documents works and decide what to keep out of it.
4. Structured output you can trust · 15 minThe learner can get structured output from a model with a schema, validate it, and handle failures with retries or a clean refusal.
5. Evaluate output at volume · 15 minThe learner can build a small evaluation set for a model feature and run it on every change to catch regressions.
6. Prompt injection in code tools · 15 minThe learner can explain how untrusted text in issues, documents or web pages can steer a coding assistant and apply defences.

Team rules for AI-written code, an agent with a person approving each write, logs and audit for AI actions, and measuring honestly whether AI helped.

1. Team rules for AI-written code · 14 minThe learner can draft team rules for AI-written code covering labelling, review requirements and what needs a human sign-off.
2. An agent with a person in the loop · 16 minThe learner can design a small agent with limited tools, permissions and an approval step for anything that writes or spends.
3. Logs and audit for AI actions · 13 minThe learner can decide what an AI feature or agent should record so a mistake can be traced and explained.
4. Measure whether AI helped · 16 minThe learner can measure the effect of AI tools on a team with a baseline and honest metrics, and write a verdict no larger than the data.

An AI-written pull request with planted bugs, invented calls, unsafe shortcuts and unneeded changes. Find and explain them before they merge.

1. What to block in an AI pull request · 13 minThe learner can recognise the four kinds of problem planted in the role lab and say why each one should block a merge.
2. Role lab: review the AI's pull request · 20 minThe learner can review an AI-written pull request with planted problems, mark every line they would block and explain why.

Fix a real bug or build a small feature with AI, and prove it with tests, a log of what you checked and a specific judgement, reviewed by a person against a six-item rubric.

1. Plan your capstone · 13 minThe learner can choose a capstone task of the right size, set up the log they will keep, and decide what they will check.
2. Capstone: ship and defend a change · 20 minThe learner can deliver a working, tested change made with AI, with a log of what was checked and a specific judgement of where AI should not be trusted.

How the exam works and the four things that earn the certificate: every lesson, the exam, the role lab and the capstone.

1. The final check, the exam and the certificate · 8 minThe learner can say how the masterclass is assessed and which four things they need to complete to earn the certificate.