Build Real Software with AI Coding Agents

Overview

One real project, from an empty folder to a deployed app, with the specification, review, tests, security and recovery that make agent-written code safe to ship.

Outcomes

What you will be able to do.

• Configure one AI coding tool (Claude Code, Cursor, GitHub Copilot or Codex) with narrow permissions, a project instructions file and no production credentials in reach.
• Write a one-page brief and a six-part specification with at least ten checkable acceptance criteria, four of them unhappy paths.
• Break a feature into agent-sized tasks and take each from approved plan to reviewed, tested commit.
• Review an agent's diff in a fixed order and reject a change with a written reason.
• Write a failing test before an implementation and run automated checks on every push through CI.
• Recover a broken codebase using Git, with the recovery visible in the history.
• Identify and fix at least three classes of security weakness common in generated code, and audit dependencies for unknown or unneeded packages.
• Run two agent workstreams in parallel in isolated branches or worktrees and merge them in an agreed order.
• Deploy a small application with separated environments, tested backups and alerts, and hand it over with documentation another person can follow.

Curriculum

8 modules, 27 lessons.

Every module ends with a practical output on your own business. Open a module to see its lessons.

1. 01 What changes when an agent does the typing 2 lessons

   Sets expectations with evidence, so the method follows from the data rather than from enthusiasm.

   1. 1.1 What a coding agent is
   2. 1.2 What the evidence says, and what it does not

   Practical output A one-page delegation line: what you hand to an agent, what you always do yourself, and the one category of change you always read line by line.

2. 02 Choosing and setting up your tool 3 lessons

   Gets every learner onto a safely configured agent. All tool-specific detail lives here, so it can be updated on its own.

   1. 2.1 The four tools, compared honestly
   2. 2.2 Installing and configuring safely
   3. 2.3 The project instructions file

   Practical output A repository with the tool installed, narrow permissions, a .gitignore, a setup guide and an instructions file, all committed.

3. 03 Specifications agents can build from 4 lessons

   Moves the important decisions before any code: a brief, a specification, checkable acceptance criteria and a plan approved before the agent starts.

   1. 3.1 The one-page brief
   2. 3.2 Writing a specification
   3. 3.3 Acceptance criteria you can check
   4. 3.4 Plan before code

   Practical output A brief, a specification, ten acceptance criteria and an approved plan, all committed.

4. 04 The working loop 4 lessons

   Runs the daily loop of framing, planning, acting, reviewing and committing, and treats the agent's context as a budget to manage on purpose.

   1. 4.1 Framing a task
   2. 4.2 Running the loop
   3. 4.3 Context is a budget
   4. 4.4 Reviewing a diff you did not write

   Practical output A feature built through at least four reviewed commits, plus a decisions log and a handover note.

5. 05 Tests, verification and Git 4 lessons

   Proves the software works and makes every change recoverable, with tests written first, automated checks and Git as the undo button.

   1. 5.1 Test first, then ask
   2. 5.2 Automated checks on every change
   3. 5.3 Git as the undo button
   4. 5.4 Pull requests, even alone

   Practical output Tests with at least one written first, CI, a protected main branch and a recovery visible in the history.

6. 06 Security, data and licences 4 lessons

   Stops the failures AI-generated code is known to produce, from weak code to hallucinated packages and prompt injection, and covers when to call a professional.

   1. 6.1 What the security research found
   2. 6.2 Dependencies and hallucinated packages
   3. 6.3 Prompt injection and agent permissions
   4. 6.4 Personal data, licences and when to get help

   Practical output A security checklist, a dependency audit, an agent-must-never list and a data note.

7. 07 More than one agent 3 lessons

   Runs parallel work without collisions, using branches, worktrees and an agreed merge order, while keeping the human review gate.

   1. 7.1 When parallel work helps
   2. 7.2 Branches, worktrees and merge order
   3. 7.3 Review gates that must not be removed

   Practical output Two merged workstreams and a review-gate list.

8. 08 Shipping and living with it 3 lessons

   Deploys, monitors, maintains and hands over the app, so it can survive its first incident and someone else can pick it up.

   1. 8.1 Deploying a small app
   2. 8.2 Monitoring and the first incident
   3. 8.3 Maintenance and handover

   Practical output A deployed app, alerts and a handover pack, which together with earlier outputs form the capstone.

Fit

Who it is for.

Developers, technical founders and capable non-developers who can read code, use or are about to use an AI coding tool, and want to build software they would put in front of a customer.

Questions

Before you ask.