Capstone — The Seven Phases of AI Development
The modules so far explained what an agent is and how one is built, connected, structured and run in production. The question left is the one you face on Monday morning: how do you actually work with a coding agent to ship software you would put your name to? The best short answer we know is an article by Matt Pocock, and this capstone sends you to it.
Read it here: My 7 Phases Of AI Development — Matt Pocock, AI Hero, March 2026.
What the article covers
It describes a workflow in seven phases, and argues that the same sequence underlies most disciplined approaches to AI-assisted development, whatever tooling they use:
- Idea — decide what you are building, and let the agent question you until it is concrete.
- Research (optional) — when the work depends on an unfamiliar API or hard-to-find documentation, gather it once into a file in the repo so each fresh agent context does not rediscover it.
- Prototype (optional) — try several throwaway versions to settle the decisions that are a matter of taste before anything is specified.
- PRD — write down the end state: what users see and how it behaves, not how it is built.
- Implementation plan — break the PRD into tickets with explicit dependencies, which shows what can run in parallel.
- Execution — agents work the tickets; with good enough inputs, without you at the keyboard.
- QA — a human tests and reviews the result, which produces new tickets, and phases 5–7 repeat until the work is done.
Read the article itself for the reasoning behind each phase, and for the tools and skills the author uses at each step.
Why it is the capstone
Every phase is the primer’s material put to work. The research file in phase 2 is the same idea as the context files in AGENTS md vs SKILLS md: knowledge the agent cannot see, written down where it can. Breaking work into dependent tickets in phase 5 is planning, one of the patterns in Choosing the Right Agentic Design Pattern, and the ticket graph is a DAG. Phase 6 is the agent loop from Building an AI Agent from Scratch running unattended, and phase 7 is a reflection pass with a human as the critic.
The thread running through all seven is the one this primer keeps returning to: an agent is only as good as the context it is given. The phases are a discipline for producing that context — a sharpened idea, cached research, a settled prototype, a precise PRD, well-scoped tickets — before asking an agent to write the code.