Get AI income methods before they spread.
Free weekly intelligence newsletter.
The three-phase AI coding system breaks software development into distinct stages — ideulation, implementation, and evolution — with a specific workflow for each. Built around Claude Code and Jira, it turns a vague idea into production-ready code through a series of structured steps, each with its own slash commands and automated handoffs.
What Are the Three Phases?
Phase one is ideulation: you brain-dump everything you know about the product, and the system turns it into a structured PRD, then automatically creates Jira tickets for each feature. Phase two is implementation: each ticket runs through a PIV loop — plan, implement, validate — one at a time. Phase three is evolution: every bug fix, every improvement, and every new pattern gets codified back into the AI layer and checked into source control.
The system does not collapse as the project grows — it improves.
Most AI coding workflows work well for greenfield projects and degrade as complexity increases. This system inverts that. The evolution phase means each session adds to a foundation of proven patterns rather than starting fresh.
How Does the create-PRD Slash Command Work?
The /create-PRD command reads your brain dump — the conversation you have had with Claude about what you want to build — and produces a structured product requirements document. The PRD covers problem statement, user personas, core features, success metrics, and technical constraints.
It takes a messy, exploratory conversation and produces a document precise enough to generate implementable tickets from. Running it forces you to notice the gaps in your thinking before any code is written — which is the point.
How Does the Atlassian MCP Create Jira Tickets Automatically?
The Atlassian MCP connects Claude Code directly to Jira. The /create-stories command reads the PRD and creates one Jira ticket per feature — with a title, acceptance criteria, and story points estimated by Claude. Each ticket is created in the correct Jira project and assigned to the current sprint automatically.
Without the MCP, this process involves copying content from a document into a Jira form, field by field, for each feature. With the MCP, it runs in one command and takes under two minutes regardless of how many tickets are being created.
What Is the PIV Loop and How Does It Run Per Ticket?
PIV stands for Plan, Implement, Validate. For each Jira ticket, the loop runs three steps:
- Plan — Claude Code reads the ticket, asks any clarifying questions, and writes a short implementation plan. No code is touched in this step.
- Implement — Claude writes the code following the plan and the codebase's governance rules from CLAUDE.md
- Validate — tests run, screenshots are taken if it is a UI change, and Claude reviews the output against the ticket's acceptance criteria
The PIV loop does not move to the next ticket until the current one passes validation. This constraint prevents the common failure mode of stacking incomplete work.
How Does the Evolution Phase Work?
After each implemented feature, Cole runs a step that asks: did anything in this implementation reveal a better pattern? If yes, that pattern gets written into CLAUDE.md, the codebase's governance document. If a bug was fixed, a rule gets added to prevent the same class of bug from recurring.
The AI layer compounds. Each session leaves the system smarter than it found it.
Over a project's lifetime, the CLAUDE.md grows into a precise specification of how this codebase works — a document that any future Claude Code session can read to get up to speed instantly, regardless of how long the project has been running.
Frequently Asked Questions
Does this system require Jira specifically?
The Atlassian MCP targets Jira, but the same pattern works with GitHub Issues, Linear, or any project management tool that has an MCP connector or API. The key is having a structured ticket per feature that the PIV loop can reference during implementation. The tool matters less than the discipline of one ticket, one loop, one validated output.
How long does a typical PIV loop take per ticket?
For a small feature — adding a field, updating an API endpoint, writing a component — the PIV loop takes 15 to 30 minutes. For a complex feature involving multiple services or significant state management, it can take two to three hours. The planning step is where most of that time goes, and it is time well spent — implementation without a clear plan takes longer and produces more bugs.
What goes in the brain dump for create-PRD?
Everything you know about the product: who it is for, what problem it solves, what it definitely needs to do, what it definitely should not do, what similar tools exist, and any constraints on technology or timeline. It does not need to be organised — the point of the brain dump is to get everything out so the PRD can structure it. A 10-minute stream-of-consciousness explanation is usually enough.
Can this system be used for existing projects, not just new ones?
Yes. For an existing project, run /init first to generate a CLAUDE.md from the current codebase. Then write a PRD for the next set of features you want to add. The system does not require starting from scratch — it can be adopted incrementally, starting with the PIV loop for new tickets while existing code stays unchanged.
Watch Cole's full system walkthrough: The Three-Phase AI Coding System on YouTube
Get AI income methods before they spread.
Free weekly intelligence newsletter.
The 10 Claude Code Habits That Will Actually Change How You Work