Skip to content
AI Code that Works
Get the Starter Harness

Start here

The five-phase system for building real software with AI.

Read this once. Then start. No prior code required.

Get the Starter HarnessRead the Framework

Phase by phase

What each phase actually closes.

Each phase exists because prompting alone leaves a specific class of failure open. Here's what each one closes, when the audience felt the failure, and how the phase works in practice.

  1. 01Context

    Give AI the full picture. Project, audience, constraints, conventions, history. Without context, AI guesses.

    The failure it solves

    AI hallucinates because it doesn't know what you're building.

    You felt this when

    You re-explained the project five times in one session and AI still missed it.

    How it works

    • Capture the full project brief, audience, and constraints
    • Document the conventions that govern the code
    • Surface the history — decisions, dead ends, what NOT to do
    • Make context discoverable for the AI on every call
  2. 02Rules

    Set the standards it must follow. Code style, schema, naming, security. Without rules, AI drifts.

    The failure it solves

    Code that drifts every time a model updates.

    You felt this when

    You shipped on Tuesday. By Friday, the same prompt gave different code. By Sunday, nothing worked.

    How it works

    • Lock the rules in version-controlled spec files
    • Tell the AI to read them every time, before generating
    • Run the rules as automated checks in CI
    • Reject any output that violates them — no exceptions
  3. 03Documentation

    Keep the project organized. Specs, decisions, change logs. Without docs, AI re-invents.

    The failure it solves

    Knowledge that lives in someone's head doesn't survive the project.

    You felt this when

    You came back to a project three weeks later and couldn't remember why anything was structured the way it was.

    How it works

    • Document specs before you build — not after
    • Capture every architectural decision as it's made
    • Keep change logs at the module level
    • Treat documentation as code — versioned, reviewed, tested
  4. 04Verification

    Test, review, improve. Don't trust the first answer. Without verification, demos masquerade as products.

    The failure it solves

    "It works on my prompt" is the new "it works on my machine."

    You felt this when

    You demoed it to a friend, it broke, you couldn't reproduce the bug to fix it.

    How it works

    • Tests before merge — written by AI, reviewed by you
    • Manual verification on every user-facing change
    • Real data, real users, real edge cases
    • Reject the demo if it doesn't pass the bar
  5. 05Feedback Loops

    Catch drift early. Every commit teaches you something. Without loops, the same bugs ship twice.

    The failure it solves

    Drift you don't catch becomes the bug your users find.

    You felt this when

    A customer reported the bug you fixed two months ago — except now it's back.

    How it works

    • Every commit is a chance to learn
    • Observability on production from day one
    • Capture every error, every miss, every surprise
    • Feed the loop back into Context, Rules, Documentation

The takeaway

This is not prompting. It is a build system.

The five phases work together. Skip one, the others don't compensate. Run all five, and you ship.

Get the Starter HarnessJoin the Free Course on Skool

We use cookies

We use cookies and similar technologies to keep this site working, measure how it performs, and (with your permission) personalise ads. Choose your preferences below.