Tools in the relay
goose DeepSeek Codex Claude

The agent relay

Every step has one of three owners: me, an agent, or an AI judge who reviews but never touches the code. The order is fixed — nothing skips ahead.

  1. 1
    Idea + Critique

    I bring the idea. An agent tries to break it.

    Every challenge it raises comes back with options, not just objections — a couple of ways to fix the problem, and I pick the one I want.

  2. 2
    PRD + Mock

    Written down, then drawn.

    The agent drafts a product requirements doc, I give feedback, it redrafts — round and round until I approve it. Then it mocks the screen so I can see the idea before anyone writes code.

  3. 3
    Set up once

    Rules first, then questions.

    A constitution sets what the game must always do and never do. The PRD gets split into small features in build order. Then one interview, in plain English — every open question gets asked once, never twice.

  4. 4
    Every feature

    The same relay, every time.

    Spec → Plan → Tasks → Build, for every feature, no exceptions. Each task is graded on its own and lands as one commit. Easy tasks go to a local model (DeepSeek); hard ones go to a stronger model (Claude, Codex).

  5. 5
    AI judge + evidence

    Nothing merges on "looks fine".

    An independent reviewer — one that can't touch the code itself — checks that every test passes and every requirement has proof before a spec, plan, or merge is approved. It only comes to me when it's genuinely unsure.

  6. 6
    Guardrails

    Rules enforced, not requested.

    Nothing merges before review passes — ever. Steps can't run out of order, the constitution can't be quietly rewritten, and every project record changes through one script, not a hand edit.

  7. 7
    Playtest

    Every few features, it's on my phone.

    I note what feels wrong in my own words — a bug, a missing piece, a change of mind, or just a feeling that something's off. The agents trace each note back to a cause and fix the spec before they touch the code again.

Merged = tests pass.

Done = I've played it.

Tests prove the logic. Only a human can tell you if a game feels good.

Curious about a specific part of the pipeline, or want to build something similar? Email support@oorukai.com.