Dev blog
Deep Keep was built almost entirely with an AI coding agent, across roughly 145 pull requests. This is the retrospective — not "look what the AI wrote," but the handful of decisions that let an AI build a coherent game: the authority boundary, a provable solvability guarantee, build divergence you can measure, and progression that unlocks options, never raw power.
The through-line of every post: give an AI agent invariants it can test, not just ones you state.
-
01 I built a roguelite mostly with an AI agent — here's what actually made it work
A dev-process retrospective: what it's really like to build a non-trivial game with an AI coding agent doing most of the typing, and the practices that kept ~145 pull requests coherent instead of collapsing into mush.
-
02 The constitution: a brief the agent reads before every session
How a short, blunt file of non-negotiable rules at the root of the repo kept an AI coding agent from drifting — the cheapest, highest-leverage practice in the whole project.
-
03 An agent has no memory: the docs that carry state between sessions
An AI agent starts every session with a blank slate. Two living documents — a backlog and a design log kept honest — became its external memory, so no session re-litigated a settled decision or re-built a finished feature.
-
04 Trust, but verify: making an AI check its own work
An LLM will assert that its code is correct, balanced, and safe with total fluency and no grounds. The load-bearing practice of the whole project was converting every such claim into a check that passes or fails the build.
-
05 When the agent was confidently wrong
Three times the AI agent shipped a plausible, well-argued change that was quietly broken — an infinite loop, a difficulty-flattening item, an unkillable build — and the gate that caught each one.
-
06 Directing the agent: what the human actually does all day
If the agent writes the code and the gates catch the disasters, what's left for the human? Setting intent, and reshaping vague or dangerous requests into the form that fits the project's rules — the highest-value work in AI-assisted development.
-
07 145 pull requests: shipping with an agent in small, verifiable steps
Why building with an AI agent came out as ~145 small pull requests instead of a few big ones — the cadence, the phased features, and the replay discipline that made each step verifiable. The series wrap-up.