It’s not that simple. You need to create a separate lib containing the rules for generating adaptive text, for instance. I needed to create a message system similar to Inform 7’s response framework, so that the standard responses can be replaced in a modular way. Then you can go about replacing the actual game text. Everything else works the same. There are one or two grammatical idiosyncrasies of French that I haven’t covered yet, though.
CLAUDE.md is a file that is injected at the top of Claude’s context, which tells it the ground rules for how you are working. For instance, mine lays down the rules for keeping the TODO file up to date, among other things.
Since there is so much noise about agentic assistance and LLM usage, generically referred as AI for simplicity, and you are making something with it, can you explain the process? The repository makes a reference to the transcript for the game Lurking Horror and includes a game transcript and a markdown file for it. How do those files factor into the process? Why are they needed for a simple game like CoD? Inquiring minds want to know.
edit - What I’m ultimately curious about, as a lifelong computer programmer - it truly is all I have ever done - is when do you cut the cord so to speak? The agentic process must at some point produce diminishing returns. Something with some sense of stability as a compute system where a small change is more disruptive than it is helpful. When does the consequences of making/asking for a change not outweigh the system and output that has been produced?
Whenever I work on something non-trivial, the process begins with discussing the work with the agent – the user-facing representation, the design, how it fits in with other pieces of the system (including those in the future – this is an important thing to ask the agent about). The agent might ask you about your preferences, but in general it’s going to be over-confident about its “ideas”. So it’s important to push back and propose alternate ideas. Plenty of times I’ve overridden the agent with far simpler and more elegant ideas than it came up with. As a lot of people have said, working with an AI coding agent is like working with an extremely well-read, inexperienced, very eager, junior developer. It’s worthwhile getting ideas from it, but it’s important to never blindly accept anything. Note that today’s models are fantastic coders. They don’t screw many things up.
The Lurking Horror transcript was for the agent to scan and find parser conventions we hadn’t covered yet. The markdown file contains its findings. We still need to work through that document.
While theoretically you’re probably right about diminishing returns, I haven’t hit that point yet. Very stable systems are always a risk to fiddle with, whether by a human or an agent, so I don’t think what you’re saying is exclusive to AI coding.
