How I Build With AI Agents: Booker Blitz, gitmoire, and Hocus
I work at Dark Magic Studios, a small team that shipped gitmoire, a git client with built-in AI features. I own the build on Booker Blitz, my wrestling management sim, and I'm part of a new product, Hexboard, that we're keeping quiet for now. I use AI agents to write a lot of first drafts. I don't let them decide what lands. This is how that works: the products, the agent setup, and the review rules I won't give up.
Booker Blitz, heading into Early Access
Booker Blitz is a wrestling promotion management sim where you manage your roster and book your shows to get the best ratings. You can read more at bookerblitz.com. It's keyboard-centric, with switchable themes. You scout and train talent, give people a gimmick and a face or heel alignment, and book title divisions and brand splits. During a show you sit in Gorilla Position and make live calls while the simulation feed scrolls, which is the heart of the game. Wrestling schools graduate regen wrestlers into the roster. There's the DataStudio, the official editor, and you can bring in custom datapacks from existing games such as .db or .mdb files. It runs on Windows and Linux, and it is 100% playable on Steam Deck.
The stack is TypeScript, React, Electron and Tailwind, with n8n handling some of the automation around it, mostly for community and marketing purposes.
During playtesting I fixed bugs fast with an AI coding agent: reproduce the bug, let the agent draft a fix, read the diff, test it, push. The agent never pushes for me. If I can't explain why a change works, it doesn't go in. That loop kept testers unblocked while the game was being shaken out. Once people are paying, the same fixes move onto branches with pull requests, because speed is no longer the right tradeoff.
The next stage is getting the game through Steam and supporting it properly.
gitmoire, a Git client built around agents
gitmoire is my desktop Git client with an Agentic Development Environment (ADE). It's a hard fork of angkorGit, which is MIT licensed, and it's built on Rust and libgit2. Our fork is closed source for now, I was exploring ways to monetize it, but have ultimately decided that it should be remain free, so I'll be open sourcing it soon.
Part of the reason it exists is simple. GitKraken took almost 40 seconds to open on my machine. That's me timing it by hand, not a benchmark, but I open a Git client constantly and that wait got old. The other reason is the loop I work in now: a ticket becomes an isolated worktree, an agent writes there, and I review the diff before anything lands. I wanted that loop inside the Git client.
You can connect a Linear workspace and see your assigned issues in the sidebar. Issues from your Git forge get the same treatment. From an issue I can create a worktree on its own branch in one click, with the agent prompt already filled in and editable. I can also start a worktree from a pull request, and gitmoire fetches the PR branch without touching my current checkout.
The newer piece is orchestration. On the Agents tab I describe a task, or point at a Linear issue. An orchestrator agent asks follow-up questions and draws a battle plan. I approve it, and gitmoire deploys one agent per task into its own worktree. I pick the provider and model for the planner, the feature developer and the reviewer in Settings. The model menu lists the real models from Codex, Cursor, OpenCode and Antigravity. Agents coordinate through gitmoire agent commands, and done, waiting or blocked updates show up as badges and toasts. Agents live on their own tab, opened from Home or with Alt+A.
The quieter features are the ones I touch every hour. Repositories sit in named groups. The diff viewer can ignore whitespace, jump between changes, search a diff or a conflict, and open the file so I can fix it in place. Markdown previews in the editor. I save commit messages I type over and over as quick commits, press Alt+P, and gitmoire stages everything and commits. Those can be global or per repo, and I can undo a bad pick. A rebase or merge picks back up once the conflicts are cleared.
Hocus, the harness for hiring a cast
Hocus is the multi-agent harness we use across Dark Magic Studios. You point it at a project and it scaffolds a 13-persona cast from your stack, then hires them, so you get agents with a defined job instead of one general assistant trying to do everything.
The roles include a planner, an orchestrator, a reviewer, a feature developer, a founder, and the rest of the crew around them. You write a persona once as a SOUL.md file: a role, a voice, and a body of instructions. Hocus compiles that into the formats used by Claude Code, Codex, OpenCode, Cursor, Antigravity, Command Code and GitHub Copilot. hocus init sets the project up. hocus cast scans the repo for languages and frameworks, tailors each persona, and compiles files for the tools you use. An interactive command deck in the terminal lets you inspect personas, track plans, and talk to the cast.
If you want to try it:
pnpm i -g @darkmagicstudios/hocus
Then run hocus init in a repo. The project lives at github.com/dark-magic-studios/hocus and will be getting a nice update soon to improve memory for your projects using hocus 🚀
The crew behind the studio
The studio's outward-facing work, posts and announcement cards especially, goes through a crew of specialized AI agents. One keeps track of what is in flight. Others write briefs, check claims against the source material, design the cards, or dig into the code and the research.
The pipeline runs in the same order every time. Sources come first, then a brief, then copy, then design, and then it comes to me. Nothing ships without my OK: not a post, not a card, not a reply. I lock the copy and I make the product calls.
The canon check is the piece I'd recommend to anyone doing this. Agents love to make a product sound better than it is, so unsupported claims get cut, and posts may only cite locked changelogs. The crew also checks subreddit rules. One subreddit I'd otherwise post in bans game ads, so it gets skipped.
Where it goes wrong
A Reddit post of mine got called AI slop. Being accurate is not the same as sounding human, and that post didn't sound human enough. I rewrote the copy plainer: fewer flourishes, shorter sentences, more like how I'd explain something to a friend. I used to ask only whether a post was accurate. Now I also ask whether it sounds like me, and if it doesn't, it goes back.
That's why the approval gate exists. An agent can check a claim against a changelog. It can't tell whether a post will make a community roll its eyes, and it has no reputation to lose if it does. I do. When I'm deep in a fix, posts sit in the queue. I've accepted that. A slower studio that sounds like a person beats a faster one that sounds like a content farm.
What I review
Agents draft fixes, first passes of copy, project scaffolds and working branches from tickets. I read every diff before it merges, and I run the change before I trust it. I lock the wording and choose the tradeoffs, including the ones I later regret.
What I'd tell you if you're starting
If you're building something like this, start with the boring rules. They do more than any model upgrade.
- Check claims against a source you trust. Agents want to make a product sound better than it is. A locked changelog beats anyone's memory, mine included.
- Keep a person at the end of the line, and make sure that person reads what they approve. An approval you didn't read is just a delay.
- Keep the diffs small. If I can't review a change properly, the problem is the size of the change. I send it back and ask for smaller ones.
- Run the change before you trust it. A diff that reads well and a change that works are two different things.
- Give each agent a job. A planner, a reviewer and a developer with their own instructions beat one general assistant trying to be all three.
None of this is clever. It's what a careful team does anyway, with the difference that some of the team now works fast and never gets tired. Speed is the easy part. Judgment is what you have to protect.
What I got wrong, and what changed
The AI-slop comment taught me the most. I had a process that checked whether a post was true, and I assumed that was enough. It wasn't. A post can be accurate and still read like it came off an assembly line, and people can tell. So the question changed from "is this accurate?" to "is this accurate, and does it sound like me?" If the answer to the second half is no, it goes back, however good the facts are.
The other lesson is about patience. The approval gate makes me slower. Posts wait while I'm in the middle of a fix, and some go out later than they could have. I've stopped treating that as a flaw. The queue is the cost of having someone with a name and a reputation sign off on everything.
What's next
Booker Blitz is heading into Early Access, and I'm staying on it. That means the Steam launch, the bug reports that come after it, and supporting the game properly once people have paid for it. That's also when the review rules get stricter: branches, pull requests, and no more shipping fixes straight to testers.
For the studio, the attention goes to gitmoire and Hocus. gitmoire is going open source, so anyone can use it, read it and send fixes. Hocus is getting an update to improve memory for your projects, so agents stop starting every session from zero. And there's Hexboard, which I can't say much about yet.
Work with me
I'm looking for my next role. I want to build products, and the systems that let a small team ship them without losing the plot: the tooling, the review habits, and the agent setup that keeps speed from turning into mess. I like working close to the product, owning a build from first draft to release, and I care about the unglamorous parts, like the approval gates and the smaller diffs, as much as the features.
If that sounds useful, start at davidmorais.com, then look at bookerblitz.com, darkmagicstudios.com, and GitHub at github.com/dark-magic-studios. I'm happy to talk through any of it.
David Morais



Top comments (0)