Make your own
AI Modding Prompts: What to Tell Your Agent, and What It Costs
Updated October 7, 2026
There are no magic AI modding prompts. Ask the creators behind the mashups on our wall and you get two camps: some send short, rambling messages and let the agent figure out the route, others write detailed briefs with every bit of context. Both camps agree on more than they disagree. This guide collects AI modding prompts you can adapt, the workflow that keeps a long project from going in circles, and an honest look at what the tools cost.
What both camps agree on
- Describe the problem, not the code. “The door doesn’t open when I press E next to it, and the log says this” beats “fix the doors.”
- Give the goal and the evidence. What you wanted, what happened, and the log from both games if there are two.
- Point at an example project. Linking SkyCraft for a passthrough mod, or the MW2 rewrite for a rebuild, does more than a paragraph of explanation.
- Watch for tunnel vision. If the agent keeps trying the same wrong approach, say so and point it somewhere else.
Five AI modding prompts to adapt
Copy these as a starting point and fill in your own games. Each one comes from a stage where creators say projects most often stall.
Prompt 1: recon before anything else
Use this before you commit to an idea. It costs little and often saves days.
I want to [your idea, e.g. put Minecraft’s building inside Valheim]. Both games are installed on this PC. Before writing any code, investigate and write a plan: each game’s engine and exact version, which mod loaders, script extenders or APIs exist for them, whether either has online play or anti-cheat (we don’t touch those), where the save files are, any existing mods that already do something similar, and the smallest thing we could build first to prove this works.
Our mod loaders guide tells you in advance roughly what the answer will be for popular games.
Prompt 2: the smallest working slice
Once the plan looks sensible:
Build only the first step from the plan. When you’re done, tell me exactly how to run it and what I should see on screen. Back up my saves before changing anything.
For a passthrough mod, the first step is usually “get your plugin loading inside the host game and writing one line to a log,” then “get both sides to agree on a protocol version.” See how passthrough mods work.
Prompt 3: a playtest report
Most of your time goes here. A good report is specific and includes evidence:
I tested build 3. Walking works. When I place a block on a slope it floats about half a block above the ground, every time. Jumping off a ledge into water kills me as if it were solid ground. Here are both logs from that session: [paste]. Fix the floating blocks first.
One problem at a time, with the steps to reproduce it, gets far better results than a list of everything that feels off.
Prompt 4: when the chat gets stuck
Long chats get slower, more expensive and more confused. Creators use a handoff:
Write a detailed STATUS.md: what this project is, how it’s built, what works, what’s broken, the exact problem we’re stuck on, everything we’ve tried, and the next three things you would try.
Then open a fresh chat, give it that file, and ask it to read the file and suggest how to troubleshoot. A fresh chat with a good handoff often solves in minutes what the old one circled for an hour.
Prompt 5: getting ready to publish
Prepare this for players. Make sure the repository contains no game files or decompiled game code, write a README with exact game versions, requirements, install steps, known issues and credits for every project we built on, and package a release that reads game files from the player’s own install.
Then read how publishers have reacted so far before you post it.
Requests to avoid
Some requests waste time or cross lines, and a good agent will push back on them anyway:
- Bypassing DRM or anti-cheat. Off limits, and agents generally refuse. Stick to single-player and offline modes.
- Asking it to download a game or its files from the web. You need your own installed copy; the agent reads from it.
- “Build the whole thing” in one message. Huge first requests produce sprawling code that’s hard to test. Ask for one working slice.
- Pasting game files into the chat. Point the agent at the install folder on your PC instead.
- Switching goals mid-chat. Finish or park one feature before starting another; mixed goals confuse the agent’s plan.
- Vague complaints without logs. “It’s broken” sends the agent guessing. Say what you did, what happened and paste the log.
Keep the project’s memory on paper
Agents forget between sessions. Three files fix that:
- A rules file (
AGENTS.mdorCLAUDE.md): what the agent must always or never do. For example: never commit game files, always back up saves, always say how to test a change. - A development log: what changed, how it was tested, what is still broken.
- A design doc: how the project works, in plain language. SkyCraft’s is widely cited as the best example in this space.
Commit to Git after every working step, so you can undo a change that broke things.
Let the agent test what it can
Modern agents can launch the game, take screenshots and read logs. Ask them to verify their own work before handing it to you, and save your playtesting for the thing they can’t judge: whether it feels right. Several creators say that is where most of their hours went, even when the AI wrote nearly all the code.
What it costs
Prices change fast, so treat these as a snapshot from early October 2026 and check the providers’ own pages before you pay.
- The AI agent. Claude Code needs a paid Claude plan: Pro at $20 a month is the usual starting point, with Max plans at $100 and $200 for heavier use. Community guides recommend maxing out a lower plan before upgrading, and note that the higher tiers multiply your short-term allowance more than your weekly one. Cheaper and free options exist, including OpenCode with free or low-cost models and pay-per-use APIs, but free tiers tend to hit limits fast on a real project.
- Generated assets. Tools like fal bill per image, model or sound. Set a spending limit before you start.
- The games. You need to own every game involved. That is not optional, legally or technically.
Two habits cut costs more than switching models: start fresh chats with a handoff file instead of dragging one huge conversation along, and use a cheaper model for routine work, saving the strongest one for the part where you’re stuck.
Where to go next
- Universal Modder packages this whole workflow for several agents.
- How to merge games with AI walks through a first project end to end.
- The wall is full of open-source projects to read before you write a line.