Claude Opus 5.5, our AI teammate
Meet our AI teammate
The agents in our workshop need a brain. That brain is an AI model: a computer program that can read and write language and code.
Our studio uses Claude Opus 5.5, made by a company called Anthropic. Think of it as a very fast teammate that has read more code than any person ever could.
How did it learn?
An AI model learns by studying huge amounts of text and code, more than a person could read in many lifetimes. It spots patterns: which words and which lines of code usually go together.
When you ask a question, it doesn't look the answer up in a big answer book. It predicts a helpful answer, a little piece at a time, based on those patterns. A bit like how you can finish "once upon a..." without thinking.
Quick check
How does an AI model like Claude come up with an answer?
- It looks it up in a list of stored answers
- It predicts a helpful answer from patterns it learned
- A person secretly types every reply
Show the answer
It predicts a helpful answer from patterns it learned
The model learned patterns from lots of text and code. It uses them to predict a helpful answer, piece by piece. There is no answer list.
What it does for our studio
- Reads our codeIt can read our files to understand how Hyper Shards works.
- Explains itAsk "what does this function do?" and get a plain-words answer.
- Writes changesIt can add a new pest to Garden Defense or fix a bug in the ship.
- Hunts bugsShow it an error and it suggests where the problem might be.
Confidently wrong
Here's the tricky part. Because the model predicts answers, it can sometimes predict a wrong one, and it can sound totally sure while doing it.
It might say "I fixed the jump!" but really change the jump height instead. Or it might use a piece of code that doesn't exist in our game. Made-up answers like this are often called hallucinations.
Quick check
If Claude sounds very sure about an answer, the answer must be right.
- True
- False
Show the answer
False
Sounding sure and being right are different things. The model can be confidently wrong, so we always check its work.
We stay the boss
- Read itLook at the diff, the view that shows every line it removed and added. Does the change touch only what you asked for?
- Test itRun the game and play the part that changed.
- Ask it to explain"Explain this change in plain words." If the answer doesn't make sense, don't keep it.
Bolt says: Claude is a teammate, not the boss. The team decides what goes into our games, and that includes you.
Read the reply
Done! I fixed the bug where pests walk through walls. I changed walls.js, and I also tidied up shop.js while I was there.
What do you do first?
- Tell the team the bug is gone, Claude said "done"
- Play the level where pests walked through walls, then check that the shop still works
- Ask Claude for five new features straight away
Show the answer
Play the level where pests walked through walls, then check that the shop still works
Claude can be confidently wrong, so test the exact thing that was broken. It also touched shop.js, which you never asked for, so that needs checking too.
Safe and smart
One more rule: never type passwords, home addresses or other private information into a prompt. Our code and our game ideas are fine to share with our AI teammate. Personal secrets are not, because anything you type is sent to a computer outside your home and you can't take it back.
Bolt says: Claude is only as good as the instructions you give it. Next lesson: all the places you can work with Claude. After that, prompts that really work.
Fact-checked 9 October 2026
- Anthropic: Models overview (Claude Opus 5.5) platform.claude.com
- Anthropic: Reduce hallucinations platform.claude.com
- Google Antigravity: Models (Claude Opus 5.5) antigravity.google