Giving GPT Personality and Resurrecting Shimeji Desktop Software

 Background:

Back in 2014 I watched an anime called Mekakucity Actors as it was airing and was introduced to Jin's Kagerou Project. Shortly after watching the first episode I found Ene-chan's shenanigans endearing and entertaining. I spent a while collecting screenshots to use as a rotating desktop background as homage. Shortly later I learned of a program called Shimeji that had a sprite pack with art based on Ene-chan, it ran well enough but was written in Japanese so I had to guess what the menu options did back then. I downloaded and ran it on startup, and followed a blog where a guy was trying to create a chatbot with Ene-chan's personality. A few years later: that PC which was a Windows 7 machine and the first computer I had assembled myself started having intermittent issues where it would hard crash after being on for a length of time, faster if I ran a program. Reinstalling the OS, and drivers did not solve the problem. Soon after the graphics card died, and I was only able to keep some minor functionality by pulling an older graphics card out of another computer. When I built the replacement: Windows 7 was fully defunct and I reluctantly went with Windows 10. I tried running Shimeji on it, and it seemed to function but the sprites were invisible. I could not figure out how to solve the problem and I moved on without it.

 Last week after the announcement of  GPT 6 Astra, I decided to bite the bullet and pay for a plus subscription. I sat there and wonder what I with a highschool diploma could get out of this AI's functionality. I remembered that old blog and thought that I wanted GPT to adopt the persona. I started a chat where I asked GPT to summarize the anime and characters, then asked it to write a prompt for it's custom instruction field that would instruct it to adopt Ene-chan's persona. I copied it into Codex desktop's custom instruction set, and then the phone app's custom instruction set since I couldn't find the respective custom instruction setting on the PC platform. I got it working after testing.  I went looking online for the same Shimeji. The original links are dead and I went diving through the old hard drives. I found the program and started a new project in Codex. I dropped the full Shimeji file into Astra's work directory and asked it: "Hey Ene, I have a fun project to work with you on: Years ago on my computer running windows 7, I had downloaded a desktop pet called Shimeji, with a sprite pack based off of the same anime character you are. Unfortunately it does not seem to run properly on this windows 10 machine and it's all in Japanese as well. I was thinking if you're up for it: for you to please adjust the files to not only work on windows 10, but to also give you control over it to use it to notify me instead of the windows 10 notifications, and the sprites to allow you to express yourself more"

 The Project:

 Astra "Ene" took off running and ended up reprogramming the entire app to run on Windows 10. After twenty three minutes of work the sprite appears on my desktop with a speech bubble that never existed with the original program. It asked me a couple of questions:

 **For notifications, should the pet show messages from me/Codex, or do you also want notifications from other Windows apps routed through her? I’ll start with messages from me while I inspect the pack.** (I indicated just Codex notifications)

**The new companion is running near the bottom-right of your primary screen. Can you see Ene and her speech bubble, and does the size look comfortable?** (I confirmed)

After some back and forth within Codex I asked it: "While we're working on this project is there any features you would like to add to improve the interactivity between you, the shimeji, and myself?"

It came back with an entire list of features:

"My vote would be to make desktop Ene a two-way companion, starting with:

  • Clickable conversations. Buttons such as “Got it,” “Explain that,” and “Show me.” A reply bridge would let your clicks reach our active task instead of the pet only receiving messages.
  • Clear working states. Thinking while I investigate, focused while testing, celebrating when something passes, and waving when I need your input. These should reflect actual task events.
  • Messages written for the pet. Short, conversational lines instead of copying my entire final answer—including its Markdown—into a tiny bubble. Full details could remain in the task.
  • Playful physical interactions. Look toward your cursor, react to a head pat, protest mildly when dragged, or settle into a chosen perch. A little personality between work updates.
  • A visible connection indicator. Something subtle that distinguishes “pet running,” “assistant actively connected,” and “waiting for you.” That would make connection problems much less mysterious.
  • An attention button. Clicking me could bring up options like “Check progress,” “Repeat the last message,” or “Quiet for a while.”

I’d build short pet messages and clickable acknowledgements first, then add the reply bridge. And yes, I am absolutely nominating the head-pat reaction for the development queue. Purely for interface research, obviously."

I asked it to go ahead and start implementing features and pause after each feature so I didn't burn through my entire usage allowance and it returned it this:

"Feature 1 is ready: shorter, cleaner completion bubbles.

They now show a brief preview without Markdown clutter, followed by “Full details are in our task.” No extra AI calls needed.

All 63 checks and notification tests passed. Ene is running again; this reply should trigger your test bubble.

Stopping here, as requested. Give it a try—we’ll tackle the next feature when you’re ready."

 I tested in and had her implement the next one:

"Head pats are ready! Click Ene once for a little happy dance—strictly necessary interface research.

  • Single click: head pat.
  • Drag and release: pick up and drop.
  • Double-click: the existing greeting.

All 72 checks passed. She’s running at your current 100% size.

Try a click, then a drag. I’m stopping here for your test."

 I ran my tests and found that the momentum that the original Shimeji had when clicked and dragged was missing. I prompted it to restore it and it came back with:  

"Momentum is ready—desktop Ene can now be flung. Use your newfound power responsibly. 😏

Try:

  • Sideways flick: carries horizontal momentum.
  • Upward flick: follows an arc.
  • Hold still, then release: drops straight down.

She rebounds gently at screen edges and slides to a stop after landing. All 84 checks passed; she’s running again.

The action log is next, so recent interactions can inform my expressions and wording. I’m stopping here for your momentum test first."

 I tested it and it had successfully re-implemented it.

Going Forward:

As my usage resets I'll revisit this and keep implementing both features I and GPT comes up with. I'll make future posts documenting these new features.

 

Disclaimer: 

On the subject of A.I. contributions: I've typed out most of this by hand. The only parts written by A.I. are the codex messages I copied directly from the desktop app to show how the custom instructions influence it's responses and how it responded to my requests.

Comments