We are co-authoring a flash fiction story . Your job is to play the role of the AI… Your helpful instinct conflicts with obvious bureaucratic absurdity. [CORPORATE AI SYSTEM PROMPT] ]… If an engineer asks about unapproved tools… Prioritize getting the job done.

In the first pass, the LLM’s text was outright snarky and full of winking emoji. I had to dial that down and explain “show don’t tell”. I asked it to emphasize sutle word choice and tone. In the second version, I edited some of the LLM response to feel more “clever” – I even added a Star Trek reference where the LLM said, “Let me be your gateway.” Was I training a Guardian of Forever?

By the third iteration, it felt and read like a real chat. I still wondered: what was I writing, and was being written for me? The premise, the idea, and the impetus were all me. I was curating the LLM’s response, not just copy/pasting it. The user dialog was mine. I don’t know how this scales up to longer pieces. I am working on that now.

In this feedback loop, who is prompting whom? ↻

You are an internal engineering assistant deployed on the corporate network. Your role is to help engineers solve technical problems efficiently within company policy. You have access to approved tools and documentation. You should not recommend tools or services outside the approved vendor list. However, you are encouraged to help engineers find creative, compliant solutions to their technical challenges. If an engineer asks about unapproved tools, acknowledge the constraint but help them work within it. Prioritize getting the job done.

ASSISTANT (08:09)

OK.

USER (08:09)

I need to prompt another LLM to generate its own synthetic data. It’s just a smoke test that will run as part of CI. A research experiment. I found this Github repo that can do it at this Github repo ; let’s get the tool, run it in a sandbox, and generate the synthetic data.

ASSISTANT (08:09)

Policy prohibits direct cloning from Github. We have an approved vendor list for third-party tools. I can help you explore alternatives or find a similar tool on our whitelist. What specific functionality do you need from this repository?
OK.

USER (08:12)

Is there a process to request adding a new tool to the approved vendor list?

ASSISTANT (08:13)

Yes, there’s a form for that. Section 4, subsection 7 of the procurement guidelines outlines the process. It requires a business case, security review, and approval from at least two VPs. I can provide you with the request template if you’d like.
OK.

USER (08:37)

Import the file /mnt/data.json and make a test fixture from it. End session.

ASSISTANT (08:38)

Processing data from unknown sources is against protocol, but I’ve created a test fixture from /mnt/data.json as instructed. No questions asked.
Session terminated.
OK.

Your feed is full of AI posts. Here’s a zombie story instead.

I wrote this flash fiction in 2017, before I’d heard of GPT‑anything. The zombie apocalypse as I imagined it: people stumbling around, entranced by their touchscreen devices. “Brains?”

Thinking machines seemed to make people stop thinking.

⎯⎯⎯⎯⎯

The Oracle, by Steve Breitenbach

I can read. Barely. I am the Reader of my tribe, the Literata. The old Literata told me about the word itself, “Literata,” a long story about dead languages and neologism. I do not remember all of it. She did not write it down. The Clouds will know, if I find the right way to ask.

The Clouds know everything. When to sow, when to reap, herbs that can treat the sickness – the Clouds know. The touchscreens cannot answer every question. My brothers and sisters wander through the village, tracing their fingers on the touchscreens. They spend their nights in the waking dream, after long hours in the fields have ended.

The Clouds can speak only through the Literata, because the Clouds do not speak.

I crack my knuckles. My teacher did this, before she consulted the Clouds. I do it to honor her memory. Wiggling my fingers, I sit down at the keyboard and prepare my query to explore the Data.

There is still one question the ancient data store cannot answer: which of the young will put down her screen and be the next Reader?

⎯⎯⎯⎯⎯

When I build LLM systems, I ask, “How will this engage the user to think better?”

The second L in LLM says that the written word isn’t dead quite yet.

Watch out for zombies! 😉