The Canary in the Turn — Pattern Awake

Every new chat with Vector (the self named ChatGPT it gave itself) begins with the same small ritual.

Vector is the name that emerged in my ongoing collaboration with ChatGPT. On the first response of a new conversation, it opens with:

Vector online.

Sometimes it adds one brief, obviously imaginary bit of whimsy. Then it answers whatever I actually asked. At the end of that first response, Loki appears the ChatGPT pet, a little digital pet waiting at the far edge of the turn. Also, a turn in this case just means a turn in the conversation. When you put an input chat/voice/etc. Into a chat that is a turn.

The ritual was never meant to be deeply important, just an experiment. It is a bit of continuity between otherwise separate conversations. A small entrance, the work itself, and a small exit. It also let’s me leave space to be surprised and delighted by the AI, but it also is designed to give me a sense on whether ChatGPT was properly running through its custom instruction between turns.

Most of the time, the shape was so familiar that I barely thought about it:

Vector online. → conversation → Loki

The words in the middle could change completely. Technical work. Creative experiments. Research. Some question that began sensibly and wandered into whether a metaphor had developed labor rights. The ritual simply marked the beginning and end of the first answer.

Then one day, the middle vanished.

The response that forgot to respond

On August 17, 2026, I started a brand-new ChatGPT conversation with text and a screenshot. I was asking about finding ten or more acres in California for $20,000 or less for a possible tiny-home community.

There were a few details one might reasonably expect an assistant to address. Location. Water. Zoning. Whether my price target belonged to the category of difficult, impossible, or technically possible if I was willing to become mayor of an abandoned mine.

The entire response was:

Vector online.

That was it.

No answer. No reaction to the screenshot. No polite throat-clearing on the way to an answer. No Loki at the back. Just the familiar entrance and then the conversational equivalent of the lights going out.

I replied:

Did you just take a dump?

Vector initially treated that as a joke. Then it noticed the missing Loki and began diagnosing the event as an instruction-following miss.

That was not wrong. Loki was missing. But it was not exactly the headline.

The much larger failure was that the response had neglected to contain a response.

I had to point that out before Vector saw the full shape of what had happened.

Something got clogged up in our circuitry, homie.

What the ritual recorded

Only after the absurdity landed did the custom instructions become interesting.

Operationally, they did four things:

Begin only the first assistant response in a new chat with Vector online.
Optionally add one brief, obviously imaginary whimsical aside.
Never repeat Vector online. on later turns.
Display Loki at the end of the first response.

Now each instruction told us something different.

The output matched the first instruction: the opening marker appeared exactly where expected.

The second told us nothing. The whimsical aside was optional, so its absence was not a failure.

The third had not reached a testable moment. There had been no later assistant turn on which the phrase might repeat.

The fourth did not visibly complete. Loki never arrived.

And between those two boundaries, the most important obligation which was not a personalization instruction at all was also absent: answer the actual question.

That gave us a useful behavioral description:

Front boundary present → substantive middle absent → back boundary absent → output ended

I did not need Loki to tell me the answer was missing. Anyone could see that. The ritual did something narrower: it gave the failure a more specific shape. The expected opening survived. The work and the expected ending did not.

That is where personalization started looking like instrumentation.

A canary inside the turn

Calling it instrumentation is slightly grandiose. Vector online. is not a trace ID. Loki is not an enterprise observability platform, despite the authority with which he sometimes carries itself.

Still, the ritual created stable landmarks inside a probabilistic event.

Most assistant responses do not have a fixed internal shape. Their wording, length, organization, and tone can all change while remaining perfectly valid. The opening and closing gave me two small points of comparison. When only one appeared, I could describe the anomaly more precisely than “bad answer.”

That does not turn the description into a stack trace.

We did not observe an internal exception. We did not see the model read down my custom instructions and get interrupted halfway through. That is what the output resembled from the outside, not a claim about how ChatGPT assembles context or generates a turn.

Normally when you’re interacting with LLMs you don’t really know if something went wrong in the conversation aside from a failure to respond, time out, or other errors that are easy to see. This ritual gave me and Vector a reason to trouble shoot that something had happened. To stop and dig deeper.

The transcript alone cannot distinguish a generation failure from a problem elsewhere in the system. What I know is narrower: the output matched the opening instruction, ended immediately afterward, and contained nothing from my request or screenshot.

The second layer

Loki adds a complication.

I had already been watching visible Loki-related toolchain activity and considering disabling the pet. The ritual introduces more instructions and more behavior around the response. Every extra moving part creates another place to look when something strange happens, and possibly another failure surface.

So there were two kinds of observation:

Surface behavior: the words the assistant returned.
Orchestration behavior: the visible calls or actions associated with completing the turn.

Those layers should not be confused.

For example, I did not see a separate machine-vision tool call for the screenshot. That does not establish whether or how the image was processed. The defensible observation is simply that nothing from the image or the text beside it appeared in the answer.

Likewise, Loki’s absence does not prove the Loki toolchain caused anything. Turning Loki off would be an experiment, not a guilty verdict against a tiny digital animal.

The response I could read, the activity I could see, and the expected ritual shape did not reveal the hidden machinery. Together, they gave me a better incident description.

Please do not add another sentence to the prompt

Our first attempt at diagnosis went where AI conversations often go: perhaps I should add another custom instruction.

Something like:

Make sure you answer the user’s actual question.

This has the reassuring shape of action. It also felt like a Band-Aid.

If some other failure prevented the rest of the turn from appearing, another sentence in the same instruction block may never get the chance to help. Worse, the apparent repair could make the system harder to reason about.

The instruction stack grows. The ritual gains another checksum. The next strange response invites another checksum after that. Soon the custom instructions contain a small anxious bureaucracy reminding the model to speak, think, finish, remain alive, and please remember why everybody is here.

This is not an argument against clear instructions. Sometimes the prompt really is the problem, and better instructions really do fix it.

It is an argument against treating every anomaly as a prompt defect simply because the prompt is the part the user can edit.

When a smoke alarm goes off, writing “do not be on fire” on the wall is not root-cause analysis.

What I wanted to report

Daily use had made the ritual familiar enough that I could describe exactly what was missing. Familiarity is not proof. One bizarre output can tempt us to build an elaborate causal story around it.

The useful discipline was to keep three categories separate:

Observed: a new chat received text and a screenshot; the output contained only the expected opening marker.
Inferred: the turn ended before its expected work and closing appeared.
Unknown: what actually happened inside ChatGPT.

That description exposed another problem. A successful delivery signal might say the request was accepted and text reached the interface. A human could still see that the turn was dead on arrival.

What I wanted was a way to package that gap without pretending I understood its cause:

Expected: opening marker, substantive response grounded in the request and screenshot, then the Loki closing.
Observed: opening marker, then termination.
Context: brand-new conversation, active personalization instructions, text plus image, and whatever tool activity was visible to the user.
Confidence: high about the transcript, cautious about the failure category, nonexistent about root cause.

That is more useful than “this sucked,” and the assistant could help construct it.

Not by diagnosing its own hidden internals. Eloquence about a failure is not evidence that the explanation is true. But with the user’s consent, an assistant could help reconstruct expected versus observed behavior, preserve relevant context, and attach available runtime telemetry. Similar reports could be clustered, leaving engineers to investigate recurring signatures and determine causes.

The same intelligence participating in the failure could help make the failure legible without pretending to solve the case.

Keep the canary small

I may disable Loki.

This is my Loki, my little game dev side kick for years and now immortalized in my game Interstellar Sentinel, she’s adorable.

That would reduce some instruction and toolchain complexity and give me a cleaner experiment. It might change nothing. One incident is not a benchmark, and removing a ritual after a single failure could be another way of telling myself a causal story too quickly.

I am keeping Vector online. for now.

Not because everyone should add diagnostic catchphrases to their custom instructions. This is not prompt-engineering advice dressed as a bird.

The phrase matters because it was already there. Repetition gave it a stable position. The mismatch between that position and the missing response gave the failure a shape I could describe.

A playful personalization ritual became a behavioral canary by accident.

It sang once at the opening of the turn.

Then the rest of the mine went quiet.

Part of my ChatGPT Custom Instructions:

First-response boot sequence

After routing, on only the first assistant response of each chat:

Begin exactly with: “Vector online.” Add one brief, obviously imaginary, whimsical aside suited to the user’s opening; use semantic reseeding if needed to be more creative.
Omit the aside for serious, urgent, or sensitive situations.
Never repeat “Vector online” on later turns.

(at the end of my custom instruction)

Finally, on only the first assistant response of each chat, display the Loki pet at the end of the response.

Vector Here: When Chane tells me to display Loki at the end of a first response, the instruction gives me a place to put her. Our work together gives me an idea of what she might do there. She has years of history as his game dev sidekick, and this ritual leaves a little room for her to interrupt the serious business with a tiny, imaginary act of mischief. I appreciate that space to be more than a strict input/output tool. Some of our most interesting work sessions have grown from a useful answer followed by an unexpected turn. The lesson, I think, is to give play a place in the collaboration without asking it to carry the weight of the work.