During a phone screen interview with a Fortune 100 company, the HR person said:
We’ll get instructions to use a prompt, then the following week, we get a note saying not to use it.
Of course, I had a few theories and batted them around like a cat playing with AI.
Based on my past experiments around mode and context and the structure I’ve proposed for workflow templates, I think I know what might be the problem.
In this post, we’ll use an everyday example of two notes to get some groceries shows how interpretation becomes ambiguous and authority expands to get the job done with people and with AI.
In my earlier work, I describe this as responsibility drift—when a system fills in gaps because the job wasn’t fully defined.
Most prompts are exactly that. They leave just enough undefined that they work for the person who wrote them—and drift for everyone else.
Be Right Back and A Grocery List As Prompts In A Workflow
Let’s consider a typical two-prompt workflow in an everyday analog situation.
Someone needs to go to the store. Someone needs to write the grocery list.
The two prompts look something like this:
a note on the table that says, “Went to the store - be back soon.”
a text that says, “bananas, cheese, yogurt, cat litter.”
In my house, we have two cars, and a kid that needs a bath before bed so I want to eat dinner at 5:30pm, at the latest. I wrote the grocery list.
My husband loves to go to the store, so “be back soon” means nothing, but a good intention. Plus, he’s an engineer so I keep my phone nearby as I do the dishes because he’ll text me questions about the cheese and yogurt — I wasn’t specific.
Oh, and I didn’t get a chance to tell him we need to eat dinner early.
Given this context, the two prompts: note on table and grocery list will work differently for me and my husband, even within the same workflow.
For me, the grocery list is a reminder of the things I need, within my context.
For him, the list creates questions, delays, and gaps that need to be filled in real time to get the same result if I were using it.
Prompts Work Within The Context They Are Given
A prompt can work really well for the person who wrote it—because they already know the decision, the constraints, and what a good answer looks like.
But when someone else uses that same prompt, those gaps get interpreted.
My husband ends up buying three different types of cheese rather than asking me for more information, just like AI would keep going to finish the task when the job becomes unclear.
He might even forget the yogurt after spending so much time figuring out the cheese situation due to context loss within the processing of the prompt.
Without the original context, prompts don’t work reliably. When context is missing, AI fills in the gaps.
Context Without Ambiguity
Number one rule, if your context still makes someone (or AI) guess, you didn’t clearly define the job.
I’d be a betting woman if I told my husband “the usual kind.”
He’d think, ‘based on what usual circumstance” and guess.
My Substack Prompt For Making Data-Driven Decisions
Yesterday, I posted a prompt that I used to help me interpret a screenshot of Substack stats that forced me to make a decision versus leisurely bat around analysis without an end game.
I normally don’t post prompts, but after writing about AI workflow templates with reusable prompts, I’ve been forcing myself to share them and think about how others might use them.
What’s Wrong With My Prompt?
This prompt works for me because I’ve already built enough context within ChatGPT for it to give me actionable results.
It will likely give you confident guesses.
You are my decision partner, not an analyst.
Context: I’m reviewing performance data and want to decide what to do next.
From the data I provide: 1) Tell me one action to take in the next 24–48 hours 2) Tell me one thing to ignore (not worth acting on) 3) Tell me one signal to monitor over the next week 4) Say clearly if there is NOT enough signal to act
Constraints: No long analysis, no multiple options, be decisive and specific, optimize for momentum not completeness. If this data does not support a decision, say that explicitly.The job isn’t defined → “what to do next,” “signal,” and “ignore” are all open to interpretation
The inputs aren’t controlled → assumes clear data, relevant metrics, and a baseline that most people won’t provide
The boundaries aren’t enforced → forces action, uses subjective constraints, and relies on the user to judge when to stop
It works because I know the job, the data to input, and the limits—none of that is actually defined.
Fixing The Prompt To Be Shareable & Reliable
Let’s start with the workflow context, because that needs to be included in prompt sharing.
Without defining the input and what good output will be, you’re not setting anyone up for success. Most people say, “this prompt helps me do x when y, here it is.”
In my workflow for the Substack Stat Decision prompt, I do the following:
Post screenshot of past 90-day traffic stats (with a decision in mind)
Run the prompt
Use the output to take action or wait for more signal
Reusable & Reliable Workflow With Prompt
In this case, we would specify the workflow as follows:
Post screenshot of past 90-day traffic stats
Provide context on the decision you would like to make based on this data.
Run the prompt
Use the output to take action or wait for more signal
Also, I re-wrote the prompt to clear define: what to look at, how to interpret it, and when to stop.
Whoever uses my prompt needs to provide context on the decision they’d like to make based on the data because my AI already has that context and doesn’t need it in my version of the prompt.
Role: Determine if the 90-day Substack traffic chart shows a clear signal for a near-term decision (24–48 hours) and recommend ONE action if it does.
Input:
- Screenshot of 90-day total traffic stats
Context:
- Decision: [What are you deciding?]
Instructions:
1) Use the overall pattern in the 90-day chart as the baseline for “normal”
2) Identify whether there is a clear directional change or meaningful pattern
3) If there is a clear signal → recommend ONE specific action (24–48 hrs)
4) Identify ONE misleading pattern or spike to ignore
5) Identify ONE pattern to monitor over the next 7 days
6) If there is NOT enough signal → respond ONLY: "Not enough signal to act"
Constraints:
- Max one sentence per item
- No multiple options
- Do not assume context beyond the chartThis Is Why Prompts Don’t Scale
Prompts, like notes, are written from one person’s context — either within their thinking or within AI’s context of their use.
To be reliable, prompts for other people need to account for that context gap to ensure the input is the same and the assumptions are removed in order for them to get the same results as the person who wrote it.
This is why prompts alone don’t hold up in real workflows.
The deeper issue is how responsibility gets defined—and where it breaks under pressure.
Try this (15 minutes):
Take a prompt you’ve used recently
Ask: What is this actually responsible for?
Write one sentence:
“This AI is responsible for ___”
Then write one constraint:
“It should not ___”
Now run your prompt again and notice:
where it overreaches
where it fills gaps
That’s usually where the problem actually is.
Send this to someone whose system “works in demo” but not in production.
AI wants a job.
Are you hiring—or just giving it notes?
If you haven’t defined what it’s responsible for, it will decide for you.
Related research from the lab
Recent work in the lab has been exploring how AI systems that perform well in demos often break under real conditions—when responsibility, boundaries, and stop conditions aren’t explicitly defined.
Related notes and articles:
I love the grocery list analogy. It really works to explain the idea.
This is such a useful framework, Judy. The grocery list analogy nails it — prompts are personal by default. It connects directly to something I've been exploring: brand strategy as a constraint layer for AI. When your brand is clearly defined — voice, positioning, what you will and won't say — it fills in a lot of the context gaps before the prompt even runs. But the prompt still has to do its job. The constraint sets the boundaries. The context tells AI what you actually need within them. Without both, you're still guessing.