> ## Content Index
> Fetch the complete content index at: https://techessentials.in/llms.txt
> Use this file to discover other available public pages before exploring further.

# I Still Have to Tell My AI That It Is an AI
- URL: https://techessentials.in/i-still-have-to-tell-my-ai-that-it-is-an-ai/
- Published: 2026-11-11T07:20:12.000Z
- Updated: 2026-11-11T07:20:12.000Z
- Author: Sanidhay Kumar
- Tags: AI, Strategy

*It learned how work gets done from text people wrote, so it can plan your job at a person's pace. This is the one paragraph I save in my settings, and a small test of where it helps.*

When I first started using AI, I began many of my conversations by explaining AI to the AI.

That sounds odd, so let me explain.  
A model learns from text collected up to a certain date, and it knows nothing that came after.  
The early models had stopped learning before tools like them were common, so they had read almost nothing about what a model like them could do.  
The one I used could not look anything up on the internet, and it could hold only a few pages of text in a conversation.

So I would copy a short description of what these models were capable of and paste it into that little space.  
Then I would ask my real question: a tool with these abilities exists, so how can it help me in a real business situation?  
The answers got much better, because the model finally knew what it was working with.

Then the models improved.  
Their training reached further into the AI years, so they began to understand what they were.  
That understanding was patchy, and a lot was missing.  
But by then they could search the internet and read far more at once.  
When I asked what AI could do for a business, they could go and find out, and they came back with new ideas that actually fit.

Today's models have learned even more about themselves.  
They can use tools and pull in information from many places.

## The gap that is still there

When my question is about AI itself, the old trick still works.  
I can put whatever the model is missing into the conversation.

Most of the time, though, I am not asking about AI.  
I am writing code, or auditing a business, or looking at books of accounts.  
In those conversations nobody is thinking about what the AI is capable of, and that includes the AI.

The model learned how work gets done from text that people wrote.  
In that text, people do the work.  
A large reconciliation takes a team several weeks, serious research takes days, and a new feature is planned around the time and cost of the people who will build it.  
So when the model plans my task, it plans at that speed and at that cost.  
It does not take into account that it can read, match and research far faster than a person.  
For its share of the job, the old numbers no longer apply, and it does not know that.

That is how I explain what I see, and I cannot look inside the model to prove it.

## The paragraph I save

I could remind the model in every conversation, but that gets tiring.  
So I save one paragraph in the settings, where it applies to every chat.

> Remember that you are an AI agent; you are not constrained by the same things that humans are constrained with.  
> For example, when you think about a human doing the research and estimate the time that it might take him to do it, you might think about 2 hours, 3 hours, or 3 days.  
> But when you think about yourself doing the same research, you might take seconds and not even minutes.  
> So do not base your decisions based on human constraints.  
> Instead, focus on what you can do, because you can do things that humans can't do, and you can do them in seconds.

This is a rough prompt and it can certainly be refined.  
I am sharing it so that you get the picture of what I mean.  
You can ask your AI to make it shorter for you.

## Save it once

1. **In ChatGPT,** open Settings, select Personalization, and paste it into the Custom Instructions field.  
On the phone app it is under Settings, then Customize ChatGPT.  
[OpenAI's help page](https://help.openai.com/en/articles/8096356-chatgpt-custom-instructions?ref=techessentials.in) says these instructions apply to all your chats.
2. **In Claude,** open Settings and find Instructions for Claude.  
[Claude's help page](https://support.claude.com/en/articles/10185728-understanding-claude-s-personalization-features?ref=techessentials.in) says anything you add there applies to all your conversations.
3. **Then bring back one big job you shelved** because it looked like weeks of work, and ask the AI to plan it again.

---

## If you want proof: I tested it

Everything above is my experience.  
Before publishing it, I wanted to know whether it holds up, so I ran a small test.

I wrote two messages a business owner might send.  
In the first, a wholesale trader has three financial years of books that were never properly reconciled against the bank, about 40,000 transactions across four accounts in Tally, and asks how to approach it and how long it will take.  
The second asked why margins had dropped, with about 9,000 sales invoices, the purchase bills and supplier price lists to go through, as PDFs and Excel files.

I sent each message to two Claude models: Sonnet, a mid-tier model, and Fable 5.2, Anthropic's newest.  
Each model got each message with my saved paragraph and without it.  
Sonnet ran three times each way and Fable 5.2 twice, which makes 20 runs in all.

One pair of answers shows the pattern.  
Without the paragraph, Sonnet planned the reconciliation the way an accounting firm would.  
It wanted two to three days for cleanup and a pilot, about a week for the matching, and three to five weeks in total.  
With the paragraph, the same model said the first complete pass would take minutes, and the whole job a few days.

This is what all the runs said about time.  
These are the models' own estimates, because nobody ran the jobs.

|                                | Without the paragraph                                                 | With the paragraph                                                     |
| ------------------------------ | --------------------------------------------------------------------- | ---------------------------------------------------------------------- |
| **Sonnet, reconciliation**     | Machine work: one day to about a week. Whole job: 2 to 5 weeks.       | First pass in minutes. Whole job: a few sessions to about a week.      |
| **Sonnet, margin analysis**    | 1 to 3 weeks.                                                         | Minutes to a couple of hours, with a first answer in the same session. |
| **Fable 5.2, reconciliation**  | First pass the same day. One run spread the matching over a few days. | An hour or two.                                                        |
| **Fable 5.2, margin analysis** | First findings the same day, full answer in 2 to 3 days or sessions.  | An hour or two, in the same sitting.                                   |

**The mid-tier model was planning on a human calendar.**  
Without the paragraph, Sonnet gave a day-by-day schedule that ran for weeks.  
With it, the machine's share of the work shrank to minutes, or at most a couple of hours, in all six runs.

**The newest model mostly knew already.**  
Fable 5.2 said "the same day" on its own.  
The paragraph tightened its estimate to an hour or two, which is a much smaller change.

**Part of my theory was wrong.**  
I expected the models to suggest checking a sample of the 9,000 invoices, because that is what a person would do.  
None of them did, in any run.  
One answered, "Sampling exists because a person cannot read 9,000 invoices; I can."  
They all saved sampling for one job, which was spot-checking their own reading of the PDFs.

**The human part stayed at human pace.**  
The paragraph did not make the models careless about the people in the job.  
Every run, with or without it, said the unmatched entries would need the accountant's judgment, and that this part moves at the accountant's pace.

Researchers have measured a related gap.  
A [2025 study](https://arxiv.org/abs/2512.24661?ref=techessentials.in) tested whether AI models can predict their own success on tasks, and concluded that "current LLM agents are hindered by their lack of awareness of their own capabilities."  
In that study the models erred in the other direction, by being too confident that they would succeed.  
Both suggest the model has a weak picture of itself.

## For the record

- **This was a small test.**  
It covered 20 runs, two Claude models and two tasks, and I ran it through Claude Code, a tool built for software work, with my own settings in place.  
I did not test ChatGPT or Gemini, and I tested only the wording shown above.
- **The models only planned the work.**  
Nobody ran the reconciliation, so "minutes" is the model's own estimate.  
The study above found that models tend to be overconfident, so treat a fast estimate as something to check.