LLMs and AI for your daily use

A large language model is built on an idea where it learns from a large body of text, it is tuned toward the answers people rated as good or they got continuation in the prompts inferring that the previous answers in the chat were acceptable. Ask it a plain question and you get the most likely answer to that plain question. That answer is often right. It is also, by construction, an average of a large set.

This post covers three things: how to ask better, how to get past the average when the average is not what you need, and how to adapt as these tools change.

What the model does with your question#

A language model predicts the next word, given everything before it. Your prompt is most of "everything before it". Where the prompt is silent, the model fills the gap with what is typical.

Two things follow from that.#

• Most poor answers are answers to a different question: the typical version of yours.
• The fastest improvement is not a clever phrase. It is context the model could not have guessed.

A short prompt is not wrong. It just hands more of the decisions to the average.

Prompting better with five habits#

1. Describe who it is for and what may happen with the answer. The subject and the decision will change the answer more than any wording trick.
2. Give it the context and material, not just the topic. Paste the email, the figures, the clause. A model working from your text guesses less than one working from memory.
3. If possible, show it what good and / or bad looks like. Two paragraphs you wrote will teach your voice faster than a list of adjectives describing it. Also describe what to leave out as well.
4. Mention the form of the answer you will need like a table if you are going to compare, three options if you are going to choose, under 100 words if it is going in a text message etc.
5. If necessary, mention not to assume anything and ask you any questions. "Before you answer, ask me up to three questions you need answered." It is the cheapest way to find the context you forgot.

One more, from the rule I build everything on: ask it to show its working. "Quote the line you relied on" turns a summary into something you can check.

Template that might be useful:#

---

CONTEXT

I am [who you are], working on [the situation]. This is for [the reader], who will use it to [the decision or action].

TASK

[One sentence: what you want produced.]

MATERIAL

[Paste the email, figures, document or notes.] Use only this material for facts. If something you need is missing, say so instead of guessing.

WHAT GOOD LOOKS LIKE

Here is an example of the style or standard I want: [Paste a sample you wrote or like.] Avoid: [words, tone, assumptions, length].

FORMAT

[A table / three options / numbered steps / under 150 words], in Australian English.

CHECKS

• List the assumptions you made about me or the situation.
• Quote the line or name the source behind each key claim.
• Tell me what would change your answer.
• Mark anything you are unsure of.

BEFORE YOU START

If anything above is unclear, ask me up to three questions first.

---

For quick, everyday tasks, a short version usually does the job:#

I'm [who], this is for [reader] to [decide/do what]. [Task.] Use this: [material]. Give it as [format]. Tell me what you assumed.

OPTIONAL: If you want to reduce the material for it to not hallucinate or go an a space walk out of the topic's domain, add a section like:

---

SCOPE:

Example - You only read the live site and the listings it links to. You do not sign in, check out or submit any form. That includes the coverage-report "free sample" search, so you do not see its output.

---

Here is the difference on a task most of us have had.#

Before: Summarise this lease.

After: I am renting my first flat in NSW and signing on Friday. Summarise the lease below for me. I need three things: what I pay and when, what I am responsible for fixing, and how I can leave early. Quote the clause for each. Flag anything a typical lease would not have. * Paste the Lease *

The first prompt gets a summary of an average lease. The second gets an answer about this one.

Trained on averages, and how to get past them#

The average shows up in four places. Each has a fix you can type.

Average writing#

Model prose drifts toward the most common phrasing: the same openers, lists of three, and enthusiasm nobody asked for. One more thing I noticed about OpenAI models. They are trained to be more of a "Yes man". While Anthropic models went against me first and then went on a data analysis to prove the claim.

Average advice#

Ask "should I refinance?" and you get the answer for the average person. Their income, loan and plans are not yours. In credit, advice is meant to be about this customer, not a typical one. A model will not hold itself to that unless you make it. Give it your numbers. Ask which assumptions it made about you. Then ask, "What would change this answer?"

Average agreement | MOST IMPORTANT#

Models are tuned on answers people liked, and people tend to like being agreed with. So a model leans toward your side of the question.

Fix: ALWAYS ask it to argue the other side, or to list what is wrong with your plan. Ask the question without revealing your preferred answer.

The confident guess#

An answer will not show how confidently it is given. The answer was provided because you asked for it. Any decision should be checked against the confidence interval it is provided with.

Fix: Ask for sources and open the ones that matter. Ask it to separate what it read from what it recalled. For anything current, such as rates, prices and laws, use a tool that searches, and check the date on the page.

One test covers all four:

Would this answer be different for someone else?

If not, which is the case for an unbelievably huge number of cases, you will have the average.

Lastly but most importantly, if you are building something concrete or training a model, you need to drift way away from the averages that is the result of an AI prompt. To breakout of this,

Make sure you ask it to provide a review for impressions as a stranger, competitor as well as a client/customer for getting a feedback for improvement. Then use it to fix all those issue. Keep repeating this process for 3 to 5 iterations. The best part is, all of this can be done in the same prompt.

What changes next, and how to keep up#

The models will keep changing. Most of what I have written above will not, because it is about context and checking, not about any one product.

Three shifts are already visible.#

• From answering to doing: These tools now search, run code and act across apps. The question moves from "what do I ask?" to "what do I let it do, how much to retreat and how do I check it?" • From the prompt to the setup: Saved skills, instructions, your own documents and memory of past chats do much of the work a long prompt used to do. Time spent setting these up pays back on every task. • From writing to judging: When a first draft costs nothing, the scarce skill is knowing what good looks like. That is what lets you notice the average.

Adapting does not need a plan. It needs a habit.#

1. Pick one task you repeat each week.
2. Write the prompt that worked, and keep it in a file.
3. Note two numbers: minutes saved, and corrections you had to make.
4. If the corrections do not fall over a month, stop using it for that task.

Treat it like a new colleague who has read everything and never met your customer. Fast, useful, and in need of a brief.

The short version#

The average of everyone is a reasonable place to start. It is almost never the real answer for you.

So give the model what makes your situation different: who it is for, the material, an example, the shape you need. Ask it what it assumed, its confidence and what would change its mind. Check the parts that matter. The model supplies the average. The working, and the judgement, are still yours.

The Future#

The way these models are evolving and the amount of news of their roughness and its extent is coming out, we will have to add guardrails in the future in our prompts along with giving them detailed context and customised prompt.

This is where the Optional Scope mentioned above becomes not so optional anymore...