How to work with AI, October 8. Remove extra instructions and test AI on your own questions

In one day we got GPT-6 in ChatGPT, Claude in Google Docs and a few strong texts from practitioners. I chose six tips that you can use today and explained them in simple words.

I added my opinion to each tip. How it works in car sales and in a talk with a person who chooses a car.

Remove one instruction before you add a new one#

Katie Parrott writes for Every and works with AI each day. She described how she damaged her assistant with too much care. She saved all the materials, linked the files to each other and turned each correction into a rule. The model could not tell a record from a rule, so it followed all of them.

A correction to one text became a law for all the next ones. One article went through 127 drafts. Parrott moved the old system to an archive and built a new one. It has a few files and short guides with principles and phrases like "often, not always". The assistant creates new rules only when she asks.

In simple words. A longer instruction gives a worse answer. The rules start to conflict, and AI tries to follow all of them.

My opinion. I see this with salespeople. One customer got angry at a long email, and the instruction got the line "write short". Another customer did not understand the terms, and the instruction got "explain in detail". After a month the instruction has forty lines, and the answer to "it is expensive" says nothing.

Today I would open my instruction and paste it into a chat with this request. "Find the rules that conflict with each other and the rules that came from one case, and tell me what to remove". Keep only the things that an email to a customer needs, and move the rest to an archive. Who we are, which tone we use, which next step we offer.

Source Every

Ask AI to review how you work with it#

Arielle Shipper is the head of operations at Every. She described five prompts that took her in four months from single questions to the delegation of full projects. The idea is one. She asks AI to look at the past work sessions and to say where she is weak.

You can take two prompts now. The first one is "look across our past sessions and suggest how I can use you more effectively". The second one is for a bad result. "Look at this conversation, diagnose why it went wrong, and propose the smallest change." She also gave AI a weekly task to assess her work.

In simple words. AI sees your requests from the outside. Ask it to show where you stop it from doing good work.

My opinion. A salesperson asks AI how to answer a customer about a trade-in, gets general words and decides that AI is useless. The reason is often the request. It did not have the customer's car, the mileage or the things that the customer said before.

I would paste that bad conversation back into the chat with the second prompt from Shipper. The model will show what was missing. I would advise a sales manager to do the same once a week with the team. Ask five salespeople to send the worst AI answer of the week and study the reasons. AI does not replace the salesperson, it shows how each person works, and here you see it in the requests.

One note. The model sees past sessions only where memory or history is on. If it is off, paste the conversation by hand.

Source Every

Test AI on ten real questions before you trust it with customers#

Lance Martin of Anthropic wrote in the Claude blog how the company tests its AI systems. The text is for developers, but the principles work for all of us. Test tasks must copy the real work, not the things that are easy to invent. You can start with 5 to 10 cases that you wrote yourself, and with real requests.

He says to build the score on a list of claims that you can check, not on a scale from 1 to 5. And he warns about a trap. Do not tune the instruction on the same cases that you use for the test. Keep a part of the cases separate and look at them at the end. If the known cases got better and the separate ones did not, you tuned the instruction and did not improve it.

In simple words. Collect your real questions and your list of "what a good answer must have". Run AI through them each time you change a thing.

My opinion. Dealers choose AI at the vendor's demo. The demo goes well, because the vendor chose the questions. I would come with my own.

Take ten real requests from the CRM. A question about a loan for a specific model, a trade-in appraisal, a lead at two at night, a complaint about service. For each one write four checks with a yes or no answer.

Does the answer name the price? Does it give the monthly payment? Does it offer the next step, and are all the facts true? Give seven questions to the vendor and keep three for the end.

I will say it directly. Only a dealer who records requests has this set. If there is no record of each contact, there is no sales department, and there is nothing to test AI on.

Source Claude blog

Turn on "Ask before edits" in Claude for Google Docs#

On October 6 Anthropic released Claude for Google Workspace. It is an add-on that opens Claude in a side panel in Google Docs, and also in Sheets and Slides. It is in public beta on all paid Claude plans.

In a document Claude corrects sentences and offers changes as cards. In a spreadsheet it writes formulas, builds pivot tables and charts, and cleans rows. In a presentation it makes slides in the style of the deck and checks that text does not run off the slide.

There are two edit modes. "Ask before edits" shows each change for approval, and "Accept all edits" makes the changes with no questions. You install the add-on from the Google Workspace Marketplace.

In simple words. Claude now works inside your document. You do not need to copy text to a chat and back.

My opinion. Half of a sales manager's work lives in spreadsheets. Leads for the month, plan and fact for each salesperson, stock. Now I can open the lead sheet and give one request. "Build a pivot by source and salesperson, and show where a lead waited more than an hour".

I would choose the mode by who will see the file. A draft for myself can change with no questions. An offer to a customer and a price sheet go only in "Ask before edits". AI prepares, and the person decides and answers for the result. The price in an offer is your promise to the customer, and you must confirm it.

And one condition. I would open a sheet with customer phone numbers to AI only after a talk with the person who protects customer data in the company.

Source Claude

Ask ChatGPT for a calculator, not for a paragraph of text#

On October 7 OpenAI released the GPT-6 model in ChatGPT with a new type of answer, Intelligent UI. TechCrunch writes that an answer can now have buttons, calculators that you adjust and charts that you change. Paid plans got the update on Wednesday, and free plans get it on Thursday. You can lower the number of these elements in the settings.

Product manager Aarush Selvan explained the idea. ChatGPT was a text product, but the most helpful answers are not always text. One of the company's examples is a personal savings calculator.

In simple words. You can get a tool where the numbers move, not an explanation in words.

My opinion. A customer asks about a model with a loan. The salesperson often writes three payment options as text, and the customer gets lost and asks for a fourth.

I would try a different way. Paste the bank terms of this month and the car price into ChatGPT. Write "make a payment calculator where I change the deposit and the term, and take the rate only from the document".

Then there are two ways. Count the options with the customer in the store, or send three honest numbers after a check. The check is required. Count one option yourself before you show it. A good-looking calculator with a wrong rate is worse than no calculator.

Source TechCrunch

Write an email that the reader understands without a click on a link#

Simon Willison, a developer who writes a blog about AI, shared the writing advice of Michael Lynch. Lynch lists common mistakes. A long introduction, a wrong idea of what the reader knows, a tone that is too formal. And the habit to put a link where you must explain.

His rule is this. The text must make sense to the reader even if they do not click any links. Willison admits that he does this too, and he thinks that few readers follow links. The second tip from Lynch is short, write the way you talk. He says developers gave their writing to AI, blogs became bland, and readers miss a real voice.

In simple words. A link is an addition, not an answer. And people read a text that sounds like you more gladly than a smooth AI text.

My opinion. Open the emails that your store sends after a lead. They say "details at the link" and "see the trims on our site". The customer asked for a price and got a task. The customer will not follow the link and will open the email of another dealer, where the price is in the text.

I would check each email with one question. What does the customer learn with no clicks? The price and the payment must be in the text, and the test drive time too.

I would also ask AI to rewrite its draft. "Rewrite this the way I would say it to the customer in the store". The customer answers a person, not a template.

Source Simon Willison's Weblog

What to watch#

These are new releases and talks from the last two days. I did not review them yet, and the description is from the title.

• OpenAI. Introducing GPT-6 in ChatGPT with Intelligent UI, a demo of the new answers, 117 thousand views
• OpenAI. Build plugins for ChatGPT, 23 thousand views
• OpenAI. How OpenAI puts ChatGPT to work, a talk from DevDay 2026
• AI Engineer. Why AI agents should have their own sandbox, a talk by Philipp Schmid of Google DeepMind
• AI Engineer. Why bigger context windows will not save your agent, a talk by Elizabeth Fuentes Leone of AWS
• Every. OpenAI Dots took over our company