whynotroman

News · Nov 09, 2026 · 5 min

How to work with AI, November 9. Four methods so AI text does not bury people

Four tips from Simon Willison, SaaStr and Anthropic in simple words. How to format a long AI answer, how to compare options, how to test a new model and what to write in agent memory. For each tip I show how to use it in car sales.

Each morning I read the people who build AI and work with it every day. I take the advice that you can use today and explain it in simple words.

This time there are four methods. Two from Simon Willison, one from the founder of SaaStr and one from Anthropic. I added how I would use each one in a dealership.

Open an AI answer with your own sentence and fold the long agent text

Simon Willison, a developer who writes a blog about work with AI, described his method on November 9. When an agent writes the description of a code change, he opens it with a short sentence that he picked himself. He hides the full agent text in a collapsible <details><summary> block and labels it with the model name and mode, for example "Claude Opus 5.5 Medium". The reader sees the point at once and opens the details only when needed.

In simple words. People do not like to read long AI text. Give one human sentence on top and hide the rest.

My opinion. In a dealership it is the same with reports and emails. AI likes to write half a page where two lines are enough.

I would ask AI for a report to the sales manager about night leads like this. "Write the conclusion for me in the first line, then the list of leads. Mark each lead that has no reply."

A salesperson opens an AI email to a customer with their own sentence. For example "Andrew, you asked about the Camry with a loan, here is the calculation". AI prepares, a person answers.

Source Simon Willison

Ask AI for two versions, compare them by your rules and improve the winner

On November 8, Willison asked Claude Opus 5.5 to write the same SQLite function in C and in Rust. He asked it to compare speed, code complexity and file size. The Rust version was faster, because its number parsing was twice as fast as the standard one in C.

Then he asked for a custom parser in C that would match or beat Rust. The new version won, at about 16 to 19 nanoseconds per row. Then he asked to write the result as a report.

In simple words. Do not accept the first answer. Ask for two paths, the rules of comparison and an improvement of the best one.

My opinion. For a salesperson this works on the objection "too expensive". One version talks about the monthly payment, the other about the cost of ownership.

I would give AI this prompt. "Write two replies to a customer who says 'too expensive' about a Kia Sportage. The first uses the monthly payment, the second uses the three-year cost of ownership. Compare them by clarity, honesty and next step, then improve the best one."

The salesperson picks and sends it. No record of each touch means no sales department, so every reply needs a next step.

Source Simon Willison

Test your work task again on each new small model

Jason Lemkin, the founder of SaaStr, said he moved the SaaStr Connect app to Haiku 5.5 and a cheaper mix of models. The cost of the same workflow fell 72%, and the result got faster and better. He says that since early 2025 nothing SaaStr ran with agents got cheaper until this change. His advice is to re-test your work tasks on real examples every time a new small model ships.

In simple words. An expensive model is not always needed. Every few weeks, run your task on the new cheap one.

My opinion. You can do it in one evening. Take ten real anonymous leads from the week and run the same prompt.

In a dealership I would take the task "read a night lead and suggest a reply". I would run it on the current model and on the new cheap one. I would score three questions.

Is the car named, is there a visit time, are there no invented prices. If the cheap model does as well, there is no reason to pay more. A person decides, and the check stays with the person.

Source SaaStr

Tell the agent what its memory is for and keep the main file short

On November 6, Anthropic published a review of how an AI agent memory works. The main advice is to tell the agent what memory is for, but not to dictate how to use it. Rigid rules like "record only failures" reduced the gains from strong models.

It is worth saving plans, checked rules and things the agent cannot see itself. Do not save a state that goes stale fast. Keep the file that always loads short, like a table of contents with links to detailed notes. Long instructions lower accuracy.

In simple words. Explain why the agent should remember. Do not write a twenty-page instruction.

My opinion. For a CRM this means three lines per customer and a link to the history.

I would give the agent this setup. "Memory exists so the customer never repeats the same thing. Record promises and dates. Do not record prices and stock, they change."

Then the agent will not repeat yesterday's price. Buying a car takes weeks, and without a record the agent forgets the customer between talks.

Source Anthropic

What to watch

Next note · 6 minHow to work with AI, November 2. Five ways to check an AI agent in sales →Five tips from Every writers and an Arca engineer in plain words. How to ask AI about its source, how to sign off on agent work, how to keep a record per customer and what to leave to a person. For each tip I show how to use it in car sales.

Roman Spitsyn. I build sales teams where AI agents do the work and people close the deals.

Read next

← All notes