How to work with AI, October 13. Five tips: split the task, keep passwords out of chat

Today I have five tips about order in your work with AI. How to cut a large task, where to keep rules, what you must not type into a chat, and how to start an agent. They come from Anthropic engineers, GitHub, Simon Willison and the founder of SaaStr.

You do not need to be a programmer to use them. I added my opinion and a dealership example to each tip.

Split a large task into parts that you can check one by one#

Anthropic engineers CJ Avilla and Harrison Stall spoke at the AI Engineer conference in New York. Latent Space covered their talk on October 12. They showed a mode where Claude splits the work between several agents and collects the result.

Stall described the principle. The system completes discrete, verifiable units of work and moves in phases. In the demo, agents scanned a stock watchlist. In the first phase, each agent got one stock and its own separate context. In the second phase, the findings were checked and put into one report.

Avilla gave two warnings. This mode costs much more than one agent, so the team sets a budget for the session. And the person who reviews the result becomes the bottleneck.

In simple words. Do not ask AI to "analyse everything at once". Give it a lot of small tasks of the same kind, and a separate task to check the answers.

My opinion. A sales manager wants a review of the week's leads, and there are two hundred. The request "analyse all leads" returns general words. You cannot check this answer, because it names no specific lead.

I would go in two steps. Here is the first request. "For each lead, write how fast we replied, whether we offered a visit, and how it ended. If the conversation does not show it, write 'not found'."

The second request is separate. "Check the table and show the rows where the conversation does not confirm the answer." After that, I would open ten rows myself.

Source Latent Space

Separate your AI instructions into standing rules and step-by-step procedures#

Christopher Harrison of GitHub wrote a guide for beginners on October 12. It explains how to teach the AI assistant Copilot about your project. The idea works for anyone who uses AI daily.

Harrison separates knowledge into levels. General rules sit in one file and always apply. Rules for specific kinds of work sit in other files and load by condition. He writes repeatable procedures as skills. A skill says when to use it, which steps to take and what to do when something fails.

He advises a small start. Choose one rule that you keep explaining, and write it into the file.

In simple words. "How we always work" and "how to do this one job" are two different documents. In one pile, AI mixes them up.

My opinion. In a dealership, rules are short and permanent. "We state the price with the date it is valid until, and we do not promise a discount in messages".

A procedure is longer, and you need it for a specific case. "The customer asked about a trade-in. Thank them, ask the year and the mileage, and offer two times for an inspection."

I would keep two documents. I would paste the rules into each chat and the procedure only for its task. ChatGPT and Claude have projects with a permanent instruction for this. I would start with one rule that I repeated to salespeople twice this week.

Source GitHub Blog

Enter passwords and keys yourself, outside the AI chat#

Simon Willison writes a blog about practical work with AI. On October 12, he released a small program called web-edit. It opens files on a computer through a browser, including from a phone.

Here is the reason. Willison ran a project from his phone through the Codex agent, and he needed to edit a file with credentials. He did not want the credentials to pass through the model itself. So he asked the agent to build an editor, and he opened the file directly, outside the chat.

The post says nothing about the security of the program itself. It is an early version, and it runs inside his private network.

In simple words. The model reads everything you type into the chat. A password does not go there, even when the agent asks for it.

My opinion. In a dealership, everything goes into the chat. The CRM login, a passport scan for a finance application, a photo of a driving licence. The salesperson does not think about where it goes. But the dealer answers for the customer's documents, not the program.

I would set one rule in the department. Passwords and customer documents do not go into an AI chat, and we enter system access ourselves in the settings.

The request still works. "Make a list of documents for a finance application, and write a message that asks the customer to send them". AI does not see the documents. AI prepares. A person decides and answers for the result.

Source Simon Willison's Weblog

Point AI at the leads that people have no time for#

Jason Lemkin, the founder of SaaStr, published the October edition of his guide to AI agents in sales. His company runs on three people and 21 agents. One rule in the guide is about who the agent writes to.

Lemkin grades leads from A to D. He advises not to point AI sales agents only at A-leads. He says the found money is in the B-leads, which got a score and then nobody touched them. Lemkin writes that this rule brought him an extra $500,000 from one event. This is his own number, and nobody outside has verified it.

A second idea stands next to it. Pretty good at scale beats perfect.

In simple words. Somebody will call the hottest customers anyway. AI is more useful where nobody answers now.

My opinion. Every dealership knows its B-leads. A customer asked for a price and disappeared, or got a trade-in value three months ago and did not come. A lead arrived at midnight. Measurements at dealers in different countries show that 40% of inquiries come outside working hours.

I would leave hot leads to salespeople and give the agent every lead older than 30 days with no visit. The request is short. "Write to a customer who asked for a Tiguan price two months ago and did not come. One message and one question, whether they still look for a car. Do not offer a discount."

To find these leads, you need a record of them. No record of each contact means no sales department.

Source SaaStr

Start an agent from a table that you open each day#

In the same guide, Lemkin explains where his agents came from. Almost none of them started as agents. They started as a dashboard, a tracker or a website. The thing became an agent later, after people used it daily.

Lemkin writes that this changes how you should start. Another of his observations stands next to it. A year ago, two agents took about 30 minutes a day, and now the agents take about 8 hours. An agent that makes decisions needs a check of each decision.

In simple words. First make a table and work with it by hand. Automate what you already use, not what you invented yesterday.

My opinion. Vendors offer the dealer a ready agent that starts on Monday. A month later, nobody knows what it does or how to check it. The table that shows it never existed.

I would start with one table of leads that wait for a reply longer than an hour. For a week, the sales manager builds it by hand and looks at it each morning.

Then the manager asks AI. "Each morning, collect leads with no reply for more than an hour from the CRM export. The table has the customer, the model, the lead time and the salesperson". A week later, AI writes drafts to these customers, and a person sends them.

Source SaaStr

What to watch#

These are new episodes from the last two days. I have not reviewed them yet, so I describe them by the title and the announcement.

• OpenAI. Code review is no longer the bottleneck, how Codex puts code changes and review comments in one place
• Lenny's Podcast. The internet is going to agents, Tibo Sottiaux, Head of ChatGPT and Codex, on agents that will take most actions on the internet
• Lenny's Podcast. Autonomy Only Works If You Own the Mistakes, Sottiaux on OpenAI culture and on taking production down on his third day
• Lenny's Podcast. How Meta designed Muse, design director Alex Cornell on an agent that people want to talk to