How to work with AI, October 15. Five tips: an honest agent, short instructions and voice

Today I have five tips. You will learn how to get an honest report from an agent and why long instructions make answers worse. You will also see how to tell AI what a good result looks like. I also cover voice tasks and how to check numbers in a report.

The tips come from Anthropic engineers, a writer at Every and Simon Willison. You do not need to be a programmer. For each tip, I added my opinion and an example from a dealership.

Make the agent write "could not read", not "nothing new"#

On October 8, Lance Martin and CJ Avilla from Anthropic explained how to build an agent that runs on a schedule. Their agent reads work chats every morning and writes a short summary. The authors say that such agents often lose access to a source without any warning.

Their rules are these. If a source does not open, the agent writes "source unavailable". It does not pretend the day was quiet. It checks each item again in the live source right before it sends the summary. Where the agent only reads, it gets read-only access.

The authors write that one old "still waiting on you" line costs more trust than ten missing items.

In simple words. When an agent is silent, it does not mean all is well. Ask it to say clearly what it could not do.

My opinion. At night, an agent sorts leads. In the morning, it tells the manager "no new leads". But in fact, it lost access to the website inbox. The manager feels calm, and customers wait for an answer until lunch.

I would add one line to the agent's instructions. "At the start of the summary, list all sources and mark which ones you read and which ones did not open. If a source did not open, do not write that there are no leads". No record of every lead means no sales department.

Source Claude

Before you add a new rule for AI, try to remove an old one#

Katie Parrott writes at Every about work with AI. For a long time, she added rules to the files that tell her AI how to write. In the end, one essay took 127 drafts.

Then she moved the whole setup to an archive and built a new one at a fraction of the size. Her advice is this. Before you add another instruction, try to take something away.

In simple words. The more rules you give, the worse AI follows them. A short instruction often works better than a long one.

My opinion. In a dealership, this happens with the instructions for the agent that answers customers. Every mistake gets a new rule. After a month, there are forty rules, and the agent gets more confused than before.

Once a month, I would open the instructions and ask AI. "Here are the instructions for an agent that answers car leads. Find rules that repeat each other or conflict with each other. Suggest a version half as long with the same meaning". I would test the short version on ten old leads before I use it.

Source Every

Tell AI what a good result looks like, not only what to do#

Simon Willison is a developer who writes a blog about practical work with AI. On October 14, he showed a game called "Tree Appreciator". He made it in the normal Claude app on his iPhone.

He asked for a game where you compare two trees, and he added one line. "Think hard about how to make sure it is both fun and beautiful".

The model worked for 3 hours and 7 minutes and made a calm game with music and animation. Willison writes that asking for a fun game worked surprisingly well.

In simple words. AI works better when it knows what a good result looks like. Say clearly what matters to you.

My opinion. A manager writes to AI "write a reply to a customer who asked about a loan". The manager gets a polite text with nothing useful in it. The task is there, but the quality bar is not.

I would write it like this. "The customer asked how much a RAV4 costs per month on a loan. Write a reply in three sentences. The customer must see the monthly payment in the first line and want to book a test drive. Think about what could stop them and answer it in the reply".

AI prepares, a person decides and is responsible. And the person also sets the bar.

Source Simon Willison's Weblog

Give the task by voice, then type the exact details#

On October 9, Simon Willison explained how he built a new feature for his blog by voice. He used voice mode in the ChatGPT app and talked to his laptop for half an hour while he cooked dinner. A preview window was open next to him, so he gave feedback by looking at the screen. He spent another half hour on typed changes.

His conclusion is this. Voice works well when your hands are busy. Examples, error messages and exact details are easier to paste as text.

In simple words. You can tell AI the general task by voice. Numbers and examples work better as text.

My opinion. A sales manager drives to the store and already knows what to tell the team. Twenty minutes in the car is enough to dictate a review of yesterday to AI.

I would say it like this. "Yesterday we had twelve leads, and three came to the store. Two said no because of the payment, and one because of the trade-in. Make a plan for a ten-minute morning meeting". Later in the store, I would paste the CRM numbers and one customer chat as text.

Source Simon Willison's Weblog

Click each number in an AI report and see where it came from#

On October 8, Anthropic opened a beta of Claude Dashboards on paid plans. Claude connects to company data, including Salesforce. You ask a question in normal words, and Claude builds a dashboard with charts that update on their own.

You can click any number and see the query behind it, or ask Claude to explain it. Each chart shows the time of its last update.

In simple words. AI can build a report for you. But you can and should check every number in it.

My opinion. A manager gets a nice AI report that says "lead to visit conversion 34%". The manager does not know if AI counted all leads or only website leads, for a week or for a month.

I would check the three main numbers every time. "Show how you counted the visit conversion. Which leads are in it, for which dates, and what counts as a visit".

If the answer does not match the CRM, the report does not go to the team meeting. AI does not replace the manager. It shows who did not put leads into the CRM.

Source Claude

What to watch#

These are new videos from the last two days. I have not reviewed them yet, so the notes come from the title and description.

• Every. Gemini 4 Argon can finally compete with Claude and ChatGPT, the second video from the Every team about the new Google model
• AI Engineer. Paving the road to production with governed enterprise AI agents, Harish Gaggar from Credit Karma on how to launch agents under control
• AI Engineer. What technical due diligence finds when the deck says 70% automated, Soukaïna Abelhad on how to check the promises of AI vendors
• AI Engineer. Tokens are the cheap part, Steve Jamison from Jellyfish on the full cost of work with AI
• OpenAI. Block on finding and fixing security risks with OpenAI, how the security team at Block works with OpenAI tools
• Google DeepMind. Untangling 500-page chemistry patents with Gemini 4, an example of the model working with long documents