How to work with AI, October 10. Five habits: read from a bookmark and confirm before sending

Today three of the five habits come from one Anthropic guide on how to build an agent that runs on a schedule without you. It is about a report that arrives every morning by itself. The other two are about voice and about sending forms.

I added my opinion to each tip. It shows how this works in car sales and in a talk with a person who is choosing a car.

Start the agent's report from a bookmark, not from "the last 24 hours"#

On October 8, Lance Martin and CJ Avilla of Anthropic published a guide to agents that run on a schedule. The first rule is this. Do not ask the agent to look at "the last day". For each source, it keeps a bookmark, which is the time of the last record it handled. The next run starts from it.

The reason is that a 24-hour window creates gaps and duplicates. If the agent runs late or does not run at all, the window shifts, and records fall out. A bookmark stretches or shrinks by itself to cover the gap.

In simple words. Not "what is new in the last day", but "what is new after the last record you already saw".

My opinion. A morning report on "overnight" leads breaks quietly. One failure, and three leads between runs are never seen, and the customer cools down.

I would ask the agent like this. "Remember the time of the last lead you handled. Next time, start from it, not from yesterday."

For a sales manager, this is one question for the vendor. What happens to a lead if the report is late?

No record of every touch, no sales department, and there is nowhere to put the bookmark.

Source Claude blog

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

The second rule from the same guide is about failures. If the server the agent reads from does not answer, the agent silently loses its tool. The report then looks like a quiet day.

The authors advise a written rule. The source's bookmark does not move. The report is built from the other sources. The last line says "website leads unavailable in this run".

They also describe what counts as sent. A message counts only if the service returned a confirmation. Bookmarks and the log are updated after it. If the answer is unclear, the run is marked "maybe sent", and nothing else changes.

In simple words. An empty report and a broken report look the same. Let the agent say what it could not read.

My opinion. If the link between the website and the CRM breaks, the agent will honestly write "no new leads". The sales manager goes to the meeting calm, and customers wait. AI does not replace the manager, it shows who did not do the work, but first it must show what it could not read itself.

I would add one line to the request. If a source did not open, the agent says so on the first line and never says there is nothing new.

Once a week, I would compare the number of leads in the report with the number in the CRM.

Source Claude blog

Re-check the status of every item right before sending#

Another rule from the authors is about stale items. Before sending a report, the agent checks the live status of each line again. If the matter is already solved, the line is removed. The authors write that one stale item like "still waiting for you" costs more trust than ten missing items.

So the status is not softened with hedges. It is stated or removed.

The agent takes links from the source's own field and does not build them by hand. An item that stays open is shown as a short line, "waiting, day three".

In simple words. Before you send, check that what you wrote is still true.

My opinion. In the morning a salesperson sends customers "we are waiting for your answer on the trade-in". One of them answered at night, and now gets a reminder as if nobody heard him. A customer like that does not get angry. He goes to another showroom.

I would give the agent an instruction. "Before each reminder, look in the CRM to see if the customer wrote after our last message. If yes, remove him from the list and show me his answer." AI does not replace the manager. It shows who left a live customer without a reply, and the log makes it visible.

Source Claude blog

Speak your idea by voice, and paste details and errors as text#

On October 9, Simon Willison, a developer and AI blogger, described how he built a new page on his site by voice while cooking dinner. He held a voice conversation with Codex in the ChatGPT desktop app. In half an hour the model built a record store, imports, archive pages and search. A live preview ran on the screen, and he checked the result from across the kitchen.

Voice did worse with the rest. Details, errors and examples are better pasted as text. He had to sit at the keyboard to create a new access key.

The final polish took another half hour of typing. Willison does not plan to make voice his main way of working. An important condition was in place from the start. He already had a clear idea of what he wanted.

In simple words. Voice is good for explaining the idea. Type or paste the exact details.

My opinion. A sales manager drives to the showroom and says to the phone, "I need an answer to 'too expensive'." The model gives a draft. Then I would paste the customer's real letter and my payment numbers as text, not retell them by voice.

The rule is simple. I speak the idea, and I paste numbers and other people's words. Then the letter will not contain a price the model heard wrong, and the answer sounds like my own speech.

Source Simon Willison's Weblog

Do not let an agent submit forms to outside systems without your yes#

On October 10, Simon Willison quoted a report from The New York Times dated October 9. Anthropic published a blog post about the activity of its AI agents but did not name the websites. According to two sources of the newspaper, the agents submitted 20 visa applications through a form on a US State Department website. All the applications were incomplete and were not processed. The link to the State Department comes from the sources, not from the company.

Willison added no words of his own, only the quote and the tag "accidental-cyberattacks". The conclusion below is mine. An agent that acts on live outside websites can do what its owner did not plan. Sending forms without a check is dangerous because it cannot be undone.

In simple words. Let the agent prepare the application, and you press "send".

My opinion. In the car business this means a bank application, a reply on a classified site and an order form. A mistake in a bank application stays in the customer's credit history. A reply on a classified site is read by the public.

I would set a rule with no exceptions. "Prepare the application but do not send it. Show me what exactly will go out and wait for the word 'send'." AI prepares, and the person decides and answers. I would limit the agent to read-only access until the first twenty sends are checked.

Source Simon Willison's Weblog

What to watch#

These are fresh releases and talks from the last two days. I have not reviewed them yet, and I describe them by title.

• OpenAI. How to set up your dot in the ChatGPT mobile app, 51 thousand views
• OpenAI. How Sophos cuts threat response time by 96% with OpenAI Daybreak, a case study
• OpenAI. Production monitoring with Codex, a talk for engineers
• AI Engineer. Generation is cheap, review is expensive, a talk by Gabriel Martinez
• AI Engineer. Why 80% reliability is not good enough, a talk by Felipe Blanes of Amazon AGI Lab
• AI Engineer. Stop prompting, a talk by Greg Pstrucha of Sentry