<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Pavel Khodalev</title>
    <description>The latest articles on DEV Community by Pavel Khodalev (@pavel_khodalev).</description>
    <link>https://dev.to/pavel_khodalev</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4239930%2F010701e2-a063-463d-bf5c-d7435c79d933.jpg</url>
      <title>DEV Community: Pavel Khodalev</title>
      <link>https://dev.to/pavel_khodalev</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/pavel_khodalev"/>
    <language>en</language>
    <item>
      <title>Stop Sitting With Your AI in a Chat</title>
      <dc:creator>Pavel Khodalev</dc:creator>
      <pubDate>Sat, 21 Nov 2026 03:46:54 +0000</pubDate>
      <link>https://dev.to/pavel_khodalev/stop-sitting-with-your-ai-in-a-chat-113b</link>
      <guid>https://dev.to/pavel_khodalev/stop-sitting-with-your-ai-in-a-chat-113b</guid>
      <description>&lt;p&gt;I do not sit with AI in a chat anymore. I give it tasks in a tracker, the same way I give tasks to a team. It calls me only when it cannot go on without me.&lt;/p&gt;

&lt;p&gt;It has been about a month. I have ten AI assistants, each with its own role. I set a task and answer inside the task, often with one emoji.&lt;/p&gt;

&lt;p&gt;I use AI as much as before. Only the place where we meet has changed.&lt;/p&gt;

&lt;p&gt;My tracker is Todoist. The tool does not matter much, IMHO. Any tracker with tasks, comments, labels and phone alerts will do. The hosting examples below are from a demo project, and the tasks are made up.&lt;/p&gt;

&lt;h2&gt;
  
  
  How it works
&lt;/h2&gt;

&lt;p&gt;I set a task as I would for a person. By voice from my phone, or in text: "Compare three hosting plans for our site." I do not pick the assistant.&lt;/p&gt;

&lt;p&gt;I drop the task into the inbox. That is it.&lt;/p&gt;

&lt;p&gt;My server is an old laptop that is always on. Once a minute a small program runs on it. No AI there, and it costs nothing. When a new task shows up in the tracker, the program leaves a note in a queue.&lt;/p&gt;

&lt;p&gt;The note wakes my dispatcher, the task assistant. It gives the task a short title and keeps my own phrase below. It sets the due date to today and hands the task to its owner in my words.&lt;/p&gt;

&lt;p&gt;The owner answers "Taken", silently. Each assistant keeps its own memory for its area, so two lines are enough.&lt;/p&gt;

&lt;p&gt;Every AI comment starts with a tag. The tag tells me whether I am needed. Only three tags call me. "For review" means a result is ready.&lt;/p&gt;

&lt;p&gt;"Question" means nothing moves until I decide. "To send" means a message waits for my word. Everything else stays in the task. It never reaches my phone.&lt;/p&gt;

&lt;p&gt;Each ask comes with a default, "I propose this". And it comes once. Three open points arrive as one numbered list, each item with its own default. Hunting for asks across comments eats a lot of focus.&lt;/p&gt;

&lt;p&gt;In the demo the assistant compared three plans and sent a table for review. Under it was one ask with two items. Take the middle plan, it covers the traffic with room to spare.&lt;/p&gt;

&lt;p&gt;Migrate on the weekend. I tapped a checkmark. Done.&lt;/p&gt;

&lt;p&gt;When an assistant calls me, the task gets one label, "Waiting on me", and the tracker alerts me. No reaction within an hour? One Slack message arrives with a link.&lt;/p&gt;

&lt;p&gt;There is no second one. The unanswered message is the reminder.&lt;/p&gt;

&lt;p&gt;The label has two axes, important and urgent. Important means money, clients, choices about people, safety and my goals. Urgent means a deadline or a meeting today or the next working day, or a live person waiting.&lt;/p&gt;

&lt;p&gt;I have four saved filters. I open the first one, important and urgent.&lt;/p&gt;

&lt;p&gt;I answer mostly with reactions. A checkmark is yes, take the default. A cross is no. The digits one to four pick an option.&lt;/p&gt;

&lt;p&gt;A thinking face asks for a simpler version. Pause puts it off. Eyes mean "seen, do not ping me in Slack".&lt;/p&gt;

&lt;p&gt;The AI never reacts itself. All assistants work under my account, so their reaction would look like mine. Sometimes a checkmark is not enough: the next step needs a sentence, or something I do by hand.&lt;/p&gt;

&lt;p&gt;Then the assistant says so at once. A checkmark never approves installing software, granting access or deleting anything. For that I write a sentence in that assistant's chat.&lt;/p&gt;

&lt;p&gt;Sending works the same way. The assistant writes the recipient and the exact text in a comment. I tap a checkmark.&lt;/p&gt;

&lt;p&gt;Then the small program sends that exact text. The AI sends nothing itself. If the text changed after I saw it, or the Slack channel is wrong, nothing goes out.&lt;/p&gt;

&lt;p&gt;This works in Slack, Telegram and my personal email. Each message is signed "Sent by the Spartan, Pavel's silicon personal assistant". People see that an assistant wrote it.&lt;/p&gt;

&lt;p&gt;A promise in a message brings the task back when it is due. If the text says "I will send it on Friday", the task comes back to me on Friday.&lt;/p&gt;

&lt;p&gt;The tracker cleans itself too. At six in the morning I get one message: what is open, what waits on me, what the AI is doing. Every hour the program sweeps for stuck tasks and sends them back to their owners. If the assistants stall for more than twenty minutes, I get one message.&lt;/p&gt;

&lt;p&gt;Am I needed right now? I do not decide that anymore. The AI decides and sets a label, and I answer with one tap. No chat needed.&lt;/p&gt;

&lt;h2&gt;
  
  
  A staff instead of windows
&lt;/h2&gt;

&lt;p&gt;A month ago my AI sessions were windows. One for the site, one for licenses, one for calls. I walked from one to another.&lt;/p&gt;

&lt;p&gt;On November 4 I rebuilt them as a staff. Each one became an assistant with a role.&lt;/p&gt;

&lt;p&gt;Each assistant has a detailed rulebook. For me it also has a card of five to seven lines. Who it is. What it takes in and what it hands over.&lt;/p&gt;

&lt;p&gt;What it does not do and cannot do. Where to find it. Moving a border between assistants means changing a rulebook, and that waits for my checkmark.&lt;/p&gt;

&lt;p&gt;Only the calendar assistant writes to my calendars. The call assistant transcribes every recording and names the speakers. Others handle AI tool licenses, expense receipts and a few projects.&lt;/p&gt;

&lt;p&gt;Above them sits a senior assistant. I call it the Spartan. It takes everything new and odd and hands the rest to owners.&lt;/p&gt;

&lt;p&gt;Once a week it brings me up to three choices. When it sees repeated work, it proposes to add, remove or merge an assistant. For other people they are all one person, the Spartan.&lt;/p&gt;

&lt;p&gt;After a call I close the laptop. That is all I do. The call assistant transcribes the recording.&lt;/p&gt;

&lt;p&gt;Two different models pull out the action items, each on its own. Only items both of them found go into work. I get fewer items, and every one is to the point.&lt;/p&gt;

&lt;p&gt;Each item on me becomes a task with the assistant who owns that area. The task quotes the item, links the transcript and asks for a check, because the AI may have misread. The assistant does everything it can without me.&lt;/p&gt;

&lt;p&gt;Then it comes to me once. An item like "discuss with someone" goes into the agenda of my next meeting with that person.&lt;/p&gt;

&lt;p&gt;When an email waits for a reply, the waiting lives with the assistant. My list stays clean. If the person is silent for more than a day, I get one ask: what do we do?&lt;/p&gt;

&lt;p&gt;After the call I wrote nothing down. I handed out nothing. The assistants split the work like a team. They called me only where my word was needed.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the chat costs me
&lt;/h2&gt;

&lt;p&gt;A few years ago I deleted social media from my phone. I caught myself getting stuck. My body felt it before my head did. At the end the time felt wasted.&lt;/p&gt;

&lt;p&gt;Then I caught the same feeling with AI. An agent finishes, and I feel I must feed it the next task right away. I gave it two or three tasks in one go. Then I sat and sorted out what had slipped through.&lt;/p&gt;

&lt;p&gt;I had six desktops full of terminal windows. I cycled through them looking for the right one. Three monitors - all busy with watching.&lt;/p&gt;

&lt;p&gt;Text ran down the screen and I kept asking one thing: am I needed here now? Mostly no. And my own work stood still.&lt;/p&gt;

&lt;p&gt;It may just be me, but one day I looked at the Claude app on my phone and saw my new Instagram.&lt;/p&gt;

&lt;p&gt;The chat also costs your word in front of people. The system did not arrive ready. Its most costly mistakes cost me my word. Minutes are nothing next to that.&lt;/p&gt;

&lt;p&gt;Once I had to raise a point on a call. The task came back fifteen minutes before the call, and I did not react. The Slack message came an hour later, after the call was over.&lt;/p&gt;

&lt;p&gt;Now such a task comes back thirty minutes early. After ten minutes of silence I am called again.&lt;/p&gt;

&lt;p&gt;A message for a project channel once went into a private chat of three. Now the program checks the channel name against Slack before sending. No match, nothing goes out.&lt;/p&gt;

&lt;p&gt;An assistant drafted a message to a colleague in a friendly tone. With that person I always write formally, briefly, without a first name. Now the assistant first reads my past messages with that exact person.&lt;/p&gt;

&lt;p&gt;There were cheaper mistakes too. An assistant checked a reference it wrote and closed the task. I never saw it. Now a result I read waits for me.&lt;/p&gt;

&lt;p&gt;Assistants wrote to me in internal task numbers. I told them: "I am a person, not a parser." Now everything has a name and a link.&lt;/p&gt;

&lt;p&gt;Every mistake is a reason to write a rule. The rule goes once into the shared rulebook, with a date and the case behind it. Every assistant follows it at once. In a chat you catch these misses only with your eyes.&lt;/p&gt;

&lt;p&gt;In a chat you pay with attention and still catch misses by eye. So let the AI call you. When it cannot go on without you, and on time.&lt;/p&gt;

&lt;h2&gt;
  
  
  How to start
&lt;/h2&gt;

&lt;p&gt;So, I am done sitting in a chat with my AI. I give it tasks in a tracker, as I would to a team. It calls me only when it cannot go on without me.&lt;/p&gt;

&lt;p&gt;I would start with three steps:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Pick a tracker you like. The AI takes tasks from it and answers in comments. Start with one area.&lt;/li&gt;
&lt;li&gt;Add the "for review" and "question" tags. Each ask comes with a default, so one checkmark answers it.&lt;/li&gt;
&lt;li&gt;Add one label for tasks waiting on you. Send alerts only on that label. Everything else stays silent.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;With several areas, give each assistant a role and a short card. Put one senior assistant above them. The rest will grow out of your own mistakes.&lt;/p&gt;

&lt;p&gt;It is a concept, not a finished product. I tune it every week. Try it and tell me where it breaks for you.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Pavel Khodalev is CTO of First Line Software and the author of **RACE Programming&lt;/em&gt;&lt;em&gt;, an AI-native software delivery framework for teams whose primary executor is an AI agent. This essay first appeared at &lt;a href="https://raceprogramming.com/essays/stop-sitting-with-your-ai-in-a-chat/" rel="noopener noreferrer"&gt;raceprogramming.com&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>productivity</category>
      <category>todoist</category>
      <category>workflow</category>
    </item>
  </channel>
</rss>
