Advertise on podcast: Kay Phillips Podcast

Thank you to everyone who tuned into my live video! Join me for my next live video in the app. My new AI assistant hearted my message. I had asked what it wanted to be called. It picked Scout. I told it that if it liked the name, I liked it, and it reacted with a little heart. Which I found disproportionately adorable for something I was about to hand a pile of admin work. By the time I went live on September 30, I’d been using Scout for about five hours. I had already thrown a property-management spreadsheet, bookkeeping questions, a camera search, my backend system, and the preparation for that very live into the mix. In the recording, I described myself as “an absolute menace.” Fair. I tend to have no fewer than ten AI chats going at once. I’ll start something useful, switch to another job while it works, go deal with whatever the dogs are doing, and come back to a collection of conversations that all need something from me. Somewhere in there, using AI becomes a job managing AI. That’s why this first day with Scout got my attention. The part I actually want help with Scout is my ChatGPT dot, an ongoing AI assistant I can give work to and keep talking with while that work is underway. In this replay, I show the desktop app, the conversation, its activity, and how I’ve started using it through Slack. The useful bit for me is being able to say what I need, let it work, and have it bring back the next result or question. I don’t want to spend my day remembering which chat was doing what and whether it ever finished. I have ADHD. I am frequently away from my desk. There are dogs. Slack gives me somewhere to send the thought when it occurs and receive an update without having to sit in front of the original conversation. (The dogs have yet to develop any respect for a focused work session.) One of the jobs I talked through was a property-management spreadsheet. I’d wanted help pulling dates. Scout made a backup and repaired formulas and date fields that had been stored as text. It also told me ongoing reminders weren’t set up yet. That last part matters. A repaired spreadsheet and a working reminder system are two separate results. I also had it investigating discrepancies for bookkeeping clients and reviewing my KayOS setup for duplication and other problems. Those were active jobs, with findings and back-and-forth along the way. The books and the wider audit weren’t all finished because I had an enthusiastic screen full of activity. At one point in the walkthrough, I showed an exchange where I’d asked whether something was done and the answer was essentially, nope, not yet. I also showed the corrections I’d made as we worked through a task. I’m excited about needing less hand-holding. I still want to know what actually happened. I’m talking to it like an assistant My messages are mostly big voice dumps. I use Wispr Flow, although the built-in microphone works too. I tell Scout what’s happening and what I want, then clarify when it needs something from me. I don’t want composing the perfect prompt to become the most demanding part of getting help. But there’s a difference between talking naturally and leaving the job completely undefined. In the live, I said the clearer you are, the less it has to sift through. That still applies even when the tool is better at connecting the dots for you. If you want to try this, pick one thing you’ve been putting off. Start with a result you can check. A short follow-up draft from your own notes is a perfectly good experiment. You could say: “Use these notes to draft the follow-up I owe. Keep it short and sound like me. Flag anything missing instead of making it up. Put the draft here for me to review, and don’t send it. Tell me what still needs my decision.” Then read the thing. Check the name, the promise, the dates, and whether you’d actually say it that way. You can try that in a regular AI chat too, even if dots haven’t reached your account. I’m still keeping my work outside the chat The second half of the live wandered into KayOS, my file-based system for keeping context and work available across the AI tools I use. That’s intentional. I’ve spent plenty of time in Claude, and lately I’ve been spending more time in ChatGPT. I want to use whichever tool works best for the job without rebuilding my context every time I switch. KayOS is a regular file structure with guidance, projects, resources, and the work we create. I keep local files and cloud backups, plus an external-drive backup. It sounds excessive until you’ve had a computer crash at two in the morning after finishing the school essay due the next day. I have lived that particular technological betrayal. Several times. So yes, I’m excited about an assistant that can carry more of the work forward. I’m also keeping the underlying work somewhere I can find it, back it up, and use it elsewhere. A new favorite AI tool doesn’t need to become the only place my business exists. Free prompts from the workshop Start with something you already need to finish. Give it the source, ask for a result you can inspect, and keep the first experiment small. You can use these prompts with a dot or another AI chat. Available tools, memory, and follow-up features will differ. Use information you are allowed to share. Replace bracketed text before pasting. If a task affects another person, an account, money, or a live document, decide exactly what you are authorizing before asking the tool to act. Before you hand it the job, get clear on these questions: * What result do I need? * What should you use as the source? * What matters most if there are competing priorities? * What may you do, and where should you stop? * What would count as finished? * What should you bring back to me? 1 Give it a clear job “Help me finish [specific task]. Use only [notes, file, or verified links]. I need [specific output]. The important constraints are [time, length, audience, or other limits]. First give me a draft in this chat. Don’t send, publish, buy, delete, or change source files. Flag missing information instead of guessing. We’re finished when [observable result].” Try it on a paragraph, a follow-up, or a short checklist before handing over an entire project. 2 Find the next step in a messy note “Read the notes below. I have [amount of time] and need to move [existing project] forward. Tell me which part has a real consequence now, what can wait, and the smallest useful action I can finish in that time. Keep promising new ideas in a separate parking list. If the notes don’t support a priority, ask me one question.” Check that its priority reflects your actual commitments rather than the item that happens to be easiest to describe. 3 Prepare a human follow up “Using only these approved or fictional meeting notes, identify what I promised and draft the follow-up in this chat. Keep it under [word count]. Use plain, friendly language. Don’t invent an attachment, deadline, appointment time, or sales pitch. Put missing information beside the draft. Don’t send anything.” Read the final draft as if you were the recipient. Does it sound like a person continuing a real conversation? 4 Ask for useful pushback “Before helping me build this, evaluate the idea against my goal of [goal] and my constraints of [constraints]. Tell me the strongest reason to do it, the strongest reason to wait, and what evidence is missing. Recommend a smaller test if it would answer the important question. Explain your recommendation and tell me what would change your mind.” You can disagree with the recommendation. The useful part is having the assumptions out where you can inspect them. 5 Keep your voice in the work “Use my notes below to draft [format] for [audience]. Keep my main point, specific examples, and phrases that sound like me. Use these writing samples only as evidence of style. Don’t invent experiences, feelings, results, or commitments. Flag any factual claim that needs checking. Show me one draft and explain only the changes that affect meaning.” Start with your own idea or observation. You do not need a bigger pile of plausible content to sort through. 6 Check whether the work is actually done “Check this result against my original request. Show what is complete, what is still open, and what needs my decision. For each important claim, calculation, or change, tell me what you checked and what you could not verify. Give me the finished output or its exact location. Don’t mark the whole task complete because one step finished.” For a file edit, add: “Keep a recoverable original and tell me how you verified the changed parts.” Your first experiment Pick one low-risk task. Save the prompt and result. After reviewing it, ask yourself: * Did this move something I already cared about forward? * What did I still have to correct or supply? * Was the checking effort worth it? * Would I give it the same job again? You are looking for a useful piece of work and a little more capacity. That is enough for the first experiment. Start with the thing you already want off your plate This replay is an enthusiastic first look after a few hours of use. It isn’t a long-term review or proof that an AI assistant can run everything without supervision. What interests me is the boring work I’ve wanted help with all along. The date fields. The follow-up. The unfinished task that stalls because I have to remember to go back and ask what happened. Give it one of those. Tell it what you need back and where it should stop. See whether the result is useful enough to hand over the next small job. What’s the thing you’ve been dreading that you’d try first? A few corrections and setup notes for the replay I got ahead of the result when I said my vehicle tags had been renewed. The renewals were still pending emissions. I also talked about data-broker removal as if the whole job could b

Get full access to The Grit & Grace Letters at thekayphillips.substack.com/subscribe