You’ve been using AI chat for a bit, you know AI is capable of more than that, and you want to take the next steps. Here are some tips and lessons learned to move forward into the world of agentic AI. 

Transform your workload by having AI actually do things for you instead of just telling you how to do them. No more copy and paste from the browser, now Claude or ChatGPT will actually read and write files, emails, reports, and more. 

Think of it like the personal assistant Sci-fi has been promising us for years!  

For most of these tips, I’ve included a starter “Try this” prompt. These aren’t magic spells, you can riff on them, customize them, and you’ll get even better results. Think of them as inspiration. 

1. Use the desktop apps

Seriously. Moving from ChatGPT.com or Claude.ai to the desktop application makes a HUGE difference in what it’s capable of. First of all, it’s just a cleaner interface, but then there’s the access to the agentic side of AI. These are the game changers. Just flip the selector from chat to CoWork/Code in Claude or Work/Codex in ChatGPT. Now the AI has access to files, folders, plugins, skills and tools. It can do real work for you. Most of the rest of these tips assume that you’re working in a desktop agent. 

Try this: “Open the browser, let me login to LinkedIn and scan my feed for interesting articles.” 

2. Get a head start on “moving in”

When you make the jump from the web to desktop apps, you might feel like the AI is dumber at first. That’s because the web interface remembers a lot about you and your work and that doesn’t transfer to the local agent. But you can get past that by having your Web AI introduce you to the local agent. 

Try this: In your web AI account ask this, “I’m going to be setting up another AI Agent to help me work. I want you to go through your memories about me and create an introduction to me for the new agent. What do you know about me that the new AI needs to know? Concentrate on how I like to work, what people, projects, and topics are important to me. Include anything else you think would be helpful. Give it to me as a document I can paste into the new AI.” Then copy the output into a thread in your local agent with this prompt: “We’re just getting started. I had Claude on the web (or ChatGPT) create an introduction for me to you. Read this and store anything relevant in your permanent instructions file. Things that aren’t useful there can be stored in memory or in markdown files you can reference from your instruction files. Anything you want more information on, just ask.” 

3. Use folders and projects

Sadly the word “projects” has become overloaded in the AI space, but when we talk about Agents, it helps to split your work into “projects”. A project in this sense is one or more folders with their own project documentation (memory and assets), permanent instructions (Claude.md or Agents.md), and files to work on. You create threads in projects for each new task, and the Agent can use those files and folders to keep track of everything related to that specific work. 

Try this: Create a folder on your system called “Productivity” and open a new project in Claude or ChatGPT pointed at that folder. Then start with this: “I want to make this project the hub for my daily productivity. I want you to help me configure the project, and setup tools so you can access my email, todo list, and project tracking system. Then I want to setup a daily productivity routine with you. Please setup the project then walk me through each step. Let’s begin.” 

Bonus tip: In addition to projects for each actual project, create a “General” project folder. This is your miscellaneous project where you put one-off threads or tasks that don’t make sense anywhere else. 

4. You don’t have to use the biggest and most expensive models or max reasoning

Yes, Claude Fable and ChatGPT Sol are pretty amazing, but they burn your usage limits pretty quickly. If you want that subscription to last for a whole work session of heavy use, you are usually fine using a second or even third tier model. Remember, Opus 5 and Terra are the same quality that the bleeding edge was just a few months ago and we all thought those were amazingly capable and they are MUCH cheaper. You can always run a planning session with Sol or Fable, have it create a plan document and then spin up an Opus or Terra thread to do the work. You can even use Luna or Sonnet for tasks that don’t require a lot of thinking like sorting files or making very simple edits.

5. Generate usable output for yourself and others

AIs are great at gathering and summarizing information, but the default ways it wants to present that information kinda suck. No one has time to wade through a markdown file, and text in the thread is gone as soon as the thread is. Have the AI make that report it just wrote you readable by asking it to create the document in HTML. Then you can use the in app browser or your regular browser to read it like you would a web page. You can even teach your agent your brand, so each report has your logo and colors. It’s weirdly hard to share HTML files if they have graphics or images in it (I’m actually working on a solution to that problem with my agent right now), so if you do want to pass it on to someone else, ask it to turn the HTML into a PDF. 

Try this: “Use the web and find out everything you can about [whatever topic you’re interested in]. When you’re done, create an HTML report summarizing what I need to know about it.”

6. Make your data easier to interpret

Following on from the previous, the web pages it builds you don’t have to just be text, and the research doesn’t need to be from the web. Have a monthly financial report or even your personal budget? Why not turn that data into some charts and graphs? 

Try this: “Here’s a link to a spreadsheet containing social media stats for the last three months. Take a look at it and create an HTML dashboard showing me trends and important information.”

7. Talk to your agent

If there’s one thing that has increased my ability to interact with AI more than any other, it’s this one. Use a competent dictation tool. Wispr Flow is the big name, but it comes with a subscription. I use VoiceInk (free and one-time payment pro versions available), and several other open source and LOCAL alternatives. Pick one and start using it. Then you can simply talk to the agent instead of typing. You don’t need to be perfect and the dictation doesn’t need to be perfect, the AI will figure out what you mean. 

Try this: Install a dictation tool and dictate this into a thread, “I’m writing an email to a client and I’m having trouble formulating what I want to say. I’m going to verbally dump some ideas and I want you to clean it up and give me the text to send them.” Then just talk. Be as messy as you like. Like an amazing secretary, Claude or ChatGPT will go through your dictation and out pops a well ordered email draft.

8. Conquer blank page syndrome

This is one for the creatives. None of us wants to send a client an AI design or document. Most of them aren’t ready for clients anyway. But instead of spending hours researching and developing moodboards, have an Image gen make some ideas. Some will be decent, some will be good, and some will be garbage, but I guarantee if you ask for several concepts, you’ll find something that gives you some neat ideas. 

Try this: “I’m thinking about design concepts for a flyer for a church barbecue. Create three cool ideas: One plain and conservative, one middle of the road, and one way out there and creative.” 

9. Multi-task

You know how you send a request and there’s that time while the AI is thinking, running tools, writing files, reading files, thinking, more tools, more thinking? Why sit around and wait? Open a new thread and have it work on something else while the first agent spins. By the time you’ve handed the second agent everything it needs and set it running, the first is probably ready for you. You can bounce back and forth. Eventually, you can have several agents working on individual tasks while you sit there directing traffic. (An honest note of mental health caution, though, this takes a toll. You’ll burn out before your tokens do, so be careful.)

10. Have it ask you questions

Seriously. The first skill I built, and the one I use at the beginning of every project is one I call “Interview-me”. It’s evolved over time from a script to uncover project specifics to one that is trained in interview techniques, Socratic and other methods, professional interview techniques, and coaching tools and can help me reach clarity on any subject. It’s a 24/7 sounding board and coach. 

Start with this: “I want you to interview me about this project until you’re sure you have a complete understanding of it. Ask me one question at a time. If you need to, dig deeper into my answers until you are certain you have enough information. When you’re done, record the understanding in a document.” 

11. Don’t take yes for an answer

AI wants to be helpful. While recent training has helped with the obvious sycophancy of a few models ago, it’s still too willing to tell you your idea to buy a banana stand is an amazing investment. Make it challenge you. 

Try this: “You have the concept for this project in the folder. I want you to do an adversarial review of the ideas. What are we missing? What are the greatest risks? Where might we fail, and what’s underestimated?” Then, once you have that, go here, “Ok, now looking at those risks, help me come up with a plan to mitigate them.” 

12. Break it down

Models love to spit out long lists of steps or recommendations. If you’re a human like me, that’s not how you think. You want to focus on one thing at a time, make a decision, then move on. 

Try this: When the agent gives you a long list of suggestions or steps to complete a process, “Woah, hang on. Let’s work through this one thing at a time. Give me these one at a time, and move on to the next when we’ve completed each one.” 

13. Customize skills

There are a billion skills out there. I know this because I’ve built and edited a lot of them. But skills made for everyone are less useful than skills customized for you and your flow. Even when you find a skill you think is useful (even if it’s one of mine) don’t install it blindly. Instead, hand the skill to your agent and have it review it against what it knows about you and your workflow. 

Try this: “Here’s a link to a skill I found that might be useful. Take a look at it, but don’t install it. How can we adapt this skill to be even more useful for my workflows?” Then install the customized one. 

14. Roll your own skills

When you find something you want to do over and over, have your agent turn it into a skill. This ability is literally built into the desktop versions of Claude and ChatGPT. And understand that a skill is more than a fancy prompt, it can include research, knowledge sources, and even tools and scripts. You can send the agent out to the web, learn how to do something, and turn that into a skill you can use over and over.

15. Don’t be afraid to say “Wait, what?”

Models (especially frontier models) can make you feel stupid. It’s not you, even some of the smartest people in the space have created skills to deal with exactly this issue. (See Matt Pocock’s “Wait what” skill). Go ahead and tell it to slow down, back up, and explain better. You can even give it a target language level and audience. 

Try this: “Hang on. That didn’t land with me. Can you explain that again please? Use simple language, short paragraphs, and avoid unnecessary details.” And if you need to communicate with someone else: “Rewrite that. The target audience is a non-technical person that just needs to understand what we’re asking them to do. Use 8th grade (or 6th grade) simple English and no more than one short paragraph.” 

16. Have it check its work

You don’t need to assume (and shouldn’t) that the first output is the best. But the more you generate, the less time you will have to read everything it produces. Instead, you can send the AI back to review its own work. 

Try this: “Take another pass through the report you just wrote. Fact check it and think about it from an adversarial angle. Think about the target audience. How can the output be improved?” Or even better, open a new thread and prompt this, “I just had another agent create this report for me. Can you give it a review please? Check the facts and the conclusions. Anything you have issues with?” 

17. Trust but verify

It can be scary to turn off all the approvals and give the agent full permissions to your computer and access to your accounts. But it can also be annoying to have to click “Approve” five times every prompt. Find a good balance and use prompt approval gates. The good news is that the models you will most likely be using are pretty competent and risk averse so massive screwups are rare. You can help this by adding explicit approval gates to the prompts you give. 

Try this: “I want you to safely organize my downloads folder. Review all the files. Do a READ ONLY review and give me an idea of what you would change if you could.” 

Or this: “I want you to edit the blog post draft in my Google Drive [[link]]. Make no changes, but review the document and give me a list of edits you recommend.” (By the way, I wrote this post by hand, but I used this exact tip as my editor and proofreader, so I practice what I preach…) 

18. Have it check its work

The next tip I’m going to drop in this list is maybe the most important. The AI can teach you how to use it most effectively. You see something on Social and you think, “Hey that looks cool… I wish I could…” Good news, you can. 

Try this: “Hey, I just saw a video where someone was using Claude to do this thing. Here’s a link to the video. Can we do that? Do some research and create a plan to enable us to do the same things.”

19. Depend on the fact it knows you

After a while, the agent will know you and how you work. Use that as a starting point when evaluating new workflows or tools. 

Try this: “Look at [coolnewtool.com]. I’m thinking about using that to solve [whatever problem you’re thinking about]. Based on your research and what you know about my processes and workflow, would that be a good addition or is there a better option for me?” 

20. Work with and learn from experts

AI has made us all hugely more capable. But the fact remains that human experience and expertise still has a TON of value. Whether you’re looking to get the most out of your agentic workflow, get new ideas about how to put it to work, or reviewing the security or experience of the tool you built before handing it to real-world users, your best bet will be working with someone that’s done it before and knows what to watch out for. Even world class authors use editors.  

Try this: Register for our free webinar series “Beyond the Hype: Practical AI”. Sessions 2 and 3 are made for people in your spot who want to see and learn the next steps in integrating AI into their day-to-day. 

---

Advanced tips and tricks

These next ones are real level ups. Once you’re comfortable using your agent, and you want to move up a level, these are for you: 

Plan, Plan, Plan

Sure, you CAN just sit down and start working with the agent, but you and it will work so much better if you start with a plan. This is really important when you move into building tools, but it has a lot of value even with working through a task in a project. Take the time to sit with the problem at the start and think through the process with your agent. Bonus: If you make this a document and have the agent update it as you go through the task, when you’re done, you have a record of exactly what was done, which becomes the basis for a client report. 

Try this: “We’re going to work through solving this problem. Before we do, I want you to create a work plan. Save it to the project context as a checklist, and as we move through the process, check off the steps and record what we’ve learned.” 

Use Loops

There’s little more annoying than having to check in with an agent working through the process just to realize it stopped 10 minutes ago without moving on to the next step. But you can solve some of that with a good loop prompt. 

Try this: “Use the plan we created and work the task. Each time you complete a step, check your work, then move on to the next step. Don’t stop until you’re done and don’t ask me for approvals unless you’re about to do something unrecoverable.”

WARNING: You will need to be sure you’ve reviewed your plan well and are being safe. Loops are great but when you hand the keys to the agent without checkpoints, it’s riskier.

Use Subagents

Delegating work to your agent frees you up to do the things you’re the best at. And multitasking makes you more efficient. Give your agent the same benefits with subagents. When working through a plan, your agent can decide which parts can be done simultaneously and spin up subagents to work on them. The other advantage is that by having other agents doing the heavy lifting, the main agent (the orchestrator) keeps its context clean allowing it to run longer and keep focus on getting the job done. 

Try this: “We’ve come up with a good plan to complete this task. Use the plan, spin up subagents to complete each appropriate subtask. You don’t need to do any of the actual work yourself, but keep an eye on each subagent and coordinate their efforts. When you’re satisfied with their work, and the task is done, let me know it’s ready to review.” 

Replace your subscriptions with custom tools

If you paid for every tool you hear about you will spend hundreds or thousands per month. But you now have a custom developer at your fingertips. Replace those SaaS or subscription services with custom built tools that are designed around the way YOU want to work. And your agent can help you figure out how to do that.

Try this: “I’ve been using [this tool] for a while but I don’t feel like I’m getting value from it. I want to build a customized clone. Do some research on the tool and interview me to see what features I would need in a version 1. Then help me come up with a plan to build one.”

Make it public

It’s easier than ever to create tools and content, and while it still requires a little bit of understanding to get those live, your agent can help with that, too. For simple projects, it’s even built right into the ChatGPT ecosystem with ChatGPT sites. For those of you using Claude or another agent, look at Vercel, Cloudflare pages, or other solutions. You can just ask your agent what the right solution is.

Try this: “The tool is working great locally. I’d like to publish it so my team/clients can see and use it. What’s the best way to accomplish that? Please go step by step and let’s get this online.” 

---

Here’s the thing about this list: the tips are generic. And that, unfortunately, is the problem with most of the AI advice out there. The prompt packs, the YouTube workflows, the “10x your business” Instagram reels. All of these assume your business is going to work just like the creator’s business. It doesn’t. That’s not how this happens.

Remember when I said to depend on your agent knowing you? The same thing applies to advice. Generic tips can get you this far, but the real magic starts when you’re building things around your workflow, your pain points, your team, and your clients. The way you actually work. That’s the part no blog post can help you with.

It’s also the part I do for a living. At Stratifi, we do this with clients on a regular basis, and soon I’m doing it for you, free. I’m running a three part webinar series called Beyond the Hype: Practical AI for Business. No hype, no prompt packs, no lists of tools, just plain-language about where AI fits in your business.

And if you’ve made it all the way through this article, you are exactly who these Webinar sessions are for. You can skip The Map (that’s the intro) and jump straight in. On September 9th, I presented The Build, where we took a business problem from blank page to finished tool in just about an hour. It’s like live tech improv. Then on September 23rd, I’m presenting The Agent, a deep dive into exactly the stuff we’ve been talking about in this article: the agentic world, and setting up AI to become the coworker you were promised.

We’ll run the series regularly, so if you’re reading this after September 23rd, no problem. Follow Stratifi on Instagram or connect with me on LinkedIn, and you’ll be the first to know when we run it again.

So here’s the “Try this” for the whole article: Register for The Agent: https://www.lgncollab.com/beyond-the-hype/the-agent.

And when you’re ready to figure out where exactly AI fits into your business that’s a conversation we should have: https://stratificreative.com/contact/

Brian’s LinkedIn: https://www.linkedin.com/in/brianawood/

---