--- title: "Coax. Persuade. Cajole. Using Generative AI Right" canonical: https://www.ianlurie.com/writing/coax-cajole-persuade-gai/ date: 2025-01-04 updated: 2025-01-06 category: "AI" description: "I used ChatGPT to generate a cycling training plan and got a valuable reminder that generative AI can’t think. But it knows a lot. Getting it to convert that knowledge into…" --- Generative AI isn’t **smarter** than you: It can’t figure out the best answer to your question, and it can’t ponder a problem (yet). But it often **knows** more than you: The average LLM has a lot of information stored in its pointy little head. Coax an AI to convert all that information into something useful. Persuade it. Cajole it, even. To do that, remember: 1. Performing tasks with generative AI is not a linear process 2. You need to prompt, test, refine, fix what’s broken, and then do it again 3. The most painful prompt/test/refine/fix process is almost always a lot faster than doing it yourself Today I followed those rules, and by golly, it worked! ## Great content and bad habits: Our story begins Two things happened to me today: First, I read this piece by Wil Reynolds: [AI survival kit for marketers](https://www.seerinteractive.com/insights/survival-kit-for-marketers-career-path-in-an-ai-world). It inspired the poop out of me. Second, I realized I’ve been overtraining. I’m a cyclist. For the last six weeks, I’ve been training absurdly hard, including riding Zwift with a bunch of lunatics called [the Knights Of Sufferlandria](https://support.wahoofitness.com/hc/en-us/articles/17537254940946-Knights-of-Sufferlandria). Yes, I’m one of them. No, it makes no difference. I’m left with my metaphorical tongue flapping in my spokes every damn time. Lately, though, I’ve been getting slower. After forty-plus years, you’d think I’d no longer be a testosterone-laden wannabe superhuman who rides his knees off every day. And yet here I am, having ridden my knees off in a testosterone-laden attempt to get stronger and riding myself into the ground instead. The math does itself: AI inspiration + bad training habits = Have AI design a training plan. ## What I wanted I wanted to create a 4-week cycling and strength training plan with: 1. 6 workouts per week and 1 rest day 2. Using workouts from a program called The Sufferest. It’s not as crazy as it sounds, I promise 3. Hard rides always on Saturdays 4. Everything imported into Todoist as a project, with weeks as sections and workouts as tasks I decided to go step-by-step. _Warning! If you don’t know your limits, the sport you’re training for, and a little training science, do not use generative AI to create a workout plan!!! This is an example based in a training system I’ve used for years. I still shared it with experienced friends to make sure I wasn’t about to wreck myself. Remember: AI provides knowledge, not intelligence._ ## Step 1: Use ChatGPT to generate the training plan This, as it turned out, was the easy part. I wrote a prompt. ### The prompt **“I am a 56-year-old cyclist with an FTP of — watts. I weigh — pounds and have a VO2 Max of –. I’m having issues with overtraining. I have 8–10 hours per week for training. I want to improve my VO2 max and FTP while allowing sufficient rest and recovery time. Create a 4-week cycling and strength training plan. Each week should include a mix of recovery rides, VO2 max intervals, endurance rides, and strength training. Please structure the plan so that each week has 6 workouts and 1 rest day. Please make sure that my hardest ride is always on Saturday. Include specific strength workouts for someone who owns a pullup bar and dumbells. Where possible, use Sufferfest cycling workouts as examples.”** Note that I included: - A little bit about me to set context. Weight, functional threshold power (FTP), and VO2 max give the AI a sense of my limits - The fact that I’m overtraining and need more rest - What I want to improve - Timing and schedule I jabbed the return key and waited. ### The result I have a great, detailed plan. It included strength workouts (which I hate), and cycling workouts (which I love) and used Sufferfest videos for examples. Here’s what one week looks like (screen capture because I’m lazy): ![The training plan, courtesy of generative ai](./coax-cajole-persuade-gai/gpt-plan-1024x719.webp) Boom. That didn’t require any cajoling. Or coaxing. Or even persuading. Look at me, Mr. AI-literate. I was feeling pretty smug. Now, I just needed to turn it into a Todoist project. Remember when I said, “Performing tasks with generative AI isn’t linear?” This next part is a perfect, not-so-perfect example. ## Step 2: Generate the Todoist CSV I wanted the entire training plan in [Todoist](https://todoist.com/). It’s my favorite productivity gadget and my portable brain. But there is no way I’m cutting and pasting dozens of tasks by hand. Todoist has a nifty import feature. I’d never used it, so why not let ChatGPT do the work? ### First try: A vague prompt **“Please convert this to a format I can import into Todoist as a project.”** ChatGPT spat out a Markdown file. I’m a Markdown guy, so, hey, neat! Todoist accepts Markdown! Wrong. Todoist requires a CSV file. ChatGPT got it wrong. Or did it? Repeat after me: Prompt, test, refine. Also coax. No persuading or cajoling necessary (yet). ### Second try: Say what I mean ChatGPT needs a little more information. Todoist **does** accept Markdown for formatting. ChatGPT knew that. But it couldn’t figure out that I wanted to import this plan as tasks. Knowledge, not intelligence. However, you need a CSV to import specific tasks. OK, time to try something different: **“Please format as a CSV for importing into Todoist. Each week should be a section, and each workout should be a task.”** Poof! ChatGPT gave me a CSV file. It had sections for each week, tasks for each workout, dates, everything. I hummed cheerfully to myself as I imported the CSV. Cue the error message: Not a valid format. The amateur developer in me wanted to punch a hole in the universe. If you know, you know. ChatGPT didn’t know Todoist’s super-specific CSV formatting rules. Fields were missing, rows were different lengths, blah blah blah. As a severely CSV-impaired human, I wasn’t going to judge. But I did need to fix it. Time to do a little persuading. ### Third try: Provide an example I needed to help ChatGPT help me. I grabbed Todoist’s official CSV template and uploaded it into ChatGPT. **“Here is Todoist’s official CSV template. Please use this as a reference to fix any formatting issues with the workout plan.”** ChatGPT added the right columns. It looked much better. Same error. Grrrr. But prompt, test, refine, right?! I tried a little cajoling, saying “Quit it!” I felt a little guilty, but I had to put my foot down. ### Fourth try: Banishing comma demons Here’s a workout description: _Focus on 85–95% FTP, short intervals._ See the comma in the description? Commas are delimiters. Todoist saw that comma and thought, “Oh, a new field!” It then split the field into two columns, creating rows with different numbers of columns, and breaking the entire file. I hate commas, and so do ChatGPT and Todoist. I’m in good company. It made perfect sense, but I’d already spent a whole **twenty minutes** on this project and was getting impatient. I am not normal. My new prompt: **“Please remove all commas within the descriptions so that the only commas in the file are column delimiters. Also, ensure there are no empty rows.”** It worked. Perfectly. My whole training plan was in Todoist. I give you generative AI, folks. Not linear. Prompt. Test. Refine. And yeah, sometimes, cajole. Be firm. Hold your ground. ## One last tweak Remember my first prompt? I wanted my hardest training day to be Saturday. After import, though, my hardest days were Wednesdays, and there’s no way I’m going to pedal myself senseless if I can’t go out for pancakes after. My theory: ChatGPT didn’t understand “Saturday” versus “Wednesday” and needed dates instead. I had one more request: **“Please readjust so that the first workout is on Monday, 1/6/2025. That will ensure the hardest rides are on Saturdays.”** All set. I re-imported and now have my entire training plan in Todoist with correct dates. Here’s how it looks: ![The final training plan in Todoist](./coax-cajole-persuade-gai/the-plan-in-todoist-1024x900.webp) ## What I learned The whole process, start to finish, took thirty minutes. Next time, it’ll take ten. Doing the entire process by hand, from research to Todoist, would’ve taken hours. Since I knew my training numbers, generating the plan was pretty straightforward. Generating the Todoist file was not so easy. If I did it again, I would: - Give ChatGPT the CSV template first - Tell ChatGPT you can’t have commas in fields, or ask it to escape the commas, or ask it to follow best practices for CSV generation - Specify the start date rather than the day of the week ## Gen AI requires patience, patience, patience But the real lesson is coax, cajole, persuade. Don’t assume Generative AI is smart. Assume it knows a lot. And yes, there’s a difference. That’s why you need to prompt, test, refine, fix what’s broken, and then do it again. It’s worth it.