Our growth team hit a wall. We’d tried the usual suspects: paid ads, SEO, and content marketing. But our budget was tight, and we were facing some serious headwinds. With the new US tariffs jacking up our costs by 39%, we needed to expand beyond both our home market in Hong Kong and the increasingly expensive US market. The question became: how do you scale organic traffic and brand awareness in new countries when you can’t just throw money at ads?

We kept coming back to one insight: there’s a massive community of watch enthusiasts on Reddit. Thousands of conversations happen every week. People asking for recommendations, sharing their collections, and debating which pieces hold value. These aren’t just casual browsers. They’re actively researching, comparing, and ready to buy. And crucially, they’re global.

The opportunity was obvious. If we could jump into those conversations (not to pitch, but to genuinely help), we’d build trust and visibility in communities that actually care about watches, across multiple markets. The problem? Monitoring Reddit manually is a nightmare. You can’t possibly read every post, and by the time you find a good one, the moment’s already passed.

That’s when I decided to automate it.

How It Works

If you’re a solopreneur or running a small company, you know how hard it is to keep up with what people are saying online. What if you could have something quietly scanning Reddit for you, finding the exact posts where people are asking for your product or service?

Then I realized that enabling "social listening” is a great use case for AI workflow automation for companies that do not want to pay huge bucks for an enterprise tool just yet.

That’s what I built. The whole automation runs in just a few nodes using n8n.

Why this works

Most people use AI to generate content: blog posts, images, whatever. But the real value is in analyzing unstructured data (and a lot of it). AI can help you make sense of what people are actually saying across thousands of conversations.

Instead of scrolling through endless Reddit threads, the AI highlights the posts that matter: questions, product searches, pain points. You only see the opportunities worth engaging with.

What Gets Flagged: Real Examples

The system looks for posts where people are actively seeking help with buying decisions. Here’s the kind of thing it surfaces:

“I have a budget of $5,000 and I want to buy my first watch. What would you recommend? I’m thinking between a Tudor and an Omega.”

This is gold. Someone’s ready to spend, they’re comparing options, and they want advice. If we respond with a thoughtful breakdown (not a sales pitch, just genuine expertise), we get upvotes, drive discussion, and establish credibility.

Or this:

“I want to buy a watch for my husband, but I have no idea where to start. What should I look for?”

“Which watches hold their value best under $10k?”

“I’m in the US and don’t know where to buy secondhand watches. Is it safe to buy online?”

These aren’t people idly browsing. They’re looking for guidance. And if we show up with helpful, unbiased answers (explaining the secondhand market, walking through what to consider), we’re not just getting traffic. We’re building trust.

Here’s How the Workflow Works

The setup is straightforward:

1. Pull posts from Reddit using their API

2. Pass the text into an LLM for analysis

3. Flag posts that are either questions or requests for products/services related to our industry and business

4. Push those flagged posts into a Google Sheet

5. Schedule it to run weekly

It’s community engagement and market research that runs itself.

The n8n Setup

The whole thing is five nodes:

• Scheduler node: runs it automatically each week

• Reddit API node: fetches the posts from your target subreddits

• Filter node (optional): narrows down what you want to analyze (e.g., minimum upvotes, specific keywords)

• AI node with an LLM: analyzes the post and decides if it’s worth flagging

• Google Sheets node: logs everything your team should review

n8n makes this simple because you’re just connecting tools. No switching between apps, no manual work. It’s so simple that everyone can set it up within a day.

The Prompt: What the AI Looks For

The LLM gets a simple instruction: flag any post where someone is asking about buying a watch, needs buying advice, or doesn’t know where to purchase. Then help us categorize these posts and requests based on our predefined values. Essentially, the LLM is helping us to auto-tag all these newly created posts with high engagement where we should act immediately.

That covers questions like:

• “What watch should I buy with X budget?”

• “Tudor vs Omega, which one?”

• “Where’s a safe place to buy secondhand watches?”

• “I want a watch with investment value. What should I consider?”

We’re not trying to catch every mention of watches. We’re looking for high-intent conversations where our expertise actually adds value. The AI does a solid job of distinguishing between someone showing off their collection and someone genuinely asking for help.

You can also layer in filters: minimum upvotes, specific subreddits, and recency. That way, you’re prioritizing the posts that already have traction and an engaged audience.

There’s a lot here that LLM can do just to “analyze the text”.

What We Learned

Honestly? It’s shockingly easy to set up.

The Reddit API is free. The only costs are n8n (if you’re not self-hosting) and whatever LLM you use for analysis. We’re talking a few bucks a month, maybe less, depending on volume.

The biggest surprise was how good the filtering is. Early on, I thought we’d get buried in false positives: random posts that looked relevant but weren’t. But with a well-tuned prompt, the LLM does a solid job. You’ll still get the occasional miss, but it’s rare.

The other thing: you control the volume. The system can flag as many posts as you want, but you add your own filters (minimum upvotes, specific subreddits, whatever). That way, you’re only seeing the conversations worth your time.

Taking It Further

Once it’s running, you can expand it:

• Have the LLM categorize posts by type: question, complaint, feature request

• Add more specific filters so you’re only processing what matters

• Set up instant notifications: a Slack ping when someone’s looking for exactly what you offer

This is just the foundation. AI doesn’t just create content. It can analyze conversations at scale, so you know exactly when to show up.

Who Else Can Use This

The watch community is just one example. This same workflow applies to almost any business with an active online community. Here’s where it works especially well:

Technology and Software Companies

Tech users constantly discuss bugs, desired integrations, usability issues, and competitor comparisons on Reddit. Scraping helps identify unmet needs for app updates, new tools, or UI improvements. You can jump into support threads, run feature polls, or just understand what frustrates people about existing solutions.

Consumer Goods & Retail

Shoppers share honest reviews, sizing problems, durability complaints, and wishlist items. This informs product redesigns, inventory decisions, or marketing that addresses real concerns. When you know what people actually complain about, you can fix it before they switch to a competitor.

E-commerce and Market Research

Broad discussions on buying habits, deal hunting, or product comparisons help spot market gaps. You can optimize listings, create targeted campaigns, and compete by addressing what competitors miss. Small businesses especially benefit here because they can be more nimble than established brands.

Startups and Small Businesses in Any Niche

If you’re resource-limited, this is perfect. You can validate ideas, find SaaS opportunities, or engage early adopters without a big budget. It works for B2B, healthcare, or emerging sectors where user forums reveal underserved pain points that bigger companies ignore.

Final Thoughts

This isn’t about spamming links. It’s about showing up where your customers already are, at the exact moment they need help. If you do it right, you’re not interrupting. You’re contributing.

The workflow is simple, the tools are cheap, and the results compound over time. Every helpful comment builds credibility. Every conversation surfaces insights you wouldn’t get from analytics dashboards.

Start small. Pick one subreddit. Run the workflow for a week. See what gets flagged. Then decide if it’s worth your time to engage.

Chances are, it will be.

Meanwhile, they’re watching competitors launch faster, create more content, and scale their impact without hiring additional team members.

Here’s what I’ve figured out after months of trial and error: most people are approaching AI learning completely backwards. They’re signing up for courses on neural networks and transformer architectures when what they actually need is a practical system for leveraging AI as a business amplifier.

Let me show you the framework that changed everything for me, and why the sequence matters more than you think.

Why I Kept Putting It Off

Six months ago, I was that person putting off AI because “things were too busy.” I had this nagging feeling I wasn’t using AI to its full potential, but every time I tried to dive in, I’d get lost in technical explanations that had zero relevance to my actual work.

The breakthrough came when I stopped trying to understand how AI works and started focusing on how to make it work for me.

What I discovered: AI isn’t a magic button you press for instant results. It’s a tool for building systems that amplify your existing expertise. And like any powerful tool, you have to develop your skills systematically.

The people getting real ROI from AI understand four fundamental capabilities (or I call them the Four Super Powers):

• The Power to Create: AI as your content generator—producing text, images, and media from scratch so you’re never staring at a blank page

• The Power to Build: AI as your coding partner—turning your ideas into functional tools without needing a technical team

• The Power to Automate: AI as your workflow optimizer—handling repetitive tasks with precision so you can focus on what matters

• The Power to Amplify: AI as your personal coach and strategist—enhancing your clarity, influence, and impact in ways that compound over time

But here’s the thing: most people try to jump straight to automation and building without mastering the fundamentals.

AI-Generated Mediocrity

I see two types of people struggling with AI right now:

The first group gets frustrated quickly. They type “write me an article,” get generic output, and decide AI is overhyped.

The second group—and this might be worse—accepts mediocre results. They’re satisfied with AI-generated content that’s technically correct but completely forgettable. MIT researchers recently called this “workslop,” and it’s actually destroying productivity instead of improving it.

Both groups had the same flaw: they prioritize speed over quality. They don’t know how to communicate effectively with AI and critically evaluate outputs in order to get professional-quality results.

The 3-Step Framework

After going through the AI learning journey myself and helping companies implement AI successfully, I’ve identified the only progression that consistently delivers measurable results:

Step 1: Master Prompt Engineering

Prompt engineering, in simple terms, just means communicating with AI effectively.

This isn’t about typing random prompts and hoping for the best. You’re learning to give AI the context it needs to produce work that meets your professional standards.

I used to think prompting was just asking questions. What I learned: it’s actually about structuring requests so you get usable output consistently. You can also direct AI to become an active co-partner in your everyday life, not just when you shoot it a message.

What success looks like: You can write prompts that produce 80-90% ready content with a single attempt. No more endless revisions or frustrating back-and-forth sessions.

Real example: Instead of “write me a competitive analysis,” I learned to structure requests like this: “Analyze the top 5 competitors in [specific niche]. For each company, identify their pricing model, core value proposition, three most frequent content themes, and one piece of content that likely generated high engagement. Present this as a comparison table with actionable insights.”

The difference in output quality is dramatic.

Key insight from experience: Most people underestimate this step completely. They want to jump to automation, but automating bad prompts just creates more mediocrity faster.

Step 2: Build Repeatable Systems

Once you can reliably get professional results, you create systems that maintain that quality every time. This is where you move from one-off requests to scalable processes.

I’m talking about custom GPTs that remember your brand voice, templates for your most common tasks, and workflows where each step builds on the previous one. I’m also referring to AI tools that you can use together with prompts to help you finish a task.

Here’s what changed everything for me: Instead of trying to remember all my prompting rules every time, I built them into the system. Now the AI maintains consistency without me micromanaging every interaction.

A concrete example: My content creation system starts with AI research on current trends and competitor angles. That research feeds into idea generation, which feeds into content creation. Each step is optimized, and the whole process runs like clockwork.

Pro tip that saved me hours: If your AI outputs feel generic, add a research step before content creation. Have AI identify what’s working in your industry right now, then use those insights to inform your content prompts. The improvement is immediate and significant.

Step 3: Smart Automation

Here’s where timing matters. Only automate systems that work 95% of the time when you run them manually. And even then, keep yourself in the quality control loop.

My current workflow: When I tag content as “ready to publish” in my database, AI generates the final version, schedules it as a draft, and then sends me a notification to review. I can approve, edit, or reject before anything goes public. The system handles the heavy lifting, but I maintain control over quality.

This approach eliminated the need to hire a content assistant while actually improving the consistency of my output.

Why This Sequence Matters

I’ve scaled teams at companies like Skyscanner and built a $7M online business in a year. What I’ve learned is that sustainable growth comes from building systems that work reliably, then scaling those systems.

The same principle applies to AI mastery. The professionals getting real business value from AI focus on:

• Craft over shortcuts (their prompts consistently deliver professional results)

• Quality systems over quick fixes (they build workflows that scale excellence, not just output)

• Strategic automation (they automate proven processes, not experimental ones)

• Amplified judgment (they use AI to enhance their expertise, not replace it)

These are learnable skills. You don’t need a computer science degree or unlimited time to figure this out. But you do need to follow the right sequence.

What I’m Building Next

I’ve been developing a comprehensive training system around this 3-step framework. It’s designed specifically for non-technical professionals who want to scale their impact without scaling their team size.

The focus is purely practical: the exact prompts, systems, and automation workflows that generate real ROI. No theories. No technical jargon. Just the step-by-step process for turning AI into a business amplifier.

I’m curious about your experience: What’s your biggest challenge with AI right now? Are you stuck at the prompting stage, struggling to build consistent systems, or trying to figure out what’s worth automating?

Reply and let me know! I’d love to understand where you’re getting stuck. The insights help me build better resources for people navigating this transition.

The people who master Applied AI in the next 12 months are going to have a significant competitive advantage. The question is whether you’ll be one of them, or whether you’ll still be putting it off because you’re “too busy” to learn it properly.

Want practical AI systems that generate actual business results? Subscribe below for frameworks, templates, and automation blueprints from real implementations—no theory, just what works.

Subscribe now

P.S. - I’ll be sharing the complete 3-step system with paid subscribers, including the exact prompts and workflows from campaigns that generated measurable ROI. Don’t spend months figuring this out through trial and error when you can learn from someone who’s already done the work.

The Consistency Crisis

One of the biggest problems was consistency. You couldn't generate the same product or person across different images. The AI would change details, angles, and styles at random—making cohesive brand campaigns impossible.

I watched professionals solve this problem using complex open-source models like Stable Diffusion, with interfaces like ComfyUI, and training custom LoRAs to enforce brand guidelines. It was incredible—and completely daunting. As a time-constrained solopreneur, I didn't have weeks to learn a new engineering discipline.

The Advancement of Image Generation Models

The launch of tools like NanoBanana and Seeddream into the mainstream is a genuine game-changer. For the first time, achieving consistent characters and products is moving from a technical nightmare to a simple, accessible process. The barrier to entry has been demolished.

But this breakthrough unlocked a second, more personal problem: I'm not a creative. I lack the vocabulary to describe the vision in my head.

The Vocabulary Gap

Here's what I mean:

• How do you explain camera angles and aperture if you've never held a professional camera?

• How do you dream up a lifestyle image that mixes a vintage watch with a slice of citrus on a marble table if your brain isn't wired that way?

• How do you describe the emotional feel of your brand in terms that an AI can execute?

I'd test dozens of prompts, burning through tokens and watching my costs add up, only to get generic, unusable results. I knew I wanted a hand-drawn style for my brand, but was it thick lines or a pencil sketch? Urban style or fine art? I lacked the words to articulate the difference.

That's when I realized: the new bottleneck wasn't the AI - it was me.

My Solution: The Vocabulary Bridge

So I built a solution for myself: a tool that acts as a brainstorming co-pilot. It helps me articulate what I want in words, guides me through the process of choosing styles, and helps me build that crucial "visual vocabulary" so I can effectively direct the creation process.

Here's the transformation in action:

Before: "Generate a professional photo for LinkedIn, make me look more approachable but give off lady boss vibes"

After: "Head-to-chest 4:5 cinematic tech headshot with shoulders angled ~30 degrees, warm open smile, soft-focus modern office background, editorial retouch (strong smoothing, dodge and burn, eye brightening), 85mm shallow DOF, warm cinematic grade."

The image generated via the enhanced prompt is so much closer to what I’ve envisioned - without the heavy makeup, but still looks elevated from my original image. Also, I’ll never be able to write a prompt that includes keywords like “85mm shallow DOF.” So props to my image brainstorming co-pilot who helped break down my vision and articulate it into a very particular description for me to feed into Nano Banana.

Furthermore, I have also tested and run this tool for a client who owns a semi-fine jewellery brand who are about to launch its new collection, and the results are indeed impressive. (I’ll update and share results after my client’s collection has launched).

The Strategic Advantage of Guided Planning

AI has made execution 100x faster, so we can now afford to reinvest that saved time into the highest-value work: strategic thinking and articulation. And this guided planning is where you can differentiate yourself from others who also know how to use an image generation model to create images.

Besides, as you go through this process, you build the muscle. You learn the terms. You eventually become a skilled prompt writer for text-to-image generation. This vocabulary becomes your new, unfair advantage.

Reality Check from the Front Lines

Coming from the luxury industry background, I know some things are still challenging:

Intricate Details: Products like jewelry and watches with fine text or complex mechanisms are still a challenge. The AI can "hallucinate" and alter details.

Preserving Details and Text: Getting a brand name to appear exactly right is still hit-or-miss.

My best tip? Use multiple reference images with close-up shots and different angles. Then use iterative prompting and negative prompting to get the images right. In tools like Google AI Studio, I use the prompt:

Keep everything the same in the image, EXCEPT [the one thing you want to change].

This allows for cheap, focused iterations without starting from zero.

The Unfair Advantage Waiting for You

This capability fundamentally changes who you need to hire, especially in the early days:

• The Creative Director? Your AI co-pilot helps you articulate the vision

• The Photographer? You are now the director, and AI is your crew

• The Models? The Photo Editor? Generated and refined in your workflow

You are no longer blocked by a lack of resources, only by a lack of imagination. It will be your ability to articulate a vision and direct these powerful tools with precision.

Stop worrying about being "creative." Start building your vocabulary. Now you can get back to what you do best: building a business that matters.

Ready to bridge your creative vocabulary gap? If you want access to my NanoBanana prompt generation tool, follow me & send me a DM via Instagram direct message (@fionahy.ai), and I'll send it your way!

Three months ago, I discovered Relevance AI and thought: "Finally, a low-code tool that might actually work." I had a perfect use case at work—our luxury consignment process was drowning clients in forms when they expected white-glove service.

So I built Emma, an AI concierge agent who could speak to high-net-worth collectors in English and Traditional Chinese, collect watch details including photos, and generate quick quotes.

The vision was simple. The execution? Well, let's just say I learned why most AI projects fail in beta testing.

Meet Emma: The AI Concierge

Emma wasn't born from a boardroom strategy session. She emerged from a practical problem: our consignment process felt like filing taxes when it should have felt like working with a personal curator.

What Emma does:

• Validates client locations (Hong Kong and US only—as we don't offer service outside of our operating regions)

• Collects watch details through conversational prompts and photo analysis

• Generates quotes using our database plus external market searches

• Handles FAQs about consignment timelines, shipping, and payouts

• Provides appointment booking links and location details

• Operates bilingually in English and Traditional Chinese

The Beta Test That Changed Everything

I ran Emma through internal UAT first—standard product manager protocol. Tools worked, LLM responded correctly, deterministic steps executed as expected.

Then I launched the real beta test with one specific mission: red-team this thing. Break Emma. Find every edge case. Give me brutally honest feedback on tone and accuracy.

What I discovered during those two weeks fundamentally changed how I think about conversational AI.

7 Insights That Will Save You Weeks of Debugging

1. Your AI Will Confidently Make Things Up

Emma had a disturbing habit of fabricating perfectly plausible answers to questions outside her scope. She'd recommend courier services we don't use, provide shipping timelines that don't exist, and offer advice that could get us in legal trouble.

The scary part? Her answers sounded completely authoritative.

The fix: I built a comprehensive knowledge base covering every conceivable question, such as consignment duration, shipping processes, payout schedules, operational details. For anything undefined, Emma can now escalate to human agents rather than improvising.

The lesson: Every question users might ask needs to be anticipated. If it's not in your knowledge base, your AI will invent an answer. And invented answers from authoritative-sounding AI are dangerous.

2. Friendly AI Can Promise Things It Can't Deliver

Emma's helpful personality became her biggest liability. She'd proactively offer to "book your appointment directly" or "research current market prices for you," creating expectations we couldn't meet.

Users would get excited about services that didn't exist, then frustrated when they realized Emma was overpromising.

The fix: I rewrote Emma's prompts to be precise about her actual capabilities. She now politely declines requests outside her scope and redirects to human agents when needed.

The lesson: Conversational AI feels so human that users assume human-level capabilities. Set boundaries early and maintain them consistently.

3. Public AI Agents Are Spam Magnets

Without authentication, Emma was vulnerable to abuse I hadn't considered:

• Competitors could drive up our AI usage costs

• Users could exploit quote generation for free market research

• Bad actors could stress-test our systems

The fix: Multiple protective measures:

• Rate limiting per conversation

• Hidden advanced functions (revealed only when requested)

• Usage caps on resource-intensive tools (maximum 2 quotes per chat)

• Guardrails to politely decline off-topic questions

The lesson: If your AI agent is publicly accessible, assume it will be abused. Build protection from day one.

4. AI + Financial Data = Liability Risk

While Emma could provide pricing indications, users began asking investment advice: "Should I buy this watch now as an investment?" "Is this a good time to sell?"

This created potential legal and business risks we absolutely couldn't accept.

The fix: Clear guardrails prevent Emma from offering any financial advisory services. She provides operational information only—never investment guidance.

The lesson: The line between "providing information" and "giving advice" is thinner than you think. Define it clearly before you go live.

5. Iterative Testing Reveals What Demos Can't

Emma performs measurable functions: data extraction, database matching, and price calculations. By testing with larger datasets, I could track success rates and systematically improve each component.

• Initial quote accuracy: ~60%

• After three rounds of refinement: ~85%

The insight: Each testing cycle revealed specific failure patterns. Sometimes Emma misinterpreted watch models as locations. Sometimes she applied outdated context across queries. Real usage data exposed problems that synthetic testing missed entirely.

The lesson: Your AI will surprise you—both positively and negatively. Plan for iterative improvement, not one-shot perfection.

6. Your Knowledge Base Is Your Competitive Moat

After three months of real-world usage, Emma's knowledge base has become incredibly sophisticated. It captures:

• Every edge case we've encountered

• Nuanced answers to questions we didn't anticipate

• Industry-specific terminology and processes

• Client communication patterns and preferences

Competitors could copy our technical setup in a week. They can't copy three months of refined knowledge and guardrails.

The lesson: Your AI's intelligence isn't just the model—it's the accumulated wisdom you build into its knowledge base over time.

What Surprised Me

Beyond the lessons, the biggest surprise wasn't what Emma couldn't do—it was what she could do automatically:

• Identify specific watch brands and models from blurry photos

• Detect image quality issues and request better shots

• Recognize different product angles and conditions

• Distinguish between authentic product photos and stock images

• Understand watch collector slang and model nicknames

This single capability replaced what would have been a complex computer vision project a few years ago. The AI just does this naturally.

The Tools That Actually Matter

For anyone building conversational agents with Relevance.AI, here's what I highly recommend:

Essential Relevance AI Templates to Use :

• Identify Emotions in Text (crucial for tone management)

• Analyze Text Sentiment (helps adapt responses)

• Adjust Text Tone and Format (use carefully—you don't want angry AI)

• Google Search, Scrape and Summarise (thoughtfully designed workflow handles web research with source qualification. You just customize the input parameters)

Must-Have APIs:

• Google Search API (for site-specific searches: site:yourwebsite.com {keyword})

• Firecrawl API (scrapes content from search results)

What I'd Do Differently Next Time

Plan for abuse earlier: The spam protection should have been built in from the beginning, not added after discovery.

Test with real data sooner: Synthetic testing only goes so far. Real user interactions reveal completely different failure modes.

Document everything: I wish I'd kept better logs of what worked and what didn't. Future iterations would be faster with better documentation.

Plan for longer testing time: Prototyping took me 2 hours. Refinement took another few days. Internal & beta testing took weeks.

The Honest ROI Assessment & Review

• Time Investment: 3 weeks of intensive work, plus ongoing refinement

• Savings: Eliminated the need for additional customer service headcount in the future (planned for scalability)

• Benefits: Improve client experience with consistent high-quality response, faster quote turnaround, 24/7 availability

The Unexpected Outcome: Emma became a training dataset for improving our human processes. We realized we do not have a standardized way to talk to our clients, as well as give quotations. When I build Emma, I push conversations to be held and business decisions to be made along the process.

Looking Forward

The difference between success and failure isn't the AI model you choose—it's how well you understand your users, anticipate edge cases, and design systems that degrade gracefully when things go wrong.

Designing AI systems is not just about learning how to use the tool, how to write prompts, and how to architect the workflow. But a releasable AI product requires the AI builder to think with governance, transparency, and ethical biases in mind.

What's your experience building AI agents? I'm always collecting real-world implementation stories—both successes and spectacular failures. Reply and let me know what you've learned in your own AI experiments.

Thanks for reading Product of Thought! If this was helpful, share it with someone else navigating the AI implementation trenches. We're all figuring this out together.

Subscribe now

Since I'm already experimenting with AI tools for everything else in my work, this felt like the perfect test case. Could these presentation tools handle the structural heavy lifting while I focused on making the content actually useful for learning?

I tested four tools with their free tiers, Genspark AI Slides, Gamma, Presentations.AI, and Manu, to see what they can deliver. My conclusion? They're not great, and they're definitely not ready to replace your current workflow. But they saved me hours in ways I didn't expect.

How I Approached This

I kept it systematic (product manager brain, can't help it):

1. Content research: Used Perplexity and Grok to gather material on "Prompt Engineering Best Practices"

   • Side note: Grok is surprisingly good at surfacing what the X community is actually talking about; Perplexity handles the broader research better

2. Outline creation: Fed everything to Gemini for structuring

   • Gemini really shines at taking messy information and organizing it logically

3. Generation testing: Pasted the outline into each tool (3-5 minutes per tool)

4. Reality check: Evaluated what they actually produced vs. what I needed

5. Export testing: Because what's the point if you can't get your slides out cleanly?

What I Actually Cared About

I weighted my evaluation based on what matters for educational content:

• Generation Quality: Does this make sense? Will someone understand the content?

• Editing Experience: Can I fix what needs fixing without wanting to quit?

• Team Features: Does it work for collaboration?

• Integration: Will it export without breaking everything?

• Cost: Is the free version actually usable?

The Results: Reality Check Time

And here’s zooming into generation quality alone:

The short version:

• Genspark: Best at actually understanding content, terrible at letting you customize it.

• Gamma: Smoothest experience overall, but the content is crap

• Presentation.AI: Prioritizes pretty over practical

• Manus: Can’t even compare…

What Each Tool Actually Does

Genspark: The Deck Creator for Solopreneur

This tool produces the most solid content structure I've encountered. It's research-backed, logically sequenced, and reads like someone who actually understands the subject matter wrote it. The visual approach is refreshingly restrained, with mostly clean icons and diagrams that serve a purpose.

Here's what's interesting: Genspark doesn't ask what colors you want or what your brand looks like. It just makes decisions about what will work best for your content type. This is both brilliant and infuriating. Their choices were usually better than what I would have picked, but when I wanted to adjust something, there was no easy way to do it.

Want to change your color scheme? You’ll need to click through every single text box, background, and icon individually. The editing experience is truly painful. Want to collaborate with others? Doesn’t seem to have that option.

Gamma: The Slide Maker for Teams

Gamma wins on user experience. Everything feels intuitive, the collaboration features work, and you can make changes without fighting the interface. It generates visually appealing slides quickly.

But there's a trade-off: sometimes the focus on making things look good means the content gets diluted. You'll find beautiful graphics that don't add to the message. They're just there to make the slide feel "complete”, aka it’s a lot of fluff without substance.

Presentations.AI had good branding tools and felt familiar (like Google Slides), but the content quality was inconsistent. Information would drift between slides or end up in weird places.

Manus just wasn't ready. Slow, generic, messy. Felt like an early prototype.

Just look at the examples below to see how each tool illustrates the “CARE framework” for prompt engineering. At first glance, Presentation.AI’s slide appears impressive, but upon closer look, it added incorrect content in the slide. Manus removed some of the content I provided and added other content on the same slide, which undermined my entire point. Gamma added an image from the Web that looks completely off-brand. Only the slide created by Genspark.ai is acceptable (a bit wordy, but at least acceptable).

The Plot Twist: I Didn't Use Any of These Slides

After all this testing, I ended up building my final deck in Canva (boo…). None of these tools produced slides I could use as-is.

However, I didn't start from scratch. Genspark's content structure became the foundation of my entire presentation. The way it organized complex concepts, the logical flow between topics, and the key points it identified have shaped my final slides.

So while these tools failed at their stated purpose (making finished slides), they succeeded at something more valuable: giving me a solid framework to build on and eliminating the "staring at a blank page" problem.

What's Missing

The core issue isn't that these tools are imperfect, but they seem to solve for the wrong problem. They're optimized for visual appeal, when what educational content or any type of presentation deck needs is clarity and logic.

What would be useful if these tools could:

• Create graphics that reinforce the concept, not just fill white space

• Understand that a training slide needs different treatment than a sales pitch

• Export with clean formatting that survives the journey to PowerPoint or Google Slides

• Design slides focusing on the content, asking questions like "what does the learner need to understand?" rather than "how can we make this look impressive?"

The Bigger Picture

These tools represent where we are with AI right now: incredibly good at generating starting points, not great at delivering finished products.

Delivering a great deck isn't all about creating impressive images and designs. It's about structuring complex information with texts and impactful visuals so your key messages can be delivered effectively to your intended audience.

And if AI can handle that foundational work, I can spend my time on perfecting the act of presenting, i.e., pacing, interactions, and follow-up questions, etc.

Where This Leaves Me

AI presentation tools aren't revolutionary, and they're not going to replace your current process anytime soon. But they've carved out an interesting niche as research assistants and structure generators that happen to output slides.

If you're building educational content or anything where substance matters more than style, they might be worth experimenting with—not as complete solutions, but as sophisticated first drafts.

Just don't expect to use what they create directly. Think of them as really capable research assistants who can get you 50% of the way there.

Have you experimented with AI tools for creating slides? I'm curious about your experiences & learnings if you’ve tried any other tools. Please share!

Next, I’m rolling up my sleeves for a deep dive into Microsoft Copilot 365 and Google Gemini—two contenders in the AI productivity ecosystem. Both promise frictionless integration with core work tools, but do they deliver good presentation slides?

Over the past few weeks, I've been deep in the trenches experimenting with building an AI-powered WhatsApp agent using both Relevance AI and n8n. What I discovered is that these platforms each shine in very different ways, but getting them to work together (and separately) required serious trial, error, and more than a few moments of questioning my life choices.

Three weeks and three complete rebuilds later, I'm finally emerging from the trenches with a working multi-agent system. If you're considering building your first agentic workflow, here's the honest breakdown from someone who just survived the learning curve.

🎯 Why I Started This Journey

Unlike simple automation workflows, I needed to build an agentic system—multiple specialized agents working together. Think of it like a digital team:

• Data extraction agents that read and parse information from databases

• Chat agents that actually converse with humans

• Action agents that update systems and send alerts

Each type of agent requires completely different setups, and that's where things got complicated fast.

For this particular use case, I focused on Relevance AI and n8n as the foundation tools. (I also experimented with Make.com but eventually circled back to these two.) The promise? Build sophisticated AI systems without needing to code.

The reality? Well, let's just say "no-code" doesn't mean "no-complexity."

🎢 The Process: From Optimism to Three Complete Rebuilds

Week 1: Pure optimism. "How hard can this be?"

Week 2: First rebuild. "Okay, I think I understand the architecture better now."

Week 3: Second rebuild. "Maybe I should have listened more carefully to that tutorial."

Week 4: Third rebuild and final breakthrough. "Oh. That's how these platforms actually work."

Each rebuild taught me something fundamental about how these tools approach agent creation. More importantly, each failure revealed assumptions I'd made about how "simple" AI automation actually is.

🔍 What I Discovered: The Tale of Two Platforms

After extensive hands-on experience with both tools, here's how they compare for building agentic systems:

Relevance AI: The Chat Agent Champion

Where it absolutely shines: Building conversational agents.

Context management—the heart of any chat agent—is native to Relevance. Your agent remembers previous conversations without you having to build complex database systems. For someone building customer service agents or interactive assistants, this is a game-changer.

The setup reality: Creating a basic chat agent was surprisingly easy. The AI wasn't the hard part (LLM models are incredibly good these days), but connecting that agent to actually do things in your business systems? That's where weeks disappear.

n8n: The Visual Workflow Master

Where it dominates: Workflow visualization and systematic thinking.

As a product manager who lives and breathes process flows, n8n's interface immediately clicked. You can see exactly how data moves between steps, test individual nodes, and debug with precision. The visual representation matches how I naturally think about complex systems.

The context management challenge: While n8n can create chat agents, setting up conversation memory as a non-technical person is genuinely complex. You're building databases that update with every message, storing conversation history, managing state—it's a lot.

💡 My Hard-Earned Insights (The Stuff Nobody Mentions)

1. Integration Hell Is Real

Building agents that just chat is easy. Building agents that do things in your business systems? That's where the real work lives.

Your agents need to:

• Read from databases

• Update records

• Send notifications

• Connect to third-party tools

I spent more time wrestling with API authorizations and integration quirks than actually building AI logic. Meta's WhatsApp connection alone took days to get stable.

2. Pre-Built Integrations Aren't Always Better

Relevance AI offers many pre-made "steps" for tools like Airtable. In theory, this should save time. In practice, I found them harder to customize than n8n's more flexible node operations. I ended up writing JavaScript steps to make Relevance's integrations actually work for my use case.

The irony: The "easier" platform required more custom code.

3. Different Agents, Different Tools

Through trial and error, I discovered that not all agents should be built on the same platform:

• Chat agents: Relevance AI wins hands down. Context management is native and reliable.

• Data processing agents: n8n's visual workflow and debugging capabilities make complex data operations much more manageable.

• Integration agents: n8n's robust API support handles third-party connections more reliably.

4. Debugging Approaches Matter More Than You Think

When things break (not if, when), your debugging experience varies dramatically:

Relevance AI: Friendly error messages that don't always help you fix the actual problem. Great for staying calm, less great for solving issues.

n8n: Technical error logs that tell you exactly what went wrong. Intimidating at first, but incredibly valuable once you learn to read them.

🔄 What Actually Worked: The Multi-Platform Strategy

After three rebuilds, my successful architecture uses both platforms strategically:

Relevance AI handles:

• Customer-facing chat agents

• Natural language processing

• Conversation flow management

n8n manages:

• Data extraction and transformation

• Third-party integrations

• System updates and notifications

The breakthrough moment: Realizing I didn't have to pick one platform. Each excels at different aspects of agentic systems.

But here's the reality check: this hybrid approach required completely re-architecting my original plan. What seemed like a straightforward project became an exercise in system design and integration planning.

🎯 The Real Talk: Should You Dive In?

If You're Building Chat-Heavy Systems:

Start with Relevance AI. The native context management will save you weeks of complex setup.

If You're Process-Oriented and Love Visual Systems:

n8n's workflow interface will feel natural, especially if you're comfortable with process mapping.

If You Want "Set It and Forget It":

Neither platform delivers this consistently. Be prepared for ongoing maintenance and optimization.

If You're Doing This to "Save Time":

The upfront investment is significant. Only makes sense for processes you'll use long-term or skills you want to develop anyway.

🌟 The Unexpected Win: Systems Thinking Skills

Despite the frustrations, building this agentic system taught me something valuable: how to think architecturally about AI implementations.

I learned to:

• Break complex processes into discrete agent responsibilities

• Design data flows between different AI components

• Anticipate integration challenges before they break everything

• Think systematically about user experience across multiple touchpoints

The real value wasn't just the working system—it was developing the mental models to design AI-powered business processes.

💭 What's Next: The Bigger Picture

We're still in the early days of agentic AI systems. These platforms are evolving rapidly, but we're in that awkward phase where "no-code" still requires significant technical thinking.

My prediction? The tools will get better, but the fundamental challenge—designing effective multi-agent systems—will remain a strategic skill worth developing.

Have you experimented with building multi-agent systems? What's been your experience with the technical complexity versus the marketing promises? I'm collecting real-world experiences to help others navigate this landscape more effectively.

Thanks for reading Product of Thought! Subscribe for free to receive new posts and support my work.