The AI Tool That Finally Stuck Wasn’t Even For Coding

I got into AI the way most developers did at the tail end of 2024. Through hype. Through Twitter threads and YouTube thumbnails promising that coding was about to change forever. Everyone talked about "vibe coding" like it was effortless. Open a window, describe what you want, and code flows out like water from a tap.

My reality looked more like a game of telephone I was playing with myself. I'd copy code from my editor, paste it into ChatGPT, type a paragraph explaining what I needed, wait for the response, copy the generated code back, find it doesn't compile, copy the error message, paste it back into ChatGPT, and repeat. With free tier limits I was rationing my questions and spending more time managing the conversation than writing code.

The worst part was the context collapse. Every time I tabbed out to paste into ChatGPT, my mental model of the code would fall apart. I'd come back to my editor and spend minutes re-orienting before I could write the next line. I was paying for a productivity tool and getting a productivity tax instead.

OpenCode

OpenCode was supposed to be the answer. An open-source AI coding agent that works directly with your codebase. No copy-paste. No context dumping. Just describe what you want and it figures out the rest. In theory.

Back then, OpenCode needed a paid ChatGPT or Claude subscription to work, and the Rupiah was sitting weak against the USD. $20/month doesn't sound like much in Silicon Valley, but convert that to IDR and suddenly you're thinking twice about whether this experiment is worth the grocery budget. I bought it anyway. Curiosity always wins in my brain, even when my wallet tells me otherwise.

The output was buggy and full of hallucinations. Code that referenced functions that didn't exist. Import paths that led nowhere. I didn't know how to steer it, how to break tasks down, how to give it the right scaffolding. I threw it away. Not with ceremony. I just closed the window and moved on.

The quiet kind of disappointment hit hardest. I wasn't angry, I was deflated. $20 was real money for me, and the tool showed nothing back. I stopped paying attention to the AI coding space for months. It just stopped feeling relevant, not through any conscious decision.

Continue

Next I tried Continue, an open-source AI extension for VSCode. Inline suggestions instead of full autonomy. Smaller scope, less to go wrong. I thought this would be the practical middle ground.

The latency killed it. Every suggestion took three to five seconds to appear. By the time the AI finished its thought, I'd already typed the code myself. You know that flow state where your fingers move faster than your conscious brain? The tool shattered it every single time. I closed that window too.

After two tools and two subscriptions down the drain, I built a comfortable mental model: AI agents are overhyped. They're not production-ready. The demos are cherry-picked and the real experience is a buggy, slow mess. I stopped caring the way you stop caring about a startup that promised to change the world and then quietly ran out of money.

The Crack

The algorithm brought me back. I was already subscribed to Theo, the t3.gg founder. He has a YouTube channel where he gives commentary on things happening in the programming space, and at some point, the thing happening was OpenClaw. The same tool I had scrolled past months earlier, now being talked about by someone whose opinion I actually trusted.

But it wasn't just Theo. I also follow security and networking people. NetworkChuck, John Hammond, LiveOverflow, David Bombal. I was an aspiring networking guy too before I leaned harder into backend development. One by one, they started talking about AI and automation. They were using it, building things with it, showing their work.

If all these dudes are talking about AI and automation at the same time, I thought, then this OpenClaw thing must be real.

The real crack came from two specific people. Neetcode, who I respected for his systematic approach to problem-solving. And ThePrimeagen, who started vibe coding and learning game development on livestream, building a game with friends entirely through AI. He was live on stream, coding with an AI agent, laughing with his friends. That was the moment I couldn't ignore anymore. You can scroll past gurus and influencers, but when people whose work you genuinely respect start doing something different, you have to ask yourself a hard question: am I the one who's wrong?

Sumopod

Right around that time, a name kept appearing on my LinkedIn feed. Ariaseta Setia Alam, building something called Sumopod. A SaaS platform for deploying code that evolved into offering AI models, containers, VPS instances. Shipping features at an absurd pace. Revenue going up every month. Partnerships with Tencent, Alibaba, Cloudeka.

I stalked his posts. Watched the platform grow from a basic idea into something with real paying customers. There was an online meetup with him. I joined just to listen. That's where he explained how he built Sumopod. None of it was manually coded. All AI, orchestrated through n8n workflows. He wasn't sitting in an IDE typing out API routes. He was designing automations in a visual builder and letting the AI wire everything together.

That got me thinking differently. A real person, not a Twitter guru, building real things with real revenue, doing what I thought was impossible. AI wasn't overhyped. I was just doing it wrong.

Hermes

I came back to the ecosystem. By now, community spinoffs of OpenClaw were popping up everywhere. One of them was promoted as an "OpenClaw killer": Hermes Agent. The pitch was an AI that learns with you, remembers what you taught it, builds skills that accumulate over time, curates itself. That promise got me. An AI that gets better the more I use it.

I tried it locally on my laptop. It took maybe fifteen minutes to get running. My first few questions were defensive, testing. "What model are you? Who are you?" I had just finished setting it up with Gemini through a complicated WARP proxy workaround, and I wanted to make sure the connection actually worked. When it answered correctly, when it understood what LLM it was running on and could explain its own capabilities, skills, plugins, and tools, I felt the difference. This was a tool that knew itself, not a chatbot pretending to be generic.

I kept using it that evening. It inferred what I wanted from partial descriptions. It remembered context between turns without me repeating myself. When I corrected it, the correction stuck, like training a junior developer who actually listens the first time.

But the Gemini connection kept giving me trouble. The API key needed WARP to work, and sometimes the proxy would drop mid-session. I'd be in the middle of a flow and suddenly get rate limit errors. Enough times to make me wonder if this whole setup was worth the hassle.

The Daily Assistant

I didn't use Hermes for coding. Not for months. After a few awkward days where I didn't know what to ask it, I started using it for the boring stuff.

CV creation was the first real task. I pasted in my raw experience and asked it to help shape a narrative. We went through fifteen iterations over several days. Each time, the assistant remembered what we'd changed before, what direction we were heading, what feedback I'd given. It felt like collaborating with an editor who had infinite patience and perfect recall.

Job fit analysis came next. I'd paste a job description and ask for a match percentage and gap analysis. This became a regular thing. Every interesting role I found, I'd run it through the assistant. It was like having a career coach on a retainer I didn't have to pay.

Interview practice followed naturally. The assistant would ask me questions, I'd answer, it'd give feedback. It remembered my resume, my gaps from the job analysis, the specific companies I was targeting. The sessions felt alive.

I started looking forward to these sessions. The tool made the parts of my day I used to dread feel productive. I was building a habit around it before I even knew what it was truly capable of.

But all of this happened on my laptop. Every session required sitting at my desk, opening the terminal, and talking to a command line. It worked, but it didn't flow into the rest of my day. The assistant was useful. It just wasn't available when I wasn't at my PC.

The VPS

I looked at Sumopod again. They rent cheap VPS instances, 60k IDR per month for 2 vCPU, 2GB RAM, 40GB storage, 20Mbps egress. About $4. They had a Hermes template. One-click deploy.

I rented one, deployed Hermes on it, connected it to Telegram. Telegram was always the plan. I use it every day, and having the assistant in my pocket changed everything.

The Telegram setup took longer than I expected. Direct messages were easy, but getting it to work with groups and threads was more hacky than I wanted. I spent an afternoon reading docs, tweaking configs, restarting the agent. My first message to the assistant in a group chat got routed to the wrong thread. I nearly gave up and went back to the terminal.

But I pushed through. By evening, I had it working. The assistant in my pocket. Ready whenever I needed it.

The Daily Log Workflow

The VPS move didn't just make the assistant always-on. It enabled something I hadn't been able to do locally: a consistent daily logging habit.

When I tried logging on my laptop, it required sitting down at the end of the day, opening the terminal, and remembering to write something. Some days I'd skip it, others I'd forget entirely because I was already in bed scrolling my phone. The friction of being tied to my desk meant the habit never stuck.

Telegram removed that friction. I could send a message from my phone while commuting, from my bed at night, from my work laptop during lunch. The assistant was always there, always remembered what we talked about yesterday. I set up a daily log workflow through Hermes. Every evening, I'd write a quick summary in Telegram. The assistant would ask follow-up questions, point out patterns, and keep a running record. After a week, I had organized logs I could reference. After a month, I had a complete picture of my professional trajectory that I could search and learn from.

Then the time issue hit.

The assistant didn't know what the actual time was. It would log my entries with wrong timestamps, sometimes overwriting older logs. I'd open my log file and find Tuesday morning's work stamped as Thursday evening, with yesterday's entry gone. I spent three days debugging this. Checking timezone configs, server clocks, cron job settings. Nothing worked the way I expected.

I nearly scrapped the whole workflow and went back to a text file. Staring at a mangled log file, asking myself if I was the fool for trusting a joke agent with my daily record. That was the closest I came to walking away for good.

I'm glad I pushed through. I figured out the timezone issue eventually, and once it was fixed, the system ran seamlessly. The process of fixing it, of having to wrestle with something that should have been simple, is what made the tool stick in a way a smooth experience never could have.

What I Actually Learned

I thought the lesson was going to be about infrastructure. About providers and context windows and deployment strategies. Those things do matter, but they're not the point.

The tools that stick in your life fit into the gaps of your actual day. The flashiest demos mean nothing if the tool isn't there when you need it.

I started by logging my day better, polishing my CV, practicing interviews, organizing my thoughts. Building an AI portfolio was never the goal. It just ended up being the side effect of solving real problems with a tool that earned my trust over months, not over minutes.

The best advice I can give someone starting out is simple: don't start with what you think you should build. Start with what you actually need help with today. A piece of writing you're stuck on. A decision you're struggling to make. A recurring task you hate doing every week. Let the tool prove itself on real problems.

Trust builds slowly, through daily utility repeated over months, until you realize the tool has earned its place in your workflow through quiet consistency.

My journey from skeptic to daily AI user took about a year, cost maybe $100 in failed experiments and subscriptions, and taught me that the best tools earn a place in your routine by being quietly consistent.