Forem: 달리기

I trusted the code AI wrote for me. My data was silently broken the whole time.

달리기 — Sat, 11 Apr 2026 02:09:00 +0000

When you're vibe coding with AI, a pattern develops quickly.

Describe the feature → get the code → run it → see no errors → move on to the next thing.

It's fast. But "no errors" and "working as intended" are different things. I learned that the hard way.

The feature I added

I built a post-alert tracking system into my crypto bot.

After each alert fires, it was supposed to record the price at 30 minutes, 1 hour, and 4 hours later — building a dataset to answer "when is the best time to sell?"

The bot ran cleanly. No errors. Data flowing into Google Sheets every day.

What I found

One day I opened the spreadsheet: 30-minute column full, 1-hour and 4-hour columns nearly empty.

Why didn't I catch it sooner? To verify the 4-hour entries, you'd need to fire an alert and come back four hours later. The bot looked fine, so I assumed it was working.

The problem was in the code AI had written:

# The broken version
if elapsed >= 30:
    record_30min()
elif elapsed >= 60:
    record_60min()
elif elapsed >= 240:
    record_240min()

Once the 30-minute condition matched, the elif branches never ran. Every alert logged 30 minutes — then silently skipped 1-hour and 4-hour entirely.

# The fix
if elapsed >= 30 and not recorded_30min:
    record_30min()
if elapsed >= 60 and not recorded_60min:
    record_60min()
if elapsed >= 240 and not recorded_240min:
    record_240min()

When I write code myself, I think through the logic line by line. When AI writes it, I check if it runs — and this one ran fine. The structural issue slipped through.

The part that bothered me more

If I'd run analysis on that dataset, I would have concluded: "always sell within 30 minutes."

The data looked clean. No errors, no flags. I would have had no reason to question it. Wrong data → wrong conclusion → wrong strategy, all from a bug that never made a sound.

What changed after this

AI-assisted coding is fast. But asking "does this work?" gets you "yes, it runs." That's not what you actually need to know.

After this, I started asking a different question when reviewing AI-generated code:

"What could be wrong with this?"

That shift caught more issues than anything else I'd tried.

I pitted Claude and Gemini against each other while building a crypto bot. Here's what I found.

달리기 — Tue, 31 Mar 2026 06:10:36 +0000

While building a crypto notification bot, I developed a weird habit: copying the same question into both Claude and Gemini, then comparing the answers.

My journal at the time: "Pasting the same questions into Gemini and Claude and watching them disagree. This is more fun than I expected."

Round 1: Setup time estimate

Asked both how long Make.com automation setup would take for a beginner.

Gemini: "20 minutes"
Claude: "That's quite optimistic for a first-timer. Expect 1-2 hours. Also, the free plan has a 1,000 operations/month limit — worth knowing upfront."

Tried it myself. Claude was right.

Round 2: Code review

Claude: "6.5/10. It works, but this code fails silently." (followed by a bug list)
Gemini: "100 out of 100... but here are 5 things to fix."

Gemini's praise felt good at first. After the third time, I stopped trusting it.

Round 3: Error fixing

Had an error related to Gemini's API. Asked Claude — didn't work. Asked Gemini directly — solved immediately.

Turns out, ask the tool about itself.

The moment that stuck

Asked Claude: "What if I charge people for bot signals?"

Claude: "Companies with billions in funding already do this. Selling investment signals for money can also be a legal issue. Honestly, the only realistic angle here is selling the process of building this — not the signals."

Didn't want to hear it. Wasn't wrong.

Journal after that: "There's this AI called Claude. More realistic than Gemini. Ended up buying it."

Takeaway

It's not about which one is better. It's about knowing where to use each.

Claude: uncomfortable truths, accurate bug spotting, realistic advice
Gemini: fast drafts, knows its own API best, friendly step-by-step

Using both at the same time showed me things I couldn't see using either alone.

## I started building bots because I was scared of falling behind on AI. Now I run 6 of them 24/7 on a Mac Mini.

달리기 — Tue, 24 Mar 2026 11:42:29 +0000

It started with FOMO.

Everyone around me was talking about using AI tools. I felt like if I didn't figure this out fast, I'd be left behind. So I thought — let me just try building something.

That "something" turned out to be 6 automated bots running 24/7 on a Mac Mini M1 server.

Here's what actually happened along the way.

The first real mistake

I was building a crypto alert bot for Upbit. Lost a full day's worth of data because I didn't handle API errors properly. The external API changed its response format without notice — my bot silently died.

Didn't notice until the next day.

I've since rewritten that single bot 48 times. v48 is running now.

What "running 24/7" actually means

Spoiler: it doesn't just run by itself.

I discovered Supervisor — a process manager that keeps bots alive with auto-restart. Before this, bots would crash and stay dead silently.

[program:upbit_bot]
command=python3 /Users/m1dongeun/bot_upbit/upbit_v48.py
autostart=true
autorestart=true
stderr_logfile=/var/log/upbit_bot.err.log

Telegram notifications helped too. Now I know about crashes the same day — not the next morning.

The infrastructure stack

Mac Mini M1 — always-on server
Supervisor — process manager, all 6 bots
Tailscale — SSH remote access from anywhere
Telegram — real-time bot notifications
Python — everything

What I'm running now

Upbit pump alert bot (v48)
Upbit auto-trade bot
Binance trading bot
News analysis bot
YouTube monitoring bot
Community auto-poster

One unexpected discovery

I built a news analysis bot assuming news would correlate with price movements. So I actually ran the numbers — 1,056 entries, average price change in the hour after news: -0.012%.

I built a whole bot around an assumption that the data didn't support.

Still figuring things out

The latest thing I set up was OpenClaw — an AI agent running on the Mac Mini. Got hit by rate limits on free models, Telegram 409 conflicts from multiple instances running at once, and a launchd vs Supervisor conflict that required going directly to the Mac Mini terminal to fix.

Eventually stabilized it with Anthropic API directly connected and Supervisor managing the process.

Is anyone else doing this solo?

I'm not a professional developer. Everything came from building and breaking things.

If you're running your own automation stack — what's the most annoying failure you've run into?