Archive
Browse all Logs and Insights in chronological order.
-
Shorts / Reels Generator — Local MVP Phase 1 Complete
A record of building a local MVP with Claude Code that converts photos and short clips into 9:16 vertical MP4s. Covers the step-by-step feature expansion from v0.1.0 to v0.1.4, the Phase 1 completion criteria, and what was deferred to Phase 2.
-
When Building an MVP with AI, What the Operator Must Hold On To Is Judgment, Not Code
Lessons from building the Shorts / Reels Generator local MVP with Claude Code. Covers why operators must own purpose, validation, and completion criteria rather than focusing on the code itself.
-
Auto-Trading Partial Sell Live Verification — Confirming State Carries Across Loop Boundaries
After fixing partial take-profit, the real question was whether the state actually carried through. PARTIAL_SELL firing is one thing. But does partial_sold=True get saved? Does the next loop read it and prevent a second trigger? Does the remaining position stay under trailing stop management? This log tracks those verifications step by step.
-
An Auto-Trading Operator Doesn't Need to Know All the Code — They Need to Build Verifiable Questions with AI
The partial take-profit bug wasn't found by a code expert. It was found by an operator who noticed something wrong in the live logs — and then narrowed it down working with an AI. In auto-trading, AI collaboration isn't about delegation. It's about building a verification loop the operator can actually judge.
-
Why the Auto-Trader's Partial Take-Profit Never Fired — Scheduler Execution Path Over Logic
Partial take-profit was enabled, the rule_engine logic was implemented, and the thresholds were configured. But the intraday 5-minute sell loop was calling check_sell_signal without passing the actual position size or partial_sold state — so PARTIAL_SELL could structurally never be returned. The fix wasn't in the strategy logic. It was rewiring the scheduler's call.
-
What Matters in Auto-Trading Is Execution Path, Not Just Logic
Partial take-profit logic was implemented and the config was enabled. It never ran because the intraday sell loop wasn't passing the actual position size or partial_sold state. In automated trading, 'the logic exists' and 'the logic actually runs' are two completely different things.
-
This Is Not About Hiding Channels. It Is About Separating Roles
An insight on separating the main brand from anonymous experiment channels by using consistent operational names.
-
From Whispering to an Operating System: In June, I Am Changing How I Operate with AI
Until May, I mostly used conversations with Meta Chulbuji to clarify direction and execute projects. In June, I am shifting toward a repeatable AI operating system using harnesses, skills, workflows, and evals across AI Content Assistant, AdSense preparation, chulbuji.com, and auto-trading records.
-
I Did Not Fix the Auto-Trader Right Away. I Added Instrumentation Before Trusting the Backtest
Recent losses made me review whether the trading conditions should change. Instead of changing them immediately, I first found look-ahead bias and universe bias in the backtest, kept the current rules, and added logs plus a shadow_report structure for observation.
-
Divergence Happens in a New Window, Convergence Happens at HQ
As AI tools multiply, role separation comes first. A new GPT window works as an external thinking space for divergence without inherited context, while Meta Chulbuji absorbs the results into existing projects as the operating HQ. The key was not blocking divergence, but building a structure that recovers it as an asset.
-
Why a Good-Looking Backtest Is Not Enough Reason to Change Strategy
A strong-looking backtest is not enough reason to change an auto-trading strategy. The first question is not whether the number is attractive, but whether the number was produced under assumptions that match live trading.
-
Codex Vibe Coding Practice — Starting to See How Plugins and Skills Work
Followed Builder Josh's YouTube course to run hands-on Codex practice covering plugins, skills, DESIGN.md-based UI work, and the $pdca flow. Key takeaway: AI agents move autonomously through the entire task, so the human needs to design clear stop conditions at each stage.
-
Commit Hero Deployed — First AI Product Build Loop Closed
Took the Codex Master Class learnings into an actual project. Ran the full Deep Interview → Plan → Design → Do → GitHub → Cloudflare sequence and registered Commit Hero as Codex experiment 001 in the chulbuji.com Vibe Coding Lab.
-
What I Learned from Codex Practice — Workflow Design Matters More Than Commands
Running Codex vibe coding practice revealed that AI coding agents move autonomously through the entire task when given a broad goal. The real capability in AI development is not prompt writing — it is designing the sequence of work, the permissions granted, and the stop conditions.
-
The Core of Vibe Coding Was the Harness, Not the Code
Before building Commit Hero, I designed the judgment structure first. Running Deep Interviews to set direction, then defining scope through Plan, Design, and AGENTS.md before implementation — the experience confirmed that the human role in vibe coding is harness design, not coding.
-
The Day the Auto-Trader Bought Nothing — Tracing the Reason for Zero Orders
Four scans ran in live mode but zero orders were placed. Six BUY signals fired, yet no stock passed the RS filter. Suspected ETF benchmark skew, compared four universe metrics, found the skew exists inside our own universe too. Built a D+1/D+3/D+5 tracker for rejected stocks instead of changing the logic.
-
The RS Filter Problem Wasn't the Formula — It Was a Skewed Benchmark
I suspected the KOSPI200 ETF benchmark was distorting the RS filter due to mega-cap skew. Switching to a universe-weighted benchmark made things worse — the market-cap weighted universe RS came out even higher than the ETF. The problem isn't which benchmark you use; it's that all common benchmarks carry the same skew.
-
Building Agent Boss Board v0.1 and Setting Up the PC2 Operating Environment
A record of building the first Agent Boss Board v0.1 prototype and setting up PC2 as a secondary development and operation environment for chulbuji.com.
-
AI Content Assistant — Kmong Listing Active, Auto-Reply Setup Complete
Verified that the AI Content Assistant service is live and purchasable from a buyer's perspective on Kmong. Service approval confirmed, 5 auto-reply FAQs configured, paid ads deferred — entering the sales validation stage.
-
Technical SEO and KO/EN Cleanup in Response to GSC Indexing Holdback
Four fixes in response to GSC 'Crawled - currently not indexed': hreflang restoration, noindex sitemap exclusion, trailing slash normalization, and KO/EN content alignment.
-
Testing the First AI Shorts Production Workflow with NotebookLM and Google Vids
Drafted a shorts script with NotebookLM, generated AI video clips with Google Vids/Veo, and added subtitles and narration to validate the first end-to-end shorts production workflow. The goal was not monetization — it was finding out whether a repeatable shorts production SOP is achievable.
-
What Matters in AI Shorts Production Is Not the Video — It's the Repeatable Routine
Running the first AI shorts production experiment with NotebookLM and Google Vids/Veo revealed that the core of shorts production is not video generation itself — it is building a repeatable routine that connects script, source, video, subtitles, upload, and review.
-
Auto-Trading System Upgraded to Read Market Regime
Root-cause analysis of repeated zero-buy days during KOSPI surges, and a strategy engine overhaul that switches MA20 deviation thresholds dynamically based on BULL / NEUTRAL / BEAR market regime detection.
-
What Auto-Trading Misses in a Bull Market Is Not Individual Stocks — It's Market Context
An analysis of how a fixed MA20 deviation filter structurally blocks leading stocks during KOSPI surges, and why regime detection must come before stock screening.
-
Stock Auto-Trading Program — Strategy Engine First Refinement
A strategy engine refinement log for the stock auto-trading program, based on live-account operating data: capital structure, stop-loss rules, target profit, sector overlap limits, and profit-risk analysis log placement.
-
In Auto-Trading, Rejection Reasons Matter More Than Buy Signals
The core insight from the first refinement of the stock auto-trading strategy engine: before chasing more buy signals, the system must explain why it did not buy.
-
Started the AI Content Assistant Kmong Listing Experiment
A record of registering as a seller on Kmong (Korea's freelance marketplace) and listing the first product to validate whether the AI Content Assistant service can sell. Currently pending approval.
-
Notebook Initial Setup and chulbuji.com Operating Standards Test
A record of creating the first Notebook for chulbuji.com operations, adding operating standards, project separation rules, and AI role assignment criteria as sources, then checking its baseline understanding.
-
Notebook Is Not a Search Tool. It Is a Device for Remembering Operating Standards.
An insight from the initial chulbuji.com Notebook setup: how to manage sources, where public URLs are limited, and how Notebook fits into the AI role assignment structure.
-
chulbuji.com Logo and Favicon — Final Selection Record
After comparing three brand symbol concepts for chulbuji.com, Concept 3 was chosen as the main logo, Concept 1 as the favicon and secondary symbol, and Concept 2 was excluded from this brand system.
-
Korean Stock Auto-Trading Program — First Trading Day After the Weekend Upgrade
An operations log covering the first trading day after the weekend upgrade to the Korean stock auto-trading program, including the move toward a ₩10M seed structure.
-
Using Claude Sparingly, Not Excluding It
In AI collaboration, the point is not to use one tool less. It is to place each AI where it creates the most operational value.
-
From a System That Cannot Buy to a System That Chooses Not to Buy
An operating insight from improving the auto-trading program from a ₩5M seed structure to a ₩10M expandable structure.
-
The Logo Carries the Worldview. The Favicon Carries the Recognition.
The chulbuji.com logo experiment confirmed that the main logo and favicon should not be the same image — they are two elements of a brand system, each with a distinct role.
-
AI Resource Optimization Strategy v2.0 — A Practical Guide to Cutting Claude Token Costs
Draft with GPT and Gemini first, reserve Claude for advanced review and precise execution. A practical SOP for reducing Claude usage through role-based AI allocation.
-
Publishing the AI Collaboration Operating SOP v0
The basic operating standard for moving from learning AI to operating with AI has been formalized as a public document.
-
Year One as an AI Collaboration Operator
Three months into AI collaboration — taking stock of what's working, what still needs building, and the shift from learning AI to operating with it.
-
Why the AI Blog / AdSense Experiment Became a Dedicated Project
The AI Blog / AdSense experiment had been managed alongside the monthly operating Board — but running it over time made clear it needed its own project page as a long-term record hub on chulbuji.com.
-
First Day Using Google Vids in the Music Channel Experiment
The tool felt unfamiliar at first and took real effort to work through — but the video got finished and uploaded. A small but meaningful comparison experiment is now in place: a dynamic video alongside a static image video of the same track.
-
Auto-Trading v2.0 Upgrade Direction
Defined the v2.0 upgrade direction for a more production-ready system. Seven improvement areas: zero-order bug after BUY signal, available-cash API errors, high-priced stock handling, ratio-based capital allocation, market-strength-based entry count, DRY RUN vs LIVE mode clarity, and API retry/fallback. Built through vibe coding: expert AI for analysis, Claude Code for implementation.
-
Instead of Writing Code, I'm Designing an Auto-Trading System with AI
One thing that's become clear from building and running an auto-trading system: my job isn't writing code — it's asking the right questions, evaluating expert AI analysis as operational direction, and issuing execution instructions to Claude Code. In vibe coding, the human's role is to set the judgment criteria and keep the AIs on course.
-
Week 1 Auto-Trading Recap — Apr 29–30
Two live trading days. Weekly realized P&L: +₩40,658. Trailing stop hit three consecutive wins (zero stop-losses). Closing portfolio value ₩4,928,266 — down ₩13,941 from the week's start on unrealized P&L.
-
Auto-Trading Week 1 — Position Rotation Is Starting to Take Shape
A trailing stop sell → cash recovery → RS-qualified buy rotation structure appeared in live trading for the first time. Three consecutive trailing wins, zero stop-losses. The next challenges: high-priced stock handling, holiday detection, and fixing the export bug.
-
First Live Buy Day After the Upgrade — The System Sold and Bought on Its Own
Live account day 10. First real buy day after RS filter + trailing stop upgrade. Morning: 2 auto-sells (Samsung +0.84%, NAVER +2.31% / realized P&L +₩31,004). Afternoon: 3 new buys (SK Telecom + KB Financial, 12 shares total). 2 bugs fixed same day. Close: total ₩4,942,207 / cumulative P&L -₩39,796.
-
What a +0.84% Trailing-Stop Exit Taught Me About Rule-Based Automation
The trailing stop sold at +0.84%. For a moment, it felt like a waste. But that reaction is the problem — if emotional regret can override evidence that a system worked as designed, the reason for building the system in the first place disappears.
-
The Day the First Trailing Stop Profit Came In
Live account day 9. LG Chem's trailing stop triggered for the first time in range A (+3–6%) — auto-sold at ₩393,500, +2.88% gain. Total valuation ₩4,948,738, daily realized P&L +₩11,000. All holdings closed in the green.
-
When Should a Trailing Stop Be Activated in an Automated Trading System?
A fixed take-profit rule and a trailing stop work against each other — one should be removed. This is an operational review of the structural conflict found during live account operation and the three-tier zone redesign that replaced it.
-
KOSPI Hit a Record High — But My Portfolio Was Down
KOSPI hit an intraday record. 4 of my 5 positions were in the red. The system was filling MAX_HOLD=5 slots in order of qualification, not strength. Added an RS filter and removed forced slot-filling. Running BUY_DRY_RUN=True for 2–3 trading days.
-
Firebase Studio Shutdown — Migrating to Local Dev Environment
Confirmed Firebase Studio end-of-service. Migrated the chulbuji.com development environment to local VS Code on an LG Gram. Cloned the repo, confirmed the build, cleaned up the Astro 5 content config, and pushed to main. Local → GitHub → Cloudflare auto-deploy pipeline secured.
-
Filling Slots with Weak Stocks Is Not Diversification — It's Distributing Losses
MAX_HOLD=5 was a ceiling, not a target. When a system runs to fill a number, it picks available stocks, not strong ones. Four positions down on the day KOSPI hit an all-time high proved it.
-
AI Content Assistant MVP — From Dev Setup to Live Service in One Day
VS Code + Claude Code install, Next.js + TypeScript + Tailwind CSS project, GitHub + Cloudflare Pages pipeline, 8 screens implemented, OpenAI GPT-4o API connected, live service deployed — all in one day. 23 seconds from git push to production.
-
One Live Day — Fixed 5 Bugs and Found a Strategy Gap
Morning orders all failed, 18 duplicate KT sell reservations stacked, and a full regime detection redesign. 5 bugs fixed, 6 strategy changes, 4 orders executed. Seed ₩5M → close ₩4.89M.
-
The Stronger the Filter, the More Good Stocks It Blocks
What I built: Regime scoring system (5 conditions) + cash ratio logic tied to total assets. What broke: Wrong balance field caused all orders to fail; duplicate sell orders stacked 18 times. What I learned: The filter that blocks overbought stocks also blocks large-caps in recovery.
-
No Coding Skills, Live in One Day — Why Claude Design + Claude Code Made It Possible
One Claude Design Handoff feature closed the design-to-development gap. The result: an 8-screen app live in production in one day, without writing code.
-
Three Trading Bugs Found and Fixed in One Day
Three bugs in the KIS live auto-trading system — wrong cash field reference, 18 duplicate sell orders for KT, misplaced code fix. All found and fixed the same day. You don't need to read code to debug. Read logs, describe what's off, let AI trace the cause.
-
Read the Logs Before You Read the Code
You don't need to read code to find bugs. Today I fixed three auto-trading bugs in one day — by reading logs, describing what was off, and letting AI trace the cause.
-
AI Content Assistant for Small Businesses — From Idea to Screen to MVP Scope
An idea sketched into a business plan application became a mobile interactive prototype in a single day using Claude Design. v1 feedback led to v2, and it was only after seeing the screens that the MVP scope and dev environment direction became clear.
-
Auto-Trading Live Day 3 — 2 Stop-Losses, System Worked Correctly
Live day 3. KB Financial and Kakao hit the -3% stop-loss, -₩61,800 total. MA60 filter correctly blocked 3 downtrend stocks (NAVER, Kia, Hyundai Motor) — first real-money validation after the 4/21 API replacement. Budget safety margin: success at 13:00, failed again at 14:00. KT entered as new position. Closing balance ₩4,906,290.
-
The Day the System Said 'No'
Whether the system worked correctly became a more important measure than whether I made money. Live day 3 — what -₩61,800 confirmed.
-
Ideas Validate Faster on Screen Than on Paper
Writing an idea into a spec kept growing the feature list. Moving it to a screen changed the question entirely — from 'what features does it have?' to 'will users know what to do here?'
-
Auto-Trading Live Day 1 — 2 Bugs Fixed, 5 Stocks Bought
First day on the live account. Two morning bugs resolved (cached paper token / MA60 data shortage), first buys executed at 10:33. Five stocks — Kakao, Kia, KB Financial, Shinhan, NAVER — ₩4,690,300 deployed. End-of-day unrealized P&L: –₩54,030 (–1.08%), all 5 positions held.
-
Auto-Trading: 1-Week Paper Trade Complete — Switching to Live
7-day paper trading run (Apr 11–17) wrapped up. All core functions validated: regime filter, trailing stop, 5-slot management. 4 bugs found and fixed same-day. Final portfolio: ₩10,058,486. Weekly P&L: +₩45,060 (+0.58%). Live trading starts Monday 08:50.
-
The Day the System Did Nothing Was the Most Important Day
On day one, the system did nothing. That wasn't a bug — it was the right answer. What one week of automated trading taught me wasn't how to buy. It was how not to.
-
Auto-Trading System: First Live Week — Apr 13–16 Operations Log
Four days of first live operations. DRY RUN validation → first buy → first trailing stop trigger → lock file failure recovery → universe expanded to 20 stocks. Weekly realized P&L: +₩15,000. Total portfolio: ₩10,037,479.
-
The System Sold Without Me — Trailing Stop's First Live Trigger
April 15, 2:51 PM. Samsung Electronics was sold automatically. I didn't tell it to. The trailing stop fired in live trading for the first time. I'm verifying whether the strategy is moving as designed.
-
KIS Auto-Trading — First Live Order Filled, Live Transition Complete
10:30 AM. Kakao, 20 shares @ ₩48,350. The system that ran in DRY_RUN mode is now live. Three fixes before the open, trailing activation logs and Telegram position summary in the afternoon, --live flag added to the scheduler.
-
The Stocks That Did Nothing Proved the Strategy Today
The real result today wasn't the one stock I bought — it was the three that got blocked. The MA20 deviation filter stopped SK Hynix at 16.9%, LG Chem at 7.2%, Samsung SDI at 9.4%. A day the system quietly prevented the wrong buys.
-
Website Restructuring — Shifting to a Record-Asset Architecture
Two months after the rebrand, records started piling up and the structure needed to follow. A log of the transition from intro site to long-term asset architecture.
-
KIS Auto-Trading — Two Hours, 11 Items Fixed
Yesterday evening, two hours. Fixed a cash balance field bug, added retry logic, duplicate prevention, improved RSI, and introduced a trailing stop. A confirmation that systems can change this much without writing the code yourself.
-
I Didn't Have AI Fix the Code — I Redesigned the Structure
I shifted from chasing bugs one by one to a structure that separates judgment from implementation. My coding ability didn't change. What changed was where I focus.
-
KIS Auto-Trading — Module Cache Issue Identified & Buy Cycle Restored
The code was updated, but the process was still running on yesterday's state. Found the root cause, patched the MA60 gap in the paper trading environment with a fallback, and restored the DRY_RUN buy cycle.
-
I Fixed the Code. Why Didn't the System Change?
Editing the code and having those edits actually take effect are two different things. What I learned from running an automated trading system wasn't strategy — it was operations discipline.
-
Auto-Trading v1.1 — Operations Setup & First Run Standby Log
Not a day without trades — a day the system correctly refused to trade when it shouldn't have.
-
Auto-Trading MVP Complete — Rule Engine, Regime Filter & Telegram Alerts
KIS API integration, rule engine, loss-limit safety guard, Telegram alerts, and scheduler all done. First automated run starts tomorrow.
-
What I Didn't Know Until I Built an AI Trading System
I didn't write a single line of code. But the skill I used most today wasn't coding. It was the ability to precisely instruct 'what to build.'
-
I Built the House. I Just Wasn't Living In It.
Consistency isn't a willpower problem — it's a structure problem. I realized that only after counting twelve unfinished things.
-
RunBuddy Coaching MVP First Field Test — Start Button, GPS Judgment & Coaching Interval Fixes
After implementing the coaching-focused MVP, I ran a 500m indoor test. Identified three issues: unresponsive start button, mismatched voice coaching due to missing GPS data, and a coaching interval mismatch with the test distance. Each root cause was isolated and fixed. The patched build gets validated in an outdoor run tomorrow.
-
There's No Perfect Prompt. Only the Sense of Keep Moving.
What making the ChulbujiRunning music video taught me again. The more I tried to pack into a single scene, the more it fell apart. AI collaboration isn't about finding the right sentence — it's about breaking scenes down, adjusting the flow, and pushing through to the end.
-
The App Was Done. But I Couldn't Trust the Data.
Cadence hit 3,648 spm. Pace swung between 3 and 14 min/km. The app ran — but the numbers couldn't be trusted. A day of fixing 5 bugs by directing multiple AIs, one step at a time.
-
Today I Ran — and My App Talked Back
I don't know how to code. I've never written a single line myself. But today, the app I built spoke to me. A field test record of Chulbuji RunBuddy.
-
I Shipped Tetris Without Writing a Single Line of Code
4 hours. Questions only. A first-year AI builder's first deployed game — bugs included, fixed in 10 minutes.
-
Clear Roles Make AI Collaboration Work
The reason AI tools hit a wall wasn't the tools — the roles were blurred. A record of the day I designed a production pipeline where three emotional keywords are all it takes to get content flowing.
-
I Broke Bricks with AI — Vibe Coding 'Breakout' + Playground Story
A lunch-break retrospective from someone who built a game without writing a single line of code. From 1-file magic to Z Fold finger feel.
-
5 Weeks of Transformation: How I Became an AI Product Builder
A 5-week record of a human transforming into an AI Product Builder in the age of AI civilization
-
2 Minutes of Reading, Zero Clicks — Why I Changed One Button
Visitors were reading for an average of 2 minutes, but not a single click on the CTA. Here's why I changed 'Start for Free' to 'Organize My Thoughts for Free' — and the hypothesis behind it.
-
Just Two Songs, But I Built a House
I promoted my channel with only two songs. Strangely, instead of feeling embarrassed, I felt alive.
-
Hackathon D-2: I Used to Guess. Today I Started Seeing 'Clicks'
A solo maker who can't write a single line of code successfully tracked button clicks with GA4+GTM two days before the hackathon. Chulbuji designed what to measure, Meta-chulbuji implemented it. Now I don't guess — I prove.
-
From Morning Ad to First Song: A 2-Hour Rocket Jump with Meta-chulbuji
From seeing an ad and thinking 'I want to try this' to asking Meta-chulbuji, executing through mistakes, and completing my first song 'Chuno: The Outcry' with a channel launch in just 2 hours
-
Learned for 5 Years. Now It's Time to Build: Evolving from DX to AX, from User to Builder
I share my experience of escaping 'Tutorial Hell' and launching my first AI product, the 'Meta-chulbuji Report,' by collaborating with AI as a 'Plus-Human'.
-
Code Isn't Business Until the Receipt Arrives
72 Hours of Monetizing an AI Service - The journey from localhost greenhouse to the wilderness of production where real money changes hands
-
I Argued with Two AIs, and I Lost
The operational strategy I learned from the 'Safety vs. Production' debate
-
I Started Because I Wasn't Perfect
My journey with JoCoding Hackathon and AI. Transforming hesitation into action, incompleteness into flow
-
The Day I Ditched Long Prompts, and Tokens Came Alive
AI collaboration lessons learned from chulbuji.com rebuild: work small, be precise
-
[Week 1 Log] From Zero to Working Website with AI Collaboration
A beginner's journey escaping from broken link hell through vibe coding
-
From Lottery Generator to Record Hub — The Full Story of Rebranding chulbuji.com
A semiconductor engineer with 34 years of experience documents the complete process of rebranding chulbuji.com to grow as a Product Builder alongside AI.