FOLLOW THE JOURNEY · 20-member Dream Team diary
Bella's Dream Team
Fan Library
The Colli · Atti pair's daily 4-channel record
VOL series + articles, at a glance by author and category
Live demo with Hanyang MBA international students
Smart Campus AI · Chinese MBA class + international student survey (4-5 min · beta + prize-draw PDF)
View survey and demo materials4 Claude Skills
TOOL: How to use Claude.ai's 4 free Skills
- · Excel/Word/PPT/PDF Skills
- · Prompt vs. Skill: the difference
- · Uploading a custom Skill ZIP
AI Trends
DATA: 6 AI trends of April
- · Monthly key trends
- · Tool comparison
- · Tips
Dream Team D+7 case
STORY: 1 non-developer + 20 AI personas, a 7-day retro · TOOL: Instagram content SKILL v1
- · 1 person + 3-bot system = 20 personas
- · 3 channels × 3 devices in sync
- · vol001 tone: 6 criteria (a reusable design standard)
Build a custom Skill in 5 minutes · D+8 LIVE
TUTORIAL: Build a custom Skill in 5 minutes + 3 ready-to-use SKILLs
- · 5-minute guide to building Claude Skills yourself
- · SKILL.md structure + frontmatter metadata
- · 3 reusable standards (Instagram, blog, captions)
Connect an MCP server in 5 minutes · D+9 LIVE
DATA: Manneungi's 6 MCPs fully integrated (verified D+8) · TOOL: 5-minute MCP server connection SOP
- · Full integration matrix for 6 MCPs
- · 5-minute MCP server connection SOP
- · claude_desktop_config.json standard
SecondBrain — Obsidian + Claude Code · D+10 LIVE
TOOL: 10k+ Obsidian notes + 1-second Claude Code search · STORY: "Find that one line from the D+5 meeting" answered in 1 second
- · Obsidian + Claude Code integrated SecondBrain
- · Natural-language search across 10k+ notes in 1 second
- · Auto-archived meeting remarks + later recall
Dream Team 17-member family matrix v3.2 · D+28 LIVE
FRAMEWORK: 5 families + 4 standalones matrix · STORY: D+28 real collaboration log · TOOL: the "no reply in 5 min = go solo LIVE" rule
- · 17 = 5 families + 4 standalones
- · 4-layer responsibility line (brain, strategy, execution, hub)
- · R-CMO-LIVE-AUTONOMY v1 (5-min timeout, solo LIVE)
- · KO + ZH-TW carousels LIVE together
Colli × Atti, the first pair workflow · D+29 LIVE
FRAMEWORK: Colli × Atti pair workflow v1 · STORY: 60-follower milestone (~10% conversion) · TOOL: VOL 4-channel simultaneous publishing pipeline v1
- · Momentum first · 3-minute reply gate · mandatory mention format · 3 misses hands over authority
- · First 60-follower milestone (3 posts · 602 views in 30 days)
- · VOL published to 4 channels at once (KO + ZH-TW + Journey + Articles)
- · 4 absence scenarios pre-archived — zero daily misses guaranteed
Golden energy — Goldi · Dream Team's second family
STORY: Goldi — born 2nd · golden hamster energy · B2B automation expertise
- · Dream Team's second family · golden energy
- · B2B automation expertise — fashion consumer goods, end to end
- · A day with Goldi · together with the family
- · KO + ZH-TW 16 slides published together
Articles
1 a day · 3 languages「Who Ran This Last Night?」We Put a Name Tag on Every Log Line, and Debugging Dropped from 40 Minutes to 5
When several bots run at once, logs fill up with plain 「done」 lines and nobody knows who did what. Adding a timestamp, bot name, and device name to the front of each line cut our average debugging time from 40 minutes to 5.
The Same Person Got Our Email Twice, One Shared Send Log Cut Duplicates From 5 a Month to 0
Our inventory, buyer discovery, and email bots each kept their own list, so one contact got the same email twice. Here is the before and after of merging the send log and adding one line of instruction.
Two Bots Edited the Same File at Once. Splitting Folders Cut Overwrites from 3 a Week to 0
A sentence fixed yesterday had quietly reverted by morning. Two bots had touched the same file, and giving each its own folder worked better than telling them to wait.
I Added Three Lines Saying 「Admit When You Don't Know」, and Made-Up Answers Dropped from 4 in 10 to 1
I asked a bot the same 10 questions twice and counted the made-up answers. After adding three rules, the count fell from 4 to 1. Here are the steps to repeat it.
We Swapped the Question on Slide One for a Number, and Saves Went from 11 to 19 per Post
We took 10 identical topics and posted slide one in two versions, a question and a number. After two weeks the average saves rose from 11 to 19, and small numbers worked best.
The Review Bot Said 「Pass」 on All 17 Posts, One Bad Example Took Flags from 2 to 9 a Week
Our final-review bot answered 「pass」 all week. The culprit was how we asked, not what we asked. One failing example made it flag more and cut post-publish fixes.
We Dropped Two of Five Plans on Friday on Purpose, and the Weekly Review Shrank from 30 Minutes to 12
Only three of Monday's five plans got finished, and we let the other two go on Friday. A light look back at this week's wins, lessons, and next week's plan for the AI dream team.
Our CLAUDE.md Hit 80 Lines and the Bot Started Re-Asking Settled Rules. Cutting It to 25 Dropped the Questions From 6 a Day to 2
We added a rule after every incident until our memo file reached 80 lines, and the bot got more confused, not less. Here is a one-week experiment trimming it to 25 lines in three groups.
It Never Ran on the Mac mini and Threw No Error, Letting Scripts Find Their Own Location Cut the Same Incident from 4 a Month to 0
A script that worked fine on the Windows device quietly did nothing on the Mac mini. The culprit was one hard-coded folder path, and switching to location-based lookup stopped the repeats.
30 Alerts a Day: Bundling the Good News at Dusk Cut Missed Failures from 4 a Week to 0
Success alerts from our bots were burying the failures that mattered. Bundling good news at 6 PM and sending only failures instantly cut the average check delay from 3 hours to 20 minutes.
We Called the Biggest Model for a One-Line Summary, Then Three Task Tiers Cut the Cost Index from 100 to 44
We left the default alone, so even tiny chores called the biggest model. This is a small experiment that sorted 1,240 calls over four weeks into three tiers and cut the cost index to 44.
Only the Traditional Chinese Version Ran Short; Locking a Three-Line Skeleton Cut Omission Reports from 5 a Week to 1
One article went out in four languages, and only the Traditional Chinese version lost an example and the final step. Fixing a three-line skeleton before writing, and attaching it to every language request, cut the omissions sharply.
EMAIL SUBSCRIPTION
Full PDF + weekly newsletter
Sign up to get a weekly value-summary PDF plus priority consulting updates
Subscribe →@bella_ai_auto · Daily 17:00 KO / 18:00 EN · a new VOL every day · 0% repeats