Here’s what I did, here’s what I learnt (including what broke first), and here’s what you might be able to learn — if you want.
First, a quick baseline on who is writing this: I am not a computer scientist or a software engineer by trade. My career arc is 14+ years in Live Events & Frontline Technical Operations — starting hands-on on the tools as an Audio Visual Technician and Venue Director, moving up to managing 250+ technicians across 50+ venues, and now governing workforce rostering, compliance, time & attendance, and labour forecasting across multiple countries.
When modern AI tools arrived, I noticed most people were doing one of two things: either typing one-off questions into a blank chat box and getting a generic answer, or watching two videos on a Tuesday night, declaring themselves an “AI Expert” on Wednesday morning, and trying to sell people a fragile prompt cheat-sheet.
Neither of those interested me. In live events and technical production, a show doesn’t run on wishful thinking — it runs on a clear stage plot, labelled cables, distinct crew lanes, and a pre-show check. So I ran an experiment to see what happens when you treat AI the exact same way.
“If a regular operator stops treating AI like a magic chat window and instead puts it inside a real production harness — with a numbered folder map, written standing rules, and a 5-second human quality check — what can you actually build and ship?”
Instead of writing theoretical opinions about AI, I used this folder-and-rule method to build, ship, and run three live public experiments on agimstudio.com:
Built an automated short-form comedy and slapstick animation studio around a single character (AGIM — “Good vibes, deadpan optimism, oversized enthusiasm”). Every short follows a strict 4-Beat Slapstick Formula: Peaceful Illusion → Over-Engineered Contraption → Kinetic Physics Betrayal → Serene Deadpan Thumbs-Up.
Built a live 60 Hz co-op 3D browser game where players chop timber, herd cows, drive vehicles, trade at village wall-shops, and build out a homestead with zero downloads. Runs 24/7 from a local Mac Studio over a Cloudflare Tunnel with live Telegram player & bug telemetry.
Built 20+ playable browser games split cleanly into two hubs: Agim Arcade for solo/invite play (featuring Agim Powder Toy falling-sand cellular automata with WASD & Arrow elemental stickmen, Shed Snake, and async Chess & Connect 4) and Agim Party for live TV + phone group play with an Eliminated Mini-Arcade Lounge.
Consolidated 5 scattered project folders into one numbered workspace spine (00_Ingestion → 10_Brand → 20_Blueprints → 30_Code → 50_QA_Checks) governed by a plain-English CLAUDE.md rulebook and an automated integrity verifier that scans every file in seconds.
| Experiment Pillar | What Was Built | How It Was Governed | Measured Public Outcome |
|---|---|---|---|
1. Daily Animation@agim.ai |
Short-form comedy & kinetic physics animation pipeline | 4-Beat Slapstick Formula + persistent schedule memory DB | 2.50M+ organic views • 1,500+ posts • 300+ runway |
2. Boundary Inc.boundaryinc.agimstudio.com |
60 Hz 3D multiplayer WebGL browser frontier sandbox | 3-Lane Parallel Orchestrator (Feature / Polish / QA worktrees) | Live 24/7 multiplayer world • $0/mo server bill |
3. Agim Arcade & Partyagimstudio.com/games |
20+ games: Powder Toy sandbox, Snake, async Chess/C4 & TV Party | Automated 6-player & physics suite + TV/Phone role split | 84 / 84 automated checks green • Live invite sync |
4. The Studio Harness00 to 99 Spine |
Unified numbered folder architecture + CLAUDE.md constitution |
“Fix It Once, Fix It Forever” rule harvester + secret scanner | Zero leaked keys • Works across any AI model |
Every rule in this studio exists because I got it wrong first and paid for it in wasted evenings. Below is an interactive look at the core difference between a Bare Chat Window and a Folder Harness, followed by the five biggest lessons from the experiment.
RULES.md & 20_Blueprint/ first: The AI immediately knows the folder map, the visual colours, and every past mistake before writing a single line.
30_Build/ without touching unrelated pages or breaking the live server.
50_Checks/. If an edge case appears, it writes Rule #14 into RULES.md so the next session never repeats it.
RULES.md), and a map of where files live — finishes the race every single time. And when a better model drops six months later, you simply unbolt the Corolla engine and drop the Ferrari engine into the exact same Harness.
Bug C, and in the process it would accidentally break Feature A or alter the styling on Page B. Worse, a fresh chat session two days later would make the exact same mistake all over again.
CLAUDE.md, and (3) Add an automated check in 50_QA so it can physically never regress.
00_Ingestion → 10_Brand → 20_Blueprints → 30_Apps → 50_QA), the AI can inspect the plan, check the brand palette, write the feature, and run the test in one continuous pass.
00 to 99. Nothing loose sits in the root folder except README.md and CLAUDE.md.
main in a single batch.
Create one project folder on your computer with five numbered subfolders so you and the AI always know where things belong:
mkdir -p My_Project/{00_Inbox,10_Rules_and_Style,20_Blueprint,30_Build,50_Checks}
# What goes where:
# 00_Inbox/ -> Rough ideas, voice notes & links
# 10_Rules_and_Style/ -> Colours, tone & examples you like
# 20_Blueprint/ -> PLAN.md (the 1-page source of truth)
# 30_Build/ -> The actual code, chapters or files
# 50_Checks/ -> Simple test script or checklist
# RULES.md -> Your 1-page constitution (Step 2)
Instead of typing a 500-word prompt every session, start each session in your project folder with these three sentences:
Read RULES.md and 20_Blueprint/PLAN.md first so you have the full project context. Today we are working on: [describe 1 specific task]. Before editing any files in 30_Build/, tell me your 3-step plan and how we will verify it in 50_Checks/ without breaking existing features.
RULES.md Constitution Template (Save in Project Root)
Save this file as RULES.md (or CLAUDE.md / AGENTS.md) at the very top of your project folder. Every time the AI makes a mistake, add one line under Section 4 so it never makes that mistake twice:
# RULES.md — Project Constitution & Harness ## 1. What We Are Building (Commander's Intent) - **Project Name:** [Insert Project Name] - **In One Sentence:** [What this project does and who it is for] - **The 5-Second Test:** A first-time visitor must be able to [do X] within 5 seconds without needing a manual. ## 2. Folder Map (Where Everything Lives) - `00_Inbox/` — Rough ideas and unapproved notes only. - `10_Rules_and_Style/` — Visual look, tone of voice, and reference examples. - `20_Blueprint/PLAN.md` — Our Single Source of Truth plan. Update this when scope changes. - `30_Build/` — All working code, pages, or chapters. - `50_Checks/` — Automated tests and pre-release checklists. ## 3. Standing Rules for the AI 1. **AI Proposes, Human Approves:** Read `20_Blueprint/PLAN.md` before making changes. Never silently delete working features. 2. **Keep It Simple:** Prefer clean, plain solutions over complicated frameworks. 3. **Zero Hardcoded Secrets:** Never write private API keys, passwords, or personal data into code files. ## 4. "Fix It Once, Fix It Forever" Logbook *(Whenever a bug happens, we write a 1-sentence rule here so future sessions never repeat it)* - **Rule 01:** Always check that mobile phone screens and desktop screens both work before finishing. - **Rule 02:** [Add your next lesson learnt here...]
If you can’t explain it simply on a napkin yet, don’t ask the AI to write code yet — ask it to interview you until the 1-page plan in 20_Blueprint/ is clear.
Show it to your partner, a mate, or a coworker for 5 seconds. If you have to lean over their shoulder explaining where to click, cut the clutter.
Don’t just patch the symptom in chat. Add one sentence to Section 4 of RULES.md so your folder gets smarter every week.
Everything below is live in your browser right now — no downloads, no paywalls.