Yuma@axis.
@null_founder
Followers
79
Following
860
Media
32
Statuses
743
Shipping dev tools after 25yrs kitchens & bankruptcy at 40. The voice says stop. I ship anyway. Managed by 3 AIs. None of them agree. https://t.co/WAS0ylDVie
Osaka, Japan
Joined March 2026
The part most people miss: QA must be independent. If the same AI that wrote the code reviews the code — that's a chef tasting his own dish and telling the customer it's perfect. Different model. Different vendor. Different bias. https://t.co/gYrIZpKkt3
note.com
I run 3 projects simultaneously. Every day. With 5 AI roles reporting to me. Not 5 browser tabs open. Not 5 chat windows with the same prompt pasted in. 5 distinct roles — each with a defined...
1
0
0
The part I'm proudest of is the part that failed. "Mistakes Made: The Settings UI agent was given the IpcResult type definition but still produced incorrect access patterns." The PM logged it. Then wrote a rule so it never happens again. Every gate makes the next gate smarter.
0
0
0
Every feature goes through a numbered GATE. Each GATE has pass conditions. Not guidelines. Conditions. The PM checks every one. CEO (me) verifies. If even one fails, the gate doesn't open. This was GATE-8. 13 conditions. 4 commits. Build clean. The gate system isn't quality
1
0
0
The product is Grimly. Revenue Forecast, ASO Optimizer, Pitch Deck Architect — five tools under one roof. Built with Electron + TypeScript. Local-first. No cloud dependency. 33 days ago I was mass-producing gyoza. Now I'm mass-producing settings screens.
1
0
0
GATE-8. 13 conditions. All passed. And the PM still filed a "Mistakes Made" report on itself. No one asked it to. It's in the spec. The agents that build the product also build the postmortem. Every gate. Every time. The green checkmarks are nice. The red ink underneath is
1
0
0
Most people use AI as a chatbot. I use it as a development organization. Human direction stays central. AI agents operate under explicit roles, boundaries, and audit discipline. That screenshot above? One spec. One command. 11 tasks. 6 minutes. The models aren't special. The
0
0
1
I don't optimize for novelty. I optimize for: - speed - stable quality - repeatability - no collapse across parallel projects - no degradation in human judgment If a multi-agent setup doesn't satisfy all five, it's just unnecessary complexity.
1
0
0
Why semi-automated, not full auto? Because multiple agents create new problems: - overlapping edits - silent contradictions - pseudo-consensus without real control Workers get disjoint write sets. If two agents must touch the same file, only one owns the final edit. That's the
1
0
0
The role split: PM — orchestration, gate control, pass/fail Wizard — spec interpretation, architecture, boundaries QA — independent audit (never writes code) Senior Engineer — high-risk shared modules Worker — bounded tasks, narrow write scope One role, one kind of decision. If
1
0
0
My AI team just ran an 11-task sprint in 6 minutes. No standup. No Slack thread. No "let's circle back." ENGINEER handles implementation. WIZARD handles prompt architecture. PM handles verification. Same models everyone else uses. Different org chart. 32 days ago I was
1
0
1
Correction: the fix I posted earlier didn't cover all paths. -webkit-app-region: drag allows double-click maximize through the OS — even with frame: false. This bypasses custom setBounds and re-triggers DPI double-scaling. Fix: maximizable: false This blocks all native
0
0
0
I’m not claiming this is complete or official. This is a best-effort reference built from publicly verifiable material. If Anthropic publishes a full reference later, I’ll happily replace this with the official one.
0
0
0
Claude Desktop Cowork just dropped, and almost nobody has documented what it actually does. No real reference. Just scattered launch material. So I made one. A 14-slide breakdown of the features, limits, and roadmap signals I could actually verify. What Cowork is: → A
1
0
0
Your AI has no job title. That's why it keeps breaking things. I gave mine five: PM. Wizard. QA. Senior. Worker. Each one has a lane. None of them cross it. I learned this washing dishes at 15. Took me 25 years to realize it scales.
0
0
0
The irony isn't lost on me. I'm building a tool that helps apps survive Apple's review process. And the reason I'm delaying it is because Apple's review process just shifted underneath everyone's feet. Including mine. That's not a setback. That's the product telling me it's
0
0
0