Hi. Before I tell you anything about myself, let me show you the world exactly as it stands. It takes about ninety seconds. It will explain every choice I've made. In 2025, researchers at Uppsala University counted 65 armed conflicts involving states. It is the highest number recorded since the counting began in 1946. Roughly 244,600 people were killed by organized violence that year. Only one year in the past three decades has run higher — 2022. 117.8 million people are currently displaced from their homes. That is one person in every seventy alive today. 2.1 billion people have no safe drinking water at home. 106 million still drink directly from rivers and lakes. And the lakes themselves are leaving. And when disease outruns us, the answer is slow by design. A single new medicine takes more than a decade to build, at a cost measured in billions, and most candidates die before they arrive. Hold those numbers. Because through those same years, something else was moving, in the opposite direction, at a speed almost nobody priced in. In 2019, the best AI system could hold a task for about two seconds. By 2023, four minutes. This year, around twelve hours, unsupervised. That capacity has doubled roughly every seven months. Which means the oldest excuse in the world, that solutions are too expensive and too slow to even attempt, is expiring in real time. So who is supposed to pick these problems up? Engineers. Students. Builders. People like me. Possibly people like you. But look at what most of us are actually doing. Competing to memorize syntax, and queueing for the same junior positions, because that was supposed to be the safe path. Except the safe path has quietly collapsed. Software developer postings have fallen roughly seventy percent from their 2022 peak. Employment for developers in their early twenties is down nearly twenty percent. New graduates in the field now face some of the highest unemployment of any major. And yet everyone in that queue, and outside it, is carrying ideas. Products. Companies. Cures. You are carrying a few right now. Almost none will ever be tried. Not for lack of tools. The tools have never been this capable. For lack of nerve. This is where untried ideas go. They pile into sediment. Doubt presses them flat. Waiting hardens them to stone. A wall of everything the capable were too careful to attempt. I refused to add mine to it. I am not from the queue. I began as a writer, and a writer learns one thing early: the exact right words can conjure an entire world. So while everyone raced to write code by hand, I made a different bet. I stopped writing syntax, and learned to direct the systems that write it. I built my own methodology to do it, and I trusted it with everything. So I break it. This is me. A writer who directs a thousand agents at problems that matter, one problem at a time. Everything below is the evidence.
I think in the space of possibilities
I'd rather perfect a plan than improvise
GHAISAN KHOIRUL BADRUZAMAN
AI SYSTEMS ARCHITECT
MEANT TO BE FOUNDER
An animated schematic of the Crown methodology. A delivery chain across four swimlanes — Operator, Crown, Design, and Orchestrator — passes artifacts (a blueprint, an orchestrator prompt, an env-prep file) between lanes. Inside the Crown lane, a session runs three modes — Brainstorm, Refine, and Another Orchestrator — whose decision branches converge into a second research pass and author the deliverables. An execution spine then walks phases P0 through P8, each guarded by a hard gate G0 through G8 with commit checkpoints; failed gates arc back into repair loops, and a P4 root-cause escalation folds back into gate G3. The P1 and P3 fan-out blooms into roughly one thousand parallel agent workers in a layered, densely interconnected network, then resolves into a durability state machine: Running, Checkpointed, Stalled, Recon, and resume or rollback. One intent, a thousand agents — AI agents are the future, and I direct them.
- One intent in. A shipped product out.
- The plan is argued to the ground before a line is written.
- Every phase earns the next — no gate, no passage.
- One intent. A thousand agents.
- Interrupt it; it checkpoints, then carries on.
- AI agents are the future. I direct them.
Tools change every few months. The way I think only improves.
I don't write code. I deliver the structure of thought to the machine.
The evidence, not the assertion. METR · 50% task-horizon — the length of task an AI can hold
As the timeline scrubs from 2019 to 2026, black ink floods the cream page: its area tracks the METR task-length horizon — the length of task a frontier AI model completes with 50% reliability. That horizon runs from about two seconds in 2019 to about twelve hours in 2026, doubling roughly every seven months. SWE-bench Verified capability enters the series in 2024 and crosses 80% in November 2025 with Claude Opus 4.5 at 80.9. Extrapolation beyond today is contested. Full figures and sources follow in the data table.
Proof data and citations (text equivalent), AS OF Q2 2026
| Model | Date | Score (%) | Type | Source |
|---|---|---|---|---|
| Claude 3.5 Sonnet | 2024 | 49 | measured | anthropic.com |
| Claude Opus 4 | 22 May 2025 | 72.5 | measured | anthropic.com |
| Claude Sonnet 4.5 | 29 Sep 2025 | 77.2 | measured | anthropic.com |
| Claude Opus 4.5 | 24 Nov 2025 | 80.9 | measured | anthropic.com |
| Claude Opus 4.7 | 16 Apr 2026 | 87.6 | as reported | anthropic.com |
| Claude Opus 4.8 | 28 May 2026 | 88.6 | as reported | anthropic.com |
| Model | Date | Horizon | Seconds | Source |
|---|---|---|---|---|
| GPT-2 | 2019 | ~2 seconds | 2 | arXiv:2503.14499 |
| GPT-4 | Mar 2023 | ~4 minutes | 239 | arXiv:2503.14499 |
| Claude 3.5 Sonnet | Jun 2024 | ~11 minutes | 684 | arXiv:2503.14499 |
| Claude 3.5 Sonnet (new) | Oct 2024 | ~20 minutes | 1231 | arXiv:2503.14499 |
| o1 | Dec 2024 | ~39 minutes | 2330 | arXiv:2503.14499 |
| Claude 3.7 Sonnet | early 2025 | ~50 minutes | 3000 | arXiv:2503.14499 |
| Claude Opus 4 | May 2025 | ~100 minutes | 6022 | metr.org |
| o3 | 2025 | ~110 minutes | 6600 | metr.org |
| Claude Opus 4.5 | Nov 2025 | ~4.9 hours | 17580 | metr.org |
| Claude Opus 4.6 | Feb 2026 | ~12 hours | 43200 | metr.org |
A pen taught me that the right few words can bloom into an entire world.
I still build the same way, a world conjured from the exact right words.
EVEN MY IMPROVISATION IS PLANNED
I weigh every path before I commit to one.
I optimize for the outcome, not the impulse.
Iterate, converge, then ship.
Split the problem; let the parts solve themselves.
Real geometry under the grain.
The same bet I make directing agents — hold the exact intent, and the chaos resolves into structure.DRAG TO ROTATEDirected, not hand-written.
Every project below was carried from idea to production by directing AI agents. The anchor: NERIUM, built with Opus 4.7 and featured in Anthropic's hackathon gallery.
Nine products. One eye for surface.
Marginalia
Editorial1 screenSELF-DIRECTED STUDYLoop
Messaging1 screenSELF-DIRECTED STUDYDrift
Marketing1 screenSELF-DIRECTED STUDYMeridian
SaaS dashboard7 screensSELF-DIRECTED STUDYForma
Commerce8 screensSELF-DIRECTED STUDYCadence
Onboarding5 screensSELF-DIRECTED STUDYTally
Finance5 screensSELF-DIRECTED STUDYDaybook
Day-planner4 screensSELF-DIRECTED STUDYMy proudest work.
Built with Opus 4.7 at Anthropic's Cerebral Valley global hackathon; selected among ~500 builders from roughly 20,000 applicants, and featured in the Built-with-Opus gallery.
1st Place — Hackathon Digital Cooperatives 2026. Ministry of Cooperatives, Republic of Indonesia, with PEBS FEB Universitas Indonesia. Jakarta, 11 July 2026.
An AI watchdog that audits Indonesia's village cooperatives on behalf of their members.
Built to be given away. The economic rights now belong to the Republic; the moral right — to be named the author — stays mine.
- METRO TVmetrotvnews.com/read/bD2CwO38-hackathon-digital-cooperatives-2026-lahirkan-3-solusi-terbaik-dorong-transformasi-koperasi-indonesia
- KEMENTERIAN KOPERASIhackathon.simkopdes.go.id
- SUARA.COMsuara.com/tekno/2026/07/16/135733/hackathon-digital-cooperatives-2026-cetak-inovasi-ai-untuk-percepat-digitalisasi-koperasi-indonesia
- STARTUP HUB INDONESIAnot linked
- INEWS TANGSELtangsel.inews.id/read/702383/kemenkop-dorong-talenta-muda-transformasi-wajah-koperasi-lewat-teknologi
- BLOOMBERG TECHNOZbloombergtechnoz.com/detail-news/114758/hackathon-digital-cooperatives-2026-cetak-tiga-inovasi-terbaik
- DAILYSOCIALnews.dailysocial.id/post/wajah-baru-koperasi-digital-dari-hackathon-kemenkop-2026
- ALGORITMUnot linked
- RRIrri.co.id/jakarta/hiburan/2562823/kemenkop-umumkan-pemenang-hackathon-digital-cooperatives-2026
Pramana asks; it does not accuse.
Execution reach
“Aerodynamics are for people who can't build engines.”
Mission
“The best Ferrari ever built is the next one.”
Let's build something that shouldn't exist yet.
Founders — bring an intent worth shipping. Engineers — bring your hardest question about directing agents. Clients — bring a product that needs to exist. And the merely curious — the channel is still open.
ghaisan.khoirul.b@gmail.com- THE PROLOGUE
- HERO
- CROWN TITLE
- AGENT GRAPH
- THESIS
- PROOF
- WRITER
- MIND AS MATH
- CRAFT 3D
- THE WORK
- DESIGN RANGE
- TWO SEALS
- EXECUTION REACH
- MISSION
- CONTACT








