Muratcan Koylan

Member of Technical Staff at Sully.ai; joined in 2026 as a Context Engineer working on clinical AI. Previously AI Agent Systems Manager at 99Ravens AI.

Background: Communication Design (HCI) and seven years in B2B marketing before engineering. Most of what I think about sits at the boundary between AI systems and the people meant to use them.

Open to advisory and speaking. Email muratcan.koylan@outlook.com or DM on X.

Current

Projects

Agent Skills for Context Engineering

Dec 2025 · Open-source collection of transferable Agent Skills for building production AI agent systems on Claude Code, Cursor, or custom stacks.

Long-Horizon Prompting

Jul 2026 · A specification method and Agent Skill for launching autonomous runs that span hours, context windows, or parallel workers, with explicit success predicates, non-counting outcomes, evidence requirements, and adversarial review.

Ralph Wiggum Copywriter

Jan 2026 · Claude Code plugin: autonomous multi-agent copywriting loop. Learns your voice, self-critiques, and rewrites until acceptance criteria pass.

digitalOS (Personal Brain OS)

Feb 2026 · A Git-based personal operating system for AI agents: 80+ plain-text files, no database, ~87% fewer tokens per task through progressive disclosure.

See all AI projects →

Writing

Long-Horizon Prompting: Specs for Autonomous Agent Runs

The File System Is the New Database: How I Built a Personal OS for AI Agents

SkillOpt: Gradient Descent for SKILL.md Files

Context Engineering: A Marketer's Guide | Zapier

Your Experts Won't Train Your AI, You Have to Interview Them | 99Ravens

Prompts Experiments with Kimi K2 Thinking

See all writing →

Lessons

"Works in the notebook" means nothing.

Shipped a multi-agent system that passed every test, then crashed in production from memory leaks. Load test, and understand exactly how state persists.

Start with the dumbest thing that works.

Built a seven-agent system for a two-agent problem. Every agent has to justify its existence.

A prompt that leans on model quirks is technical debt.

Over-optimized a prompt for one model and it broke on another. Favor structure over per-model cleverness.

The architecture you pick at 100 users compounds at 10,000.

JeezAI reached 5,000 users but was architected for a demo, not for scale.

In fast markets, speed beats perfection.

Held the Butterpath idea too long while looking for validation.

Two things at 100% beats five things at 60%.

Ran everything at once and burned out.