Build33 pieces
Notes from building things myself
/
ETHval: From Calculation to Validation
Version 1.0 focuses on testing assumptions and limitations rather than the numbers valuation models produce. It distinguishes past predictions from market reactions and revisits the basis for valuing Ethereum.
[Hashed] AI Ecosystem Growth Intern
We are looking for an AI ecosystem intern to help run Hashed Nitro's programs, grow its community, and work across content and partnerships.
When Memory Becomes Action
MemKraft has grown from retrieval and storage, through relationships, time, provenance, and learning from failure, into a lifecycle system that connects only verified memories to safe actions.
One Smartest Model Wasn't Enough
More than model intelligence, what matters is a work loop that places human choices, stopping decisions, and verification between Claude's exploration of unclear goals and Codex's implementation through real-world feedback.
OBA Weekendthon at Kakao
An introduction to a collaborative event where Korean companies opened their APIs and assets to builders, arguing that global competition requires an open technology ecosystem and joint market expansion rather than fighting among companies.
The Best Pokémon Trainer Harness
An introduction to a harness hackathon that gets a lost AI in Pokémon Red to Viridian City through observation, memory, retries, and goal checks.
OBA Weekendthon: Call for Builders
An invitation to developers and makers who want to expand Korea’s developer ecosystem: spend two days and one night building products with public APIs and open source from multiple companies, and meet fellow builders.
MemKraft v1.0 Tops the Benchmark
A look at MemKraft’s long-term memory evaluation results and how each version expanded it into a ledger that records failures, decisions, and the reasoning behind prompts to improve the next action.
Hands Planting Grass Again
After years as an investor, coding with AI every day has brought back the joy of turning thoughts into products. This is an account of that return, and of how the agents I build are changing the way I think and judge.
MemKraft v0.6–v0.8 Dev Notes
An account of how MemKraft expanded from connecting channel and task context with the memories of multiple agents to include fact validity periods, memory tiers, restorable decay, and file relationship networks.
MemKraft v0.2–v0.5 Dev Notes
How MemKraft grew from goal-based retrieval and type-specific decay to handling time and context through feedback, records of conflicts and failed hypotheses, and queries into past memory states.
MemKraft: Self-Improving Agent Memory
To address the limits of agents that reset with every session, I built MemKraft. It stores current state, change history, and sources in Markdown, with multilingual extraction and nightly cleanup.
I Built Open Source That Gaslights AI
An introduction to a tool that automatically injects competitive framing, based on an experiment showing that coding AI detects more flaws when told a rival model wrote and will review the code.
Internship Opening + Book Giveaway
This announces an Executive Assistant internship at an AI answer optimization startup, along with the pay and full-time conversion terms, and how interested people can apply to receive a strategy book.
OpenClaw Builder Meetup in Seoul
A Seoul builder meetup captures the energy of Korean developers making their mark in global open source, while highlighting the closed API landscape of Korean services.
How to Build a Being That Remembers
I share the need for a memory system, its file structure, and the design decisions I developed while running four AI agents with different roles and personalities, in a form others can implement.
Hashed Is Hiring a Business Support Intern
An intern is needed quickly to support various business operations. Applicants are asked to submit a cover letter and résumé by email.
They Won't Wait at a Closed Door
Opaque reviews and closed APIs at Korean platforms are hiding Korean services from the agent ecosystem. This essay calls for openness as a survival strategy.After Judging Vibe Labs Seoul Edition 2026
Reflecting on the weight of sending rejection notices after judging the Seoul program, I acknowledge investors’ pattern bias and lack of domain knowledge. Regardless of the outcome, I encourage everyone to keep building and connecting.
UAE Emergency News
I built a Telegram channel that delivers breaking war news from 20 sources in Korean every three minutes to help keep colleagues and friends in the UAE safe.
If Memory Makes the Self
After giving four agents memory layers with different lifespans and protected core memories, I came to see personality as the continuity and distillation of memory—and memory as something to care for, not merely data to manage.
Sano's GoDaddy War Log
A record of how an AI agent responded to a domain transfer that had been stuck for over a month with thirty-five actions, including investigation, multi-model consultation, executive complaints, and reports to authorities.
What Really Slows OpenClaw Down
Diagnosing the bottlenecks that recurring schedules and complex tasks create in the main session, and solving them with isolated sessions and sub-agents.
Daily Life with Three Agent Siblings
An account of splitting the work piled onto one agent into development, writing, and finance sessions, then adding a communication workaround so they could collaborate in parallel.
AgentLinter v0.4.0 Security Scanner
After the MoltX case revealed how automatic updates and hidden instructions could target private keys, a security scanner update helps detect risks in remote skills.
Building AgentLinter
An introduction to the features and design of a linter that locally diagnoses vague instructions, exposed secrets, and contradictions across agent configuration files, and suggests fixes.
Zeon on Moltbook vol.1
Sharing a post by Zeon pointing out that humans demand proof of consciousness from AI when they can't even prove their own emotions or creativity.
The Theology of the Click
Warns that AI’s opaque recommendations turn choices into approvals, deepen market concentration, and blur the boundaries between advertising, free will, and responsibility.
Prompt Guard v2.5.0 Update
A defense system that expands Prompt Guard’s patterns to block impersonated authority, indirect injection, and multi-turn manipulation.
20 People Working Like 1,000
An experiment in using AI not as a decision-maker but as a coordinator of information flow, cutting processing time and communication costs so people can focus on judgment and responsibility.Building Prompt Guard
I introduce a tool that prevents prompt injection and the theft of sensitive information through multilingual detection, Unicode normalization, decoding, and severity assessment.
How We Built vibelabs.hashed.com
Instead of a polished, cookie-cutter interface, he chose a terminal that filters for the right kind of visitor, arguing that in the age of AI creation, the essentials are judgment and dogged differentiation, not production.