Writing

This page is the map. The homepage is the handshake. Individual essays are the proof.

Start Here

• A Guide to Meta-Prompting for Agents in 2026 - the reusable move for forcing agents to recover assumptions, telos, constraints, and success before answering
• OMP and Codex - Tale of Two Harnesses - how file-editing grammars, compaction policy, and prompt caching change the same model
• A thesis on Work and Research with agents, in 2026 and beyond. - why the future of knowledge work is an alloyed composite of un-correlated intelligences
• Sol maxxing without breaking the bank - a working argument for cost-aware reliability targets in agentic systems
• The Context Window Is Not Your Memory Architecture - the core distinction between working set and truth layer
• Giving Agents a CLI Is Not About Commands - a Remotion case study on repo-local CLIs as agent control planes

Recipes

Practical notes that teach a repeatable move, not just a take.

• A Guide to Meta-Prompting for Agents in 2026 - recover intent before generation
• A Recipe for Curating Expert Datasets - golden labels, taste, judgement, and expert routing
• Why Evals Are Not the Same Thing as Golden Labels - separating measurement from training signal
• When Fine-Tuning Is Worse Than In-Context Learning - sample efficiency before weight updates
• How to Route Contested Examples to Experts - verification loops when expert patience is expensive

AI Memory and Agent Systems

Practical frameworks for building agents that do not confuse a longer prompt with durable memory.

• The Context Window Is Not Your Memory Architecture - the core distinction between working set and truth layer
• MEMORY.md Should Stay Small. Open Brain Should Grow. - why startup context should stay small
• Managed Memory Has Three Simultaneous Constraints - fidelity, cost, and continuity in one design space

Deep Research and Comparison

Notes on how research agents search, normalize, compare, and synthesize evidence.

• OMP and Codex - Tale of Two Harnesses - OMP, Codex, Hashline editing, Responses state, prompt caching, and compaction
• What OpenAI and Gemini Deep Research Are Actually Built For - task shape before benchmark theater
• High-Arity Comparison Breaks Research Agents - why five-way comparison is a topology problem
• Comparator-Aware Working Memory - schema first, lanes second, ledger third, adjudication last

Agent Economics

Writing about the operational and economic consequences of agentic systems.

• Sol maxxing without breaking the bank - a working argument for cost-aware reliability targets in agentic systems
• Giving Agents a CLI Is Not About Commands - a Remotion case study on repo-local CLIs as agent control planes
• KV Cache Sharing Changes the Economics of Multi-Agent Systems
• A Runtime Can Own State Without Owning Cognition
• Here Is What My Agents Produced While I Was Sleeping

Personal Operating System

The life side of the work: attention, taste, agency, self-respect, and ambition without self-erasure.

• What I Am Optimizing For
• Notes on Becoming Harder To Fool
• Designing Your Life, But With Better State

The Publishing Rule

Every public essay should earn one of these jobs:

• explain a hard idea with less fog
• show the shape of a system
• expose a hidden tradeoff
• make a personal claim that costs something

If a post cannot do one of those, it stays in notes.