# Hasan Arief · LLM research notes

> Personal lab notebook of Hasan Asyari Arief, Senior Researcher (machine
> learning) at NORCE Research AS. Agentic coding practice, open-weight LLMs,
> curated benchmarks, deployment costs. Every factual claim links its source.
> No sponsors, no affiliates. Entries are dated; guides update in place.

## Guides (long-form, evergreen)

- [Karpathy's LLM wiki: a knowledge base your coding agent maintains](https://notes.hasanarief.dev/guides/llm-wiki-knowledge-base/): Karpathy's LLM wiki: point a coding agent at a markdown folder and let it maintain the knowledge base. What it is, why it earns its place, and how to set it up. (raw markdown: https://notes.hasanarief.dev/guides/llm-wiki-knowledge-base.md)
- [Claude Code in 2026: the agentic loop, capabilities, and architecture](https://notes.hasanarief.dev/guides/claude-code-reference/): A maintained third-party reference to Claude Code: the agentic loop, built-in tools, context window behavior, environments, safety model, and extension layer. (raw markdown: https://notes.hasanarief.dev/guides/claude-code-reference.md)
- [Open-weight LLMs for coding in 2026: hardware and real costs](https://notes.hasanarief.dev/guides/open-weight-llms-for-coding/): Open-weight coding LLMs as of July 2026, GLM-5.2, Kimi K2.7, DeepSeek-V4, with the VRAM math, GPU rental prices, and when self-hosting beats an API. (raw markdown: https://notes.hasanarief.dev/guides/open-weight-llms-for-coding.md)
- [OpenAI Codex in 2026: capabilities, surfaces, and how it works](https://notes.hasanarief.dev/guides/openai-codex-reference/): A maintained third-party reference to OpenAI Codex: the CLI, IDE, cloud and app surfaces, AGENTS.md configuration, skills, MCP support, plans and access. (raw markdown: https://notes.hasanarief.dev/guides/openai-codex-reference.md)
- [Setting up Claude Code in 2026: install, CLAUDE.md, skills, hooks, memory](https://notes.hasanarief.dev/guides/claude-code-setup/): Practical Claude Code setup for 2026: install on Mac, Linux or Windows, authenticate, run /init, add skills and hooks, and stay safe with git. (raw markdown: https://notes.hasanarief.dev/guides/claude-code-setup.md)
- [Agent orchestration patterns in 2026: subagents, teams, and pipelines](https://notes.hasanarief.dev/guides/agent-orchestration-patterns/): A map of agent orchestration in 2026: five coordination levels from a single session to agent teams, with failure modes and selection criteria. (raw markdown: https://notes.hasanarief.dev/guides/agent-orchestration-patterns.md)
- [Codex CLI vs Claude Code: what I learned running both daily](https://notes.hasanarief.dev/guides/codex-cli-vs-claude-code/): An honest Codex CLI vs Claude Code comparison from running both daily on one codebase: strengths, config interop, and the two-model workflow that emerged. (raw markdown: https://notes.hasanarief.dev/guides/codex-cli-vs-claude-code.md)
- [AGENTS.md and CLAUDE.md best practices: one instruction file for every coding agent](https://notes.hasanarief.dev/guides/agents-md-best-practices/): How to write AGENTS.md and CLAUDE.md files: what to include, the 200-line size target, monorepo precedence, imports, and one-source-of-truth setup. (raw markdown: https://notes.hasanarief.dev/guides/agents-md-best-practices.md)
- [Agentic coding best practices: what works in 2026](https://notes.hasanarief.dev/guides/agentic-coding-best-practices/): Field-tested agentic coding best practices from daily Claude Code and Codex work: verification loops, context discipline, hooks, and review gates. (raw markdown: https://notes.hasanarief.dev/guides/agentic-coding-best-practices.md)

## Notes (short, dated)

- [1M-context Opus on the $200 Max plan: what is included and what is capped](https://notes.hasanarief.dev/notes/max-plan-1m-context-value/): The $200 Max plan is not unlimited, Anthropic says so plainly. Yet 1M-context Opus is included, and at six concurrent agents I have yet to hit a cap. (raw markdown: https://notes.hasanarief.dev/notes/max-plan-1m-context-value.md)
- [Steering agents from my phone: a terminal stack vs Claude Code Remote Control](https://notes.hasanarief.dev/notes/mobile-agent-steering-vs-remote-control/): Tailscale, tmux, mosh, and a mobile terminal let me steer long agent runs from my phone. Why that plain stack still outperforms Remote Control for me, for now. (raw markdown: https://notes.hasanarief.dev/notes/mobile-agent-steering-vs-remote-control.md)
- [tmux as a transparent orchestrator for long agent sessions](https://notes.hasanarief.dev/notes/tmux-agent-orchestration/): Why a plain terminal multiplexer beats bespoke agent frameworks for long generation runs, and five tmux practices each learned from a real failure. (raw markdown: https://notes.hasanarief.dev/notes/tmux-agent-orchestration.md)
- [When skipping Claude Code permission prompts is safe, and when it is not](https://notes.hasanarief.dev/notes/skipping-permission-prompts-safely/): Claude Code users approve 93% of permission prompts, so manual mode and bypass converge. What prevents catastrophe is the environment I run in, not the flag. (raw markdown: https://notes.hasanarief.dev/notes/skipping-permission-prompts-safely.md)
- [The AI capex bubble, and what I am building while it lasts](https://notes.hasanarief.dev/notes/ai-bubble-good-thing/): Trillions committed, Nvidia the most valuable company ever, and the measured productivity gains are tiny. Notes on using the overbuild and surviving the crash. (raw markdown: https://notes.hasanarief.dev/notes/ai-bubble-good-thing.md)
- [Karpathy's autoresearch, and its loop for parameter optimization](https://notes.hasanarief.dev/notes/autoresearch-parameter-optimization/): Karpathy's autoresearch runs an AI agent in a keep-or-discard loop. I examined its techniques, then applied the loop to parameter search where a grid stalls. (raw markdown: https://notes.hasanarief.dev/notes/autoresearch-parameter-optimization.md)
- [Codex cloud vs Codex CLI: which surface for which work](https://notes.hasanarief.dev/notes/codex-cloud-vs-cli/): Where OpenAI Codex cloud tasks actually run, what the CLI keeps that cloud cannot, and a decision split for routing work between them. (raw markdown: https://notes.hasanarief.dev/notes/codex-cloud-vs-cli.md)
- [How Codex CLI discovers skills: the .agents filesystem layout](https://notes.hasanarief.dev/notes/codex-skills-discovery/): Where OpenAI Codex looks for skills on disk, in what order, how SKILL.md files load, and how the same format ports between coding agents. (raw markdown: https://notes.hasanarief.dev/notes/codex-skills-discovery.md)
- [Fable 5 versus Opus 4.8 with /goal: pricing and persistence compared](https://notes.hasanarief.dev/notes/fable-5-vs-opus-48-goal-loop/): Fable 5 returned at double the price of Opus 4.8, with days-long autonomy as the headline capability. A comparison with Claude Code's /goal loop, sourced. (raw markdown: https://notes.hasanarief.dev/notes/fable-5-vs-opus-48-goal-loop.md)
- [The two MCP servers I use daily](https://notes.hasanarief.dev/notes/mcp-servers-i-use-daily/): Two MCP servers remain in my daily agentic coding configuration: Playwright MCP and mobile-mcp. Notes on the selection rule that filters the rest. (raw markdown: https://notes.hasanarief.dev/notes/mcp-servers-i-use-daily.md)
- [Why this notebook exists](https://notes.hasanarief.dev/notes/why-this-notebook-exists/): Opening note: what I research, why these notes are public, and the rules this site follows, dated entries, sourced numbers, no sponsors. (raw markdown: https://notes.hasanarief.dev/notes/why-this-notebook-exists.md)

## About

- [About the author](https://notes.hasanarief.dev/about/): who writes this and the rules it follows
- [Author CV](https://hasanarief.dev/): Hasan Asyari Arief, NORCE Research AS