Posts
All the articles I've posted.
-
The CLI's Second Spring in the AI Era: Three Structural Reasons Behind the Claude Code / Codex / Charm / Ink Surge
Starting in 2025, Claude Code, Codex CLI, Charm Bubble Tea, and Ink all took off at once. After 30 years of GUI dominance, the CLI suddenly became the default form factor for AI tooling — Claude Code itself even ships as a Bun binary. This isn't nostalgia. Three structural constraints — LLM I/O, controllable long-running tasks, and cross-tool interop — are pushing engineers back to the terminal. Includes a 30-line Claude Agent SDK sample.
-
After the Loop Is Running: A Playbook for Verification, Comprehension, and Cognitive-Surrender Debt
blog191 walked through the five components of Loop Engineering and named three debts—verification, comprehension, cognitive surrender—without offering a treatment plan. This post fills that gap: a Wharton 1,372-person experiment, Anthropic's 52-engineer study with a 17% comprehension gap, a real $4,200 overnight bill, Claude Code's PreToolUse rejection hook, and three checkpoint patterns—the full playbook in one place.
-
Stop Re-Introducing Your Project to AI Every Session: A Project Passport with AGENTS.md, CLAUDE.md, and memory
AGENTS.md is the de facto standard in 2026, CLAUDE.md is still Claude Code's richer format, and project memory holds a different class of information altogether. Stitch them into a "Project Passport" so every new session begins with the AI clearing customs in seconds instead of asking you to paste context again.
-
A 2026 Monorepo Setup From Zero to Production: pnpm catalogs, Turborepo 2.x, changesets
The previous post argued monorepos are an organizational problem. This one walks through the actual setup, end to end. pnpm catalogs for version alignment, Turborepo 2.x task graphs, changesets + OIDC publishing, remote cache, CODEOWNERS — every command and config you need, current as of 2026.
-
Monorepo Is an Org Decision, Not a Tech One: Lessons from Babel, Lerna, and Mercari
Monorepo isn't a yes/no technical choice. Babel went in and then quietly walked half of it back, Lerna spent a stretch with no maintainer, and Mercari's cross-region CI burned real money for a year. All three cases point at the same thing: what decides whether a monorepo works isn't Turborepo or Nx, it's your team and how it's organized.
-
Loop Engineering: From Writing Prompts to Designing Loops That Run Agents for You
The "Loop Engineering" term Addy Osmani popularized in June isn't a replacement for prompt engineering — it's about swapping you out as "the person who hits enter" and turning you into "the person who designs the loop". Walks through the five components plus a state, and the three debts Osmani is really worth remembering for (verification debt / comprehension debt / cognitive surrender) — and along the way, why I disagree with him placing loop above the harness.
-
Two Days with Claude Fable 5: The 5 Things Every API Integrator Actually Has to Change
Anthropic shipped Fable 5 on 6/9, swapping the Opus/Sonnet/Haiku naming for Fable/Mythos. But the things that actually force every Claude API integrator to touch code aren't the names — they're the new stop_reason refusal, the forced adaptive thinking, and the "you must wire up a fallback model" architecture. After two days driving it inside Claude Code, here's the integration-side detail you need.
-
Prompt, Context, Harness, Agentic: The Four Nested Layers of LLM Apps — and Knowing Which One You're Stuck In
Prompt engineer, context engineer, harness engineer, agentic engineer — these aren't four competing job titles. They're nested layers of concern, from a single instruction to an entire autonomous system. Understand how the four layers relate, and you'll know exactly which one you're optimizing every time you get stuck.