I rewrote my CLAUDE.md this week


I rewrote my CLAUDE.md this week
Plus the 1M context bill, an 11-day design-tool fork, and the release that survives Mac sleep.  ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌
WotAI

I rewrote my CLAUDE.md this week

Hey Reader,

Two weeks since the last email. A lot happened. The thread that ties it all together: Opus 4.7 dropped on April 17, and almost every Claude Code release since has been about hardening the runtime around the new model. Memory leaks, context calculation, OAuth resilience, Mac-sleep recovery, managed settings, gateway integration. None of it flashy. All of it exactly what a model migration needs.

Four reads from the last 14 days. I'd go in this order.

The one Claude Code trick that cuts context by 64% (Episode 1)

Claude Code Episode 1 cover - 1M context window deep dive

This is the start of the 10-episode "Production Claude Code" series. Opus 4.7 ships with a 1M token context window – and every coverage piece sold it as "load your whole repo." In production, that's the most expensive way to use it. Every token you load gets billed every turn until compaction. Load 100K tokens, do 20 turns, you've paid for 2M input tokens of rent.

The skill isn't fitting more in. It's loading less of what doesn't matter. Episode 1 walks through the four levers that actually move the needle on a real $200 Max workflow: line ranges, `.claudeignore`, `/context` as a dashboard, and manual `/compact` before auto-compact fires at 97%.

Read Episode 1 →

Compound Engineering for Claude Code + AGENTS.md

Compound Engineering + AGENTS.md cover - splitting CLAUDE.md into AGENTS.md for process anchoring

Your CLAUDE.md is doing two jobs and badly at both. It holds code conventions (where do imports go, what's the test framework). It also holds your dev process (Linear conventions, branch naming, when to brainstorm vs ship). Those have different change cadences. Stuffed into one file, neither reads cleanly.

The fix is structural. CLAUDE.md keeps codebase context. AGENTS.md carries the development process. Claude Code reads AGENTS.md first, so process anchors every prompt. This post is the actual structure I'm running on a real production codebase – the 12-step pipeline, the blast-radius tier rules, Linear conventions, and how the Compound Engineering plugin runs the show.

Read the AGENTS.md breakdown →

Claude Design vs Open Design – did "design tool" just stop being a category?

Claude Design vs Open Design cover - 11-day fork from launch to open-source clone

Anthropic launched Claude Design on April 17 – Opus 4.7-powered, hosted, $20/mo on Claude Pro. Eleven days later, Open Design shipped under Apache 2.0 with the same artifact-first loop and 71 brand-grade design systems, but BYOK on whichever coding CLI you already pay for.

Eleven days from launch to OSS clone. That fork tells you something about every vertical AI-native SaaS category right now. The post is an honest comparison plus a decision framework for picking between them based on your existing stack.

Read the comparison →

Claude Code 2.1.126 – sessions survive Mac sleep, OAuth works over SSH

Claude Code 2.1.126 cover - Stream idle timeout struck through in phosphor green with FIXED in warm gold

Five releases shipped in two weeks (2.1.116 → 2.1.126), and 2.1.126 is the biggest one. 33 fixes in a single release, 80+ since the last broadcast.

The two that matter most: three Stream idle timeout fixes that stop sessions from dying after Mac sleep or during long thinking pauses, and an OAuth login flow that finally works over SSH, WSL2, and containers (paste the code into the terminal). Plus `claude project purge`, gateway-aware `/model` picker, and a managed-sandbox enforcement security fix worth knowing about if you run Claude Code in an enterprise.

This recap consolidates everything from 2.1.117 through 2.1.126 – the 1M context window calculation fix, the 4 memory leaks, `claude ultrareview` in CI, the AGENTS.md / process file convention, plus today's 33-entry hardening pass.

Read the 2.1.126 recap →

Reading order if you want the through-line: start with Episode 1 (load less context), then AGENTS.md (split the process file), then 2.1.126 (the runtime that backs all of it), and finish with Claude Design vs Open Design (where the market is heading).

Reply and tell me which one moved a needle for you. I read every reply.

– Alex


P.S. – Episode 2 of the production series drops this week. It covers steering with modes – plan mode, auto mode, and `--dangerously-skip-permissions` (which got more permissive in 2.1.126, btw – the post explains when that's actually OK to turn on). If you want the full systematic version, it's the 5-day workshop.

Production-Grade Claude Code in 5 Days – June 1 cohort. Early bird is $297 through May 15 (saves $200 off the $497 regular price). 30 seats, still open.

Same setup I use to ship daily, unpacked in 5 days.

Save my seat →

Read Episode 1 →

Reply to this email anytime – I read every one.

WotAI · AI Automation that Actually Ships

wotai.co

WotAI · 440 N. Barranca Ave., Unit 3777, Covina, CA 91723

WotAI

If you're building with Claude Code, n8n, or AI agents in production, this is the email I'd want. Weekly recap of what shipped, what broke, and how I'm running it. Specific over vague. Replies welcome.

Read more from WotAI
Your CLAUDE.md File Is Too Long

1,204 lines down to 151 The four-bucket diet, and the 20 lines no model can rediscover. Plus 3 more reads. 1,204 lines down to 151 Hey Reader, Four reads this week, and each one spends its best section on the same question: which part of this should you leave alone? The lines you don't delete. The invoices you keep typing by hand. The decisions you never route through software. Every line gets paid for on every session (most don't earn it) Anthropic cut roughly 80 percent of Claude Code's own...

When not to buy an AI agent The three questions that decide it, and what a missed call is actually worth. Plus 3 more reads. When not to buy an AI agent Hey Reader, Two people in your business describe the same process differently. That is the clearest signal it is not ready to automate. You do not have a workflow, you have two workflows and a disagreement. Automate it anyway and you lock in whichever version the person configuring it happened to believe. This week's first two reads are about...

110 free AI models, tested 16 providers, what's actually free, and the catches each one hides. Plus 3 more reads. 110 free AI models, tested Hey Reader, Four reads this week. The first two go together: where the free AI models actually are, and how to put Claude Code in charge of them. 110 free AI models, and what each one actually costs you Sixteen providers hand out real API access right now. Not trials. Free tiers, running today. The catch is that "free" is doing different work at each...