Untitled @noor-mohammad - Tumblr Blog

Your 40-rule CLAUDE.md is making your Claude Code agent worse.

The pattern every dev falls into: agent does something annoying → add a rule. Repeat. 40 rules later, the agent is ignoring half and hallucinating conventions that aren't in the file.

The instinct is "more rules = more control." The data says the opposite.

The most-starred CLAUDE.md on GitHub — tens of thousands of stars, more than most frameworks in the space — contains essentially 4 sentences.

→ No dependencies → No build step → No plugins → Just 4 lines describing how an agent should BEHAVE

The principle: stop telling the agent what NOT to do. Tell it who it IS.

10-min read with before/after code from a real project:

The most-bookmarked resource in AI coding isn’t a framework or a plugin — it’s four sentences in a markdown file. Here’s why they…

Claude Fable 5 broke the tier ceiling.

On June 9, 2026, Anthropic pulled its Mythos-class tier — the rung above Opus — down to general API access. The headline most coverage stopped at.

The two strings they attached are the real story:

→ A rerouting classifier that can silently hand your request to a weaker model → A 30-day data retention rule that some enterprises literally can't accept

Fable 5 is the most capable Claude you can call today. It's also the most conditional. Both are true — and the gap between them is where your decision actually lives.

11-min read:

Anthropic just shipped its most capable public model ever, pulled a once-restricted capability tier down to the API, and quietly attached…

Anthropic Didn't Win the Model Race. It Won the Workflow

Opus 4.8 shipped 41 days after 4.7 — and the cadence, not the benchmark, is the real story.

The moat isn't the model (models leapfrog every six weeks). It's the workflow layer: the 20 hours a week devs spend inside Claude Code, and the trust that lets them walk away from the keyboard.

Leaderboards are for press releases. Trust is for Tuesdays.

Full piece → https://medium.com/@reactjsbd/anthropic-didnt-win-the-model-race-it-won-the-workflow-af455ff8b8c0

You wrote off Grok too early

You wrote off Grok too early.

xAI quietly shipped ~15 consequential things in 2026. The three that matter:

Grok 4.3 — 2-million-token context window (largest among Western closed models)

Grok Build — agentic coding CLI, 8 parallel sub-agents, compatible with Anthropic Skills + MCP

Grok Imagine 1.0 — 720p video with native synchronized audio

Default Grok hides almost all of this. The 12 setups in the article are the gap — plus an honest "vice-versa alternative" for each, because no, Grok isn't always the right answer.

17-min read:

https://medium.com/@reactjsbd/the-best-12-grok-setups-a-complete-2026-guide-to-getting-real-work-out-of-xai-61e079dba443

Airbnb owns zero properties. Banked $4.6 billion in cash last year.

That's the Airbnb business model in one line. $91.3B in bookings flowed across the platform. $12.2B kept as revenue. $4.6B fell to free cash — a 38% FCF margin.

Marriott and Hilton own or finance buildings. Airbnb owns a database and a brand. Same industry. Wildly different economics.

Wrote up the full teardown — the take rate, the matching layer, and why the machine didn't work for the first decade: