jungard

claude ultrareview — a CLI subcommand for CI pipelines.

Prints findings to stdout. --json for raw output. Exit 0 or 1.

Drop it into GitHub Actions like eslint. My engineer added it Thursday. The PR was four lines.

Claude Code now assumes you might NOT be at a keyboard.

The CLAUDE.md at the root of that repo says, in plain English, on line eleven: never modify migration files without explicit approval. The session loads the file. Claude reads the rule. Claude modifies a migration file fourteen minutes later.

What the field is teaching, and where it stops I went back through the most-trafficked CLAUDE.md articles on Medium from the last six months. Not to dunk on anyone. To look for the gap. The advice clusters into two dominant patterns and one good outlier, and almost every new piece is a remix.

Pattern one is the personal business OS. A solo operator publishes a beautiful directory structure. There is a decisions/ folder. There is a priorities/ folder. There is a meetings/notes/ folder with date-prefixed transcripts. The author runs it for themselves and the system clearly works for them. I have read at least eight versions of this article. Every published example is one person operating on their own machine.

That was my first real session with OpenAI’s codex-plugin-cc — an official plugin, published March 30, 2026, that puts Codex directly inside Claude Code. Not a community hack. Not a bash script wrapper. OpenAI built this, packaged it as a plugin, and shipped it through Claude Code’s plugin marketplace.

Every tech outlet has already covered what it is. This article covers what to actually do with it.

What This Is (And What It Is Not) The plugin gives you six slash commands inside Claude Code:

/codex:review — standard read-only code review of your current changes /codex:adversarial-review — steerable review that challenges design decisions /codex:rescue — delegates a task entirely to Codex (bug investigation, fixes, parallel work) /codex:status, /codex:result, /codex:cancel — job management for background tasks It also includes an optional review gate — a Stop hook that intercepts Claude’s output and runs a Codex review before Claude can finish. More on that in Use Case 3.

Karpathy released autoresearch last week. 31,000 stars. 100 ML experiments overnight on one GPU.

Everyone wrote about the ML training loop. I saw something different: a pattern.

One file. One metric. One loop. Modify → Evaluate → Keep or Discard → Repeat.

That pattern has nothing to do with machine learning.

So I built a skill that applies it to: → API response time (benchmark_speed evaluator) → Bundle size (benchmark_size evaluator) → Headline click-through (LLM judge evaluator) → System prompt quality (LLM judge evaluator) → Test pass rate, build speed, memory usage

Works across 11 tools: Claude Code, Codex, Gemini CLI, Cursor, Windsurf, OpenClaw, and more.

The hardest problem: evaluating things that are not numbers. Headlines do not come with a val_bpb metric.

Solution: LLM judges using the agent's own subscription. Critical constraint: the agent cannot modify its own evaluator. (The alignment problem in miniature.)

What I have not done yet: run 100 experiments overnight. The skill shipped this week. The architecture is solid. The validation is ahead of me.

Full architecture + honest limitations: https://github.com/alirezarezvani/claude-skills/tree/main/en...

What manual optimization loop are you running that should be automated?

I maintain a skill library that started as my own Claude Code setup and grew into something bigger (177 skills, MIT-licensed).

Here's what I actually use for writing, research, and productivity — not the full list, just the ones that make a real difference.

Would love to collect your feedback on this project. Happy to get connected with me.

Reza

5 practical scenarios where questions save time, tokens, and sanity — My takes as a practitioner who almost missed this feature entirely.

What Claude Code /btw Actually Does The /btw command in Claude Code lets you ask a side question without adding anything to the conversation history. You type /btw followed by your question, get an answer in a dismissible overlay, and the main conversation continues as if nothing happened.

That sounds simple. It is. But simple is not obvious, and the implications only clicked after I started using it daily.

180+ production-ready skills and plugins for Claude Code, OpenAI Codex, Gemini CLI, and OpenClaw — reusable expertise bundles that transform AI coding agents into specialized professionals across engineering, product, marketing, compliance, and more.

Skills are modular instruction packages (plugins) that give AI coding agents domain expertise they don't have out of the box. Each skill includes a SKILL.md (instructions + workflows), Python CLI tools, and reference documentation — everything the agent needs to perform like a specialist.

One repo, four platforms: Works natively as Claude Code plugins, OpenAI Codex CLI and agents, Gemini CLI skills, and OpenClaw skills. All 237 Python tools are stdlib-only and verified working.

Most engineers hear “scheduled tasks” and think cron jobs. A time trigger, a script, a log file nobody checks. The Claude Code loop command is not that. It is closer to handing your night shift to a colleague who already knows your codebase, your conventions, and your PR review standards — then finding three pull requests waiting for you at breakfast.

Boris Cherny, the engineering lead behind Claude Code, runs five parallel Claude instances daily. Twenty to thirty PRs per day. His CLAUDE.md is a living document that evolves with every session. When Anthropic shipped /loop, they did not build a scheduler.

Peter Steinberger — the creator of OpenClaw, formerly Moltbot — put it bluntly at the first ClawCon event in San Francisco this February, where 700 developers showed up (compared to 20 at the inaugural meetup just weeks earlier): it’s about building systems that do things on your computer. Not chat. Not suggest. Do.

Claude Code handles your codebase — reading files, writing code, running tests, managing git. It lives in your terminal. OpenClaw handles everything else — email triage, calendar management, messaging, browser automation, smart home control. It lives in your messaging apps, running as a persistent daemon with long-term memory.

Why Anthropic Made Their Smartest Model 67% Cheaper (It’s Not Desperation) Anthropic released Claude Opus 4.5 yesterday. It scored higher than any human candidate ever on their notoriously difficult engineering take-home test. But that’s not the only headline.

The headline: they slashed pricing from $15/$75 to $5/$25 per million tokens. A 67% price drop on their flagship model. Opus now costs less than Sonnet did six months ago.

I’m running Opus 4.5 right now in a multi-tool session with web searches, file operations, and extended context.

What follows: news analysis, real-time observations, and strategic breakdown of what Anthropic is actually doing here.