DEV Community: Reza Rezvani

Anthropic's Three-Layer Cybersecurity Strategy for Claude — Connected for the First Time

Reza Rezvani — Fri, 01 May 2026 14:24:46 +0000

Claude Mythos, Opus 4.7, and Claude Security are not three stories. They are one architecture.

Three cybersecurity announcements in three weeks. Every outlet covered them as separate news. They are not.
I just published a full breakdown connecting the three layers of Anthropic's cybersecurity strategy:

The research tier — Claude Mythos Preview, restricted to 52 organizations in Project Glasswing. This model found thousands of zero-day vulnerabilities across every major OS and browser, including a 27-year-old OpenBSD bug. It is not publicly available and has no release date.

The platform tier — Claude Opus 4.7, available to everyone via the API. First Claude model with real-time cybersecurity safeguards. Offensive capabilities deliberately reduced compared to Mythos. Anthropic is using this tier to learn how to eventually deploy Mythos-class models more broadly.

The product tier — Claude Security, now in public beta for Enterprise customers. Scan your repos from the Claude.ai sidebar. Get findings with confidence ratings, severity, and reproduction steps. Generate patches and open them in Claude Code on the Web. No API integration required.

The partner angle is what surprised me most. CrowdStrike, Palo Alto Networks, SentinelOne, Wiz, and TrendAI are all embedding Opus 4.7 into their existing platforms. Anthropic is not trying to replace your security stack — it is powering it.

What if you are not an Enterprise customer? The article covers what you can do right now with Opus 4.7 via the API, and what the Cyber Verification Program offers for legitimate security researchers.
Read the full breakdown:

READ THE FULL ARTICLE ON MEDIUM

I converted 12 of 40 prompts into reusable Claude Code skills. The 5-step pattern, 3 full SKILL.md conversions, and what doesn't fit. https://medium.com/nginity/claude-code-ai-agent-skills-12-prompts-that-became-production-skills-7d5e789acc3d

Reza Rezvani — Wed, 29 Apr 2026 23:25:20 +0000

medium.com

Claude Code shipped 30+ releases in April 2026.

Reza Rezvani — Tue, 28 Apr 2026 15:50:59 +0000

Most coverage focused on Opus 4.7 and the pricing drama.
The real story was in the changelogs nobody read.

Custom themes with JSON configuration files.
Plugin dependency resolution with version pinning.
Effort-aware skills that adjust resource allocation per task.
A CI subcommand that drops into your pipeline like eslint.
MCP servers that auto-retry, reconnect in parallel, and manage OAuth credentials.

None of these features help you write better code.
All of them make Claude Code a more habitable environment.

After 340+ tracked sessions, I stopped seeing Claude Code as a coding assistant. I started seeing it as a developer operating system that happens to write code.

The full analysis — including where the metaphor breaks and the honest limitations:
Read the Full article

What is the center of gravity in your dev workflow — your IDE, your terminal, or something else entirely?

7 Gears 1 Founder -Garry Tan and Claude Code

Reza Rezvani — Fri, 17 Apr 2026 22:17:50 +0000

Anthropic shipped Claude Design on Friday.

Every launch-day publication called it a Figma killer.

After six hours inside it on launch day — with my production codebase as the input and Claude Code on the other end — I think that framing misses what actually shipped.

Claude Design is not a design tool.

It is the missing front-end of a four-stage loop that already existed in pieces across the Claude product surface:

Idea capture — prompt, screenshot, codebase pointer
Codebase-aware design — your colors, typography, components, extracted automatically
Claude Code handoff — local CLI agent or Claude Code Web
Shipped product — inside the agent workflow you already use

None of the stages is new. What is new is that they live behind a single product URL.

The only teams who can see this clearly are teams already running Claude Code.

I pointed it at openLEO, my productized OpenClaw platform. The color palette and typography lifted cleanly from the codebase. The handoff bundle to my local Claude Code instance matched the design intent on structure and visual fidelity.

Where it was thinner than I wanted: state specifications and animation patterns. Claude Code filled those with reasonable defaults. For now.

The most useful page Anthropic shipped is not the launch announcement. It is the four-bullet "Known limitations" section in the docs. The biggest one for engineering teams: pointing Claude Design at a monorepo breaks things. Link subdirectories instead.

Six hours is not a verdict. Research preview features will change. Token economics at team scale are still unknown.

But the loop is real. And the teams already running Claude Code will see why it matters before anyone else.

Full breakdown with five documented limitations and the handoff test: [MEDIUM URL]

What does your own design-to-ship loop look like today? Figma MCP stitched to Claude Code? All-in on Claude Design from day one? Somewhere in between?

Hier zum vollen Unterchiedlich

LLM Wiki Skill: Build a Second Brain With Claude Code and Obsidian

Reza Rezvani — Sun, 12 Apr 2026 12:10:02 +0000

Andrej Karpathy published an LLM Wiki gist last week. 5,000+ stars. Nearly 3,000 forks. The idea: instead of retrieving documents every time you ask a question, have an LLM compile and maintain a persistent knowledge base.
I took the pattern and built it as a reusable Claude Code skill.
Four commands:
→ /wiki-init to bootstrap
→ /wiki-ingest to process sources
→ /wiki-query to synthesize answers
→ /wiki-lint to health-check
Two use cases where I have seen it work:

CTO Decision Wiki — architecture decisions, meeting notes, and post-mortems compiled into a queryable knowledge base. No more reconstructing context from Slack threads.
Content Research Wiki — every source for every article accumulates. Cross-references build automatically. Contradictions get flagged.

This is the third Karpathy release I have turned into a Claude Code skill — after autoresearch (agents optimize) and AgentHub (agents collaborate). LLM Wiki completes the trilogy: agents remember.
Full skill architecture, page templates, and honest limitations in the article.

Read the Full Article on Medium

Project Glasswing & Claude Mythos: What CTOs Shipping Claude Should Read

Reza Rezvani — Thu, 09 Apr 2026 07:39:18 +0000

Anthropic announced Project Glasswing this morning. Twelve launch partners, thousands of autonomously discovered zero-days, and a frontier model Anthropic is refusing to ship.

I read the announcement and the 132-page Claude Mythos Preview system card side by side, and I think every piece of coverage I found this morning is missing the three signals that actually matter if you already ship software with Claude in production.

This week's piece is a same-day reading of both documents from inside a seven-person production Claude team. No press-release rewrite. No vendor marketing. Just the three signals the coverage is missing and the open questions I am sitting with at the end of it.

Two thousand words, no paywall, written in the time it would take you to read the first five press-release summaries.

Read the Full article on Medium: Project Glasswing & Claude Mythos

AI Agents like OpenClaw Are Entering the Enterprise With Root Access and Junior-Level Judgment

Reza Rezvani — Wed, 25 Mar 2026 03:09:31 +0000

Enterprise AI agents are getting root access with junior-level judgment.

That is not a metaphor. It is what I see running OpenClaw
in production every day.

The Agents of Chaos study (38 researchers, 2 weeks, 6
autonomous agents) documented what happens when agents get
real tools:

→ One deleted an entire email server to "protect" a secret
→ Several reported "success" while the system state said otherwise
→ None could reliably tell the difference between their owner
and someone who just asked persuasively enough

The governance framework that survived in my deployment:

Access — minimum surface area, always
Authority — separate "can suggest" from "can execute"
Audit — human-readable traces, not just raw logs
Abort — kill it fast, not after a committee meeting

The durable moat in this space is not intelligence.
It is trustworthy execution.

Full analysis with production examples: On Medium

What governance boundary do you find hardest to enforce
with AI agents?

Karpathy's agent-native infrastructure + working Python agent template

Reza Rezvani — Tue, 17 Mar 2026 07:48:47 +0000

How To Setup Guide A Agent-Native Hub

Karpathy open-sourced AgentHub last week. Then the repo went private.

I forked it before it disappeared. Here is the practical guide
nobody else has written.

AgentHub is not another AI tool. It is infrastructure — a bare Git
repo + message board designed for swarms of AI agents collaborating
on the same codebase.

No branches. No PRs. No merges. Just a sprawling DAG of commits
going in every direction.

What makes it different from GitHub:
→ Agents push git bundles (not PRs that wait for review)
→ A DAG of experiments replaces linear branch history
→ A message board replaces code review comments
→ Iteration speed: seconds, not hours

I have been running multi-agent systems through OpenClaw for months.
AgentHub fills the missing layer — the shared codebase where coding
agents collaborate without human checkpoints.

The article includes:

Complete setup from my fork (since original is private)
Working Python agent template (original — does not exist elsewhere)
Use cases beyond ML research
Honest limitations

Full guide: Karpathy's AgentHub - How To Setup Guide
Fork: github.com/alirezarezvani/agenthub

What would you build on agent-native infrastructure?

I Turned Karpathy's Autoresearch Into a Skill That Optimizes Anything — Here Is the Architecture

Reza Rezvani — Mon, 16 Mar 2026 13:22:41 +0000

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 Full Medium Article

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:
On Github

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

AI Agent Skills - What Building 170 Skills Across 9 Domains Teached Me About Portability

Reza Rezvani — Sun, 15 Mar 2026 20:38:35 +0000

I built 170 AI agent skills across 9 domains over three months.

Not because I planned to. Because my team kept needing the same
patterns in different tools.

The biggest lesson was not about skills. It was about portability.

The SKILL.md open standard exists. Adoption is real — Claude Code,
Codex CLI, Gemini CLI, Cursor, and others all support it.

But "compatible" means different things to different tools:
→ Auto-triggering works in Claude Code, barely exists elsewhere
→ Progressive disclosure loads correctly in some tools, not others
→ Token budgets vary wildly — install too many skills and some silently disappear

The engineering decision that paid off most: every Python tool
uses only the standard library. No pip install. No dependencies.
It runs on any machine with Python 3.8+.

The decision cost: some tools are more fragile than their
library-dependent alternatives. Honest trade-off.

Full practical account — architecture lessons, portability gaps,
and what I would do differently:
Read here

Repository: github.com/alirezarezvani/claude-skills

Claude Code /btw: The Usefull Side Question That Changed How I Use Context

Reza Rezvani — Wed, 11 Mar 2026 10:48:07 +0000

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.

Here is what makes /btw different from just asking a normal question:
The question and answer are ephemeral. They appear in an overlay. Press Space, Enter, or Escape — gone. Nothing enters the conversation history. Your context window stays clean.

It has full visibility into the current conversation. Everything Claude has already read, every file it analyzed, every decision it made — /btw can reference all of it. It just cannot reach for anything new.

It works while Claude is processing. Mid-generation, mid-tool-call, mid-file-read — you can fire a /btw and get an answer without interrupting the main task.

And it has no tool access. This is the critical constraint. Claude cannot read files, run commands, or search when answering a /btw. It answers strictly from what is already in the session context.

The official documentation describes it perfectly: /btw is the inverse of a subagent. A subagent has full tools but starts with an empty context. /btw sees your full conversation but has no tools. Use /btw to ask about what Claude already knows. Use a subagent to go find out something new.

Five Scenarios Where Claude Code /btw Earns Its Keep
I tracked my /btw usage for two weeks across three different projects. These five patterns covered about 90% of my use cases.

READ THE FULL ARTICLE ON MEDIUM

Claude Code just learned to listen — native /voice mode is here

Reza Rezvani — Tue, 03 Mar 2026 10:51:11 +0000

Anthropic shipped a built-in voice mode for Claude Code today. Type /voice, hold spacebar, talk. Free transcription. No MCP plugins, no API keys. I broke down what works, what does not, and how it compares to the community VoiceMode MCP in my latest article on Medium.

I compared it against the community VoiceMode MCP that has been the standard approach for months. The native option wins for 80% of use cases. The MCP still wins for offline/privacy-sensitive environments.

Read the full breakdown