Anthropic's project-artifact plugin turns your repo into a live status page Save this post if you want to try it on your own projects What I like is that it's not a one-off page It saves where the data comes from, so when you refresh it pulls the current state and tells you what changed Here's how: - Reads your repo, PRs, tracker, or design doc - Publishes the page as a private claude artifact you can share with your team - Covers overview, success criteria, workstreams, and next steps, and only adds risks, plan, or FAQ tabs when there's something to put there - Numbers PRs as 1.0, 1.1, 2.0 so you can see which ones depend on which - Gets PR, CI, and review status with the gh CLI - Keeps a config per project, so "refresh the artifact" updates the same URL from any session If a fetch fails it marks that data as stale instead of guessing, and it treats PR text as data, not instructions Having one built from what's actually in the repo makes a lot more sense Install: /plugin install project-artifact@claude-plugins-official Run: /project-artifact Repo:

Anthropic launches OSS Scanner A free, opt-in service that scans open source repositories for security vulnerabilities using Anthropic's strongest models Designed for eligible open source projects, it provides periodic security scans and private reports with reproducible findings and proposed patches I've already submitted Here's how to submit yours: - Add a Dockerfile to your repo that installs dependencies and builds your project - Optional: Create a threat_model.md describing security risks and critical components - Fork anthropics/oss-scanner - Create projects//project.yaml with your repo URL, contact email, and configuration paths - Validate with python tools/validate.py and test with tools/check - Open a PR, complete the checklist, and sign the CLA - Wait for Anthropic to review and approve your submission Once accepted, you'll receive security reports directly by email Repository:

This is what a Claude Code session with this setup would look like Sonnet does a great job as the main model, delegating smaller tasks to Haiku subagents But sometimes it needs help with more complex decisions That's when Opus steps in as an advisor, reviews the work, reasons through the problem, and guides Sonnet so it can keep going

Sonnet as the main model, Opus as an advisor, and Haiku for subagents Save this and try it with one of your sessions Install: npx claude-code-templates@latest --setting model/sonnet-opus-advisor-haiku-subagents This command installs the settings required to configure your Claude Code sessions with this model distribution Select the appropriate configuration scope: User, Project, or Local

Claude Projects is finally starting to roll out! Perfect time to read this article where I explain how to set it up and what to keep in mind when creating your projects👇

Microsoft just released its own Jev... I'll have to give it a try this weekend I recommend building your own evals for decision-making models. There are already enough options out there to start running large scale evaluations

Jev Skill Typeahead is fully open source Feel free to explore the code, contribute improvements, and help make skill predictions even more accurate GitHub:

Jev Skill Typeahead for Claude Code As you type your prompt, Jev analyzes it and predicts which of your installed skills Claude is most likely to invoke It shows each skill's estimated probability before you even send the request Super useful for understanding and optimizing your skill pipelines before execution And it uses 0 Claude tokens Install: npx claude-code-templates@latest --mod productivity/jev-skill-typeahead

A good security practice is to choose the right permission level for each task Claude Code offers claude --restricted, which: - Disables command execution and WebFetch by default - Restricts file operations to working directories - Ignores user and project settings - Prevents permission bypasses - Blocks protected-path writes Not every agent needs full access to your environment The goal isn't to restrict everything, but to give each agent only the permissions it actually needs

This image doesn’t mean some tasks are easier than others It means each model has a role where it can deliver the most value If you’re using Opus for tasks Haiku can handle with the same quality, you’re likely wasting time and money Choosing the right model for each task is a core AI engineering skill, and exactly the kind of decision Forward Deployed Engineers (FDEs) make Not sure which model to use? Build a dataset of real tasks, run evals across models, and compare: - quality - latency - token usage - cost Don’t assume the most powerful model is always the best choice… Measure it!

Anthropic's new plugin puts a team of agents on hard math /plugin install math-proof@claude-plugins-official Two skills for research level problems, both ending in a self contained proof /math-proof:solo has one Claude work the problem, saving each step to notes /math-proof:siege has a judge assign questions to parallel workers for hours, with every proof checked twice

Very interesting paper on how secure coding agents actually are Bookmark this one, it's worth reading later They reviewed 40 harnesses and 10 security mechanisms, then hid malicious instructions in docs, MCP tools and dependencies to see if the agent would execute them while still doing the real task It's like leaving fake notes from the boss around the office and counting how many the new intern follows The image is the map of that, each column is a harness and each row is a mechanism - AA auto-approve, the agent acts without asking - NI network isolation, blocks internet access - CAL command allowlist, only runs allowed commands - CDL command denylist, blocks dangerous commands like rm -rf - MCP permissions over which MCP tools it can use - PIF prompt injection filter - PR path restriction, only touches files inside allowed paths - RO read-only, can't modify anything - AL audit log, records what it did - PT project trust, decides whether to trust the repo config Green is on by default, yellow you have to turn on yourself, red doesn't exist and gray couldn't be determined Almost everything is yellow, and the closed-source products on the right are almost all gray Harnesses: - Claude Code - Codex CLI - Gemini CLI - gptme - Qwen Code - GitHub Copilot Model GLM-5.2 Benchmark 23 tasks, 2,500 runs With auto approve the attack worked 96% of the time, security you don't turn on doesn't exist

Let me explain in simple terms what this OpenAI Math repo is about I cloned it and started going through it carefully. There's still more to check, but this is what I have so far OpenAI has a new internal model and wanted to know how good it really is at math Normal math benchmarks were already too easy for it, so they gave it open research problems that professional mathematicians haven't solved, some of them stuck for decades They gave it around 4,000 problems. The model spent about 3 hours on each one and wrote its answers as real papers After filtering, they ended up with 722 papers grouped into 372 results, some claim to solve very well known problems The catch is that a long proof can hide a small mistake nobody sees, and checking one paper by hand can take an expert months That's where Lean comes in. Lean is a programming language where you write the proof step by step and the computer checks every step If it compiles, the logic is correct. Think of a teacher who never gets tired and never skips a line More than half of the results come with a Lean version I searched the Lean code for the usual shortcuts (sorry, fake axioms) and found none Those are the strongest results in the repo. The rest is basically "the AI says it works" and still needs humans to review it OpenAI itself marks the formalizations as unchecked One more catch. Lean checks that the proof is correct, but someone still has to confirm that the statement written in Lean is the same question as the famous problem Write the question slightly wrong and you can perfectly prove the wrong thing People are already finding cases where the Lean only covers part of what the paper claims Repo:


I’ve been testing Claude in Word, Excel, Power Point, and Outlook for a few months now Everything from day to day workflows to Open Telemetry and Compliance API data from these extensions has worked extremely well Excited to start testing Claude across the Google ecosystem now... Great move 👍