Hey everyone,
How much work would you let an AI agent do without watching it?
That’s the big question this week. Anthropic just announced Claude Duet, an agent that can keep working while your laptop is closed. There’s also a way to use Claude Code with a local model, plus a simple workflow for getting a second AI to challenge your first draft.
Let’s get into it!
AI News This Week: Claude Gets a Background Job
Anthropic has announced Claude Duet, a personal agent that runs on its own cloud computer and can keep working while your laptop is closed.
It can check connected accounts throughout the day, put together a morning brief, and draft next steps from your inbox, calendar, and workspace. You can keep talking to it in one ongoing conversation through text or voice.
Think meeting prep, scheduling, or turning scattered information into a presentation.
Here’s what I care about: when does it stop and ask? Anthropic says Duet asks for approval before sending a message, making a payment, signing a document, or deleting anything.
Show me exactly what you’re about to do, then let me decide. That’s what I’d want before handing an agent more responsibility.
For now, Duet is a research preview with a waitlist for US Pro and Max subscribers. Its background work counts toward your plan’s existing limits, and Anthropic says it can still make mistakes.
Tool of the Week: A Local Model Inside Claude Code
Want to try a coding model on your own machine? Ollama lets you use Claude Code with a local model.
You keep the coding interface, while a local model handles the requests. To be clear, this does not put Anthropic’s Claude model on your Mac.
Once you’ve installed Ollama and Claude Code, you can launch it with:
ollama launch claude --model qwen3-coderAnother option is Flint from Recursive, an open coding model available through Ollama:
ollama launch claude --model flintCheck your hardware first. The default Qwen3-Coder download is about 19GB; Flint’s is about 23GB. Your computer also needs memory for the model’s working context and everything else you have running.
The appeal: no per-token API fee for local inference. Start with a small coding task and see how it does on your machine before giving it a bigger job.
Free Sauce: Give Your AI Draft a Second Opinion
If your AI keeps giving you another version of the same mediocre draft, try this.
Have Claude write the draft, give ChatGPT the job of finding the weak spots, then take the useful feedback back to Claude. You can swap the tools around.
Here are the three prompts to save.
1. Write the draft and expose the weak spots
Task: [What you need, who it’s for, and what it should achieve.]
Write the complete first draft. Below it, add a handoff note for another reviewer. Explain the objective, the assumptions you made, the two weakest parts of the draft, and any important objections or missing evidence.
Clearly separate facts you can support from assumptions that need checking.
2. Give a second AI the review
Review the draft and handoff note below against the original goal. Do not rewrite it yet.
Identify unclear reasoning, unsupported claims, generic filler, and anything the intended reader would question. Include blind spots the writer missed.
Rank the changes by importance and explain why each one matters. If something works, say so. Don’t invent problems just to be critical.
[Paste the original task, draft, and handoff note.]
3. Bring the useful feedback back
Review this critique of your draft. Make the changes that improve accuracy, clarity, and usefulness for the intended reader. If you reject a suggestion, briefly explain why.
Return the finished version, followed by a short list of what changed. Flag any facts that still need verification.
[Paste the critique.]
Try this on your next important email or proposal. Give the reviewer the original goal so it can judge whether the draft actually does the job. You still need to check the facts yourself.
A Quick Security Check for Website Owners
If your website runs on PHP, send this to your hosting company:
Can you confirm that our server has the PHP2Shell security fixes installed and that you’ve reviewed PHP’s October 23 follow-up guidance? Please also confirm that the updated configuration is active in the running web workers.
PHP’s official advisories cover critical issues in the software underneath your website. Updating WordPress or your plugins alone may not address them. Your host can confirm whether it has applied the fixes or equivalent vendor patches.
And if you install third-party AI skills, NVIDIA’s SkillSpector helps flag suspicious instructions and unsafe behavior. Treat it as an extra check; a clean scan doesn’t guarantee a skill is safe.
Three Things Worth Saving
prompts.chat
A community prompt library you can connect to compatible AI tools through MCP. Grab a useful starting point, then adapt it to the job you need done.
Superpowers
If your coding agent jumps straight into code before you’ve agreed on the plan, look at this. It gives the agent a process for working through the idea, getting the design approved, and building and testing it.
Cloudflare’s security-audit skill
A security-review workflow that has another agent challenge suspected issues before they’re reported. Useful when you want the agent to check its findings before handing you a list of problems.
See you next week!
— Thomas