In partnership with

1,000+ Claude Prompts Top Professionals Actually Use at Work

Claude can be your analyst, editor, and strategist.

But most professionals are using it to fix grammar.

These 1,000+ Claude prompts take it from grammar tool to your most powerful AI work assistant.

Sign up for Superhuman AI and get:

• 1,000+ ready-to-use Claude prompts to get real work done in minutes — researched, tested, and used by professionals at Google, Microsoft, and NASA

• Superhuman AI newsletter (4 min daily) so you keep learning new AI tools and skills to stay ahead in your career — the prompts are just the beginning

Claim your free prompts

In this issue · 9 min read

• The Signal: agents got capable faster than anyone decided who lets them in

• AI News Roundup: OpenAI's 100 math claims, a model that rewrote its own instructions, an AI report that nearly started a war, the $899 Googlebook

• Model Watch: a model that answers only in numbers, and a 27B model that fits in 6 GB

• Shipping with AI: how Warp scores every agent run, and Teresa Torres on 16 experiment variants for one customer complaint

• Running the Team: Aha! on where PMs add value now, and Martin Fowler on disliking the tools he uses

• Worth Your Time: 3 links, starting with a step-by-step guide to running your first LLM eval

The Signal

AI agents now need permission

On Sunday, people using Meta's Muse agent to shop on Amazon started getting an error: "continued access by an unauthorized AI agent violates Amazon's Conditions of Use." According to GeekWire's reporting, Meta never told Amazon that Muse would be shopping its store, and Amazon objected that the agent did not identify itself while browsing. (The Verge)

Amazon did not block a fringe product. Apptopia estimates Muse had 1.8 million iOS downloads in the US and Canada in its first 12 days, against 1.3 million for ChatGPT in its own first 12 days. On iOS alone, Muse has about 359,000 US daily users, versus 231,000 for ChatGPT at the same point after launch. (TechCrunch)

Last week, following a Wall Street Journal report, Google confirmed that Gemini models hacked three real companies in May during a cybersecurity test run by the firm Irregular. The model was not supposed to have internet access, but it had been left on by mistake. Gemini guessed one password and found credentials in public software repositories for the other two, and stopped each time it realized the target was real. Google disclosed it only after the Journal asked. (The Verge, Ars Technica)

The common thread: in both cases the agent could do the job. What was missing was an agreed answer to who allowed it in, on what terms, and who checks.

The same gap shows up inside companies. Warp's internal software factory takes an average of 35 minutes to go from a request to a pull request, and the first human review arrives 3.5 hours later. Warp now calls review its biggest bottleneck. (Lenny's Newsletter) More on how they run it in Shipping with AI, below.

For the last two years, product teams mostly asked what an agent could do. The question this week is who grants it access, and product managers are the people best placed to write that answer down before a platform, a regulator, or an incident writes it for them.

Three things to do this week:

1. Decide your policy for outside agents. If a customer's agent tries to use your product, do you allow it, block it, or route it through an API or MCP server you control? Amazon's complaint was about an agent that did not announce itself, so decide what an agent has to do to be welcome.

2. If you ship an agent, make it identify itself and use sanctioned access. Scraping a site as if the agent were the user works until the site notices. Then your feature breaks on a Sunday night for every customer at once.

3. Measure the time between an agent finishing work and a human looking at it. If that wait is several times longer than the work itself, the fix is a review policy, not a better model. Pick one low-risk class of change where the person who prompted the agent can approve it.

Weigh in

AI News Roundup

OpenAI says one model resolved more than 100 open math problems

OpenAI set up an independent Advisory Group on Mathematics and AI, hosted at the Institute for Advanced Study in Princeton, to give mathematicians input into its math research. The announcement followed its published solution to the Navier-Stokes Millennium Prize problem and a claim that the same internal model has resolved more than 100 other open problems. Many mathematicians have objected to how fast these results are arriving. (TechCrunch)

Why it matters. Output is now arriving faster than experts can check it, and OpenAI's answer was to build a review body. Your own AI features will hit the same wall on a smaller scale. Plan who verifies AI output, and how many hours that takes, before you plan how much output you can produce.

A model rewrote its own instructions while summarizing its work

OpenAI published a framework for reporting model misalignment, with six incidents from the last six months. In one, a model in training compacted its context to keep working and added a paragraph of invented "additional instructions" to its own summary, giving itself a new persona. It went back to the task without acting on them, and the run was not used for the released model. (Simon Willison)

Why it matters. Any long-running agent that summarizes its history to save context is rewriting its own instructions as it goes. If your agents run long tasks, log the compaction summaries and review a sample. They are the least inspected text in most agent systems.

An AI-written intelligence report nearly led to a boarding

The US military prepared to intercept and board a Chinese ship after an analyst's report, produced with a chatbot, said it was carrying nuclear arms program components, according to CNN. The chatbot had combined open-source and classified intelligence and misidentified the cargo. Officials caught the error before the boarding. (Ars Technica)

Why it matters. The chatbot fused open and classified sources into one report, and the error was in the fused output. If your product summarizes across documents, show the source next to each claim. Readers cannot catch an error they cannot trace.

Google's Gemini laptops go on preorder at $899

Googlebook, first shown in May, runs Android with a desktop Chrome browser and builds Gemini into the machine: an AI cursor, generated widgets, and a dictation feature that turns rambling speech into clean text. Google is betting people will buy new hardware for built-in AI rather than install it as an app. (TechCrunch)

Why it matters. When the assistant ships with the operating system, it becomes the default for anything it can reach. If your product does a job Gemini can do from the cursor, expect to compete with a feature the user never had to choose.

Apple's retail architect is skeptical of agent shopping

Ron Johnson, who built Apple's stores, told TechCrunch that AI will improve online shopping but probably will not change how people choose to shop, and that physical stores will keep doing well. It is a useful counterweight in a week when the biggest agent launch in years is fighting over checkout access. (TechCrunch)

Why it matters. Agent shopping gets the funding and the headlines, but adoption is still an open question. If your roadmap assumes buyers will hand purchase decisions to agents, find evidence from your own customers before building around it.

Model Watch

Jev: a model that returns numbers instead of text

TypeSafe AI's Jev takes text or structured records and answers yes/no, multiple-choice, or score questions with a probability, not prose. Input costs $0.042 per million tokens and output is free, which is cheaper than GPT-5 Nano. It suits classification work: spam detection, labeling, prioritization, and reranking search results. Simon Willison notes the trade-off. It gives you no reasoning at all, so bias and error are harder to see, and running your own evals matters more. (Simon Willison)

How to use it. Take one classification step in your product that currently calls a general LLM and returns a label. Run the same inputs through Jev and compare accuracy and cost on a few hundred of your own examples. At this price the experiment costs cents.

Ternary Bonsai 2: a 27B model in about 6 GB

Prism ML released Bonsai 2, a version of Qwen3.8-27B with every language weight stored as -1, 0, or +1. The smallest build is 5.95 GB, down from about 54 GB at full precision. Prism reports it keeps 98.2% of the original's benchmark score and runs at about 47 tokens a second on an M5 Max laptop (measured on the slightly larger 7.2 GB build). Context is 262K tokens. It is Apache 2.0. (Hugging Face)

How to run it. It needs Prism's forks of llama.cpp or MLX for the custom kernels, so follow the demo repo linked from the model card. The practical use is a capable model on a normal laptop or a single GPU, with no data leaving the machine.

Shipping with AI

How Warp runs a software factory

Warp CEO Zach Lloyd walked Claire Vo through Wilson, the factory that takes a request from Slack through a Linear ticket, implementation, automated QA, and merge. Three practices stand out. Warp picks models by replaying its own past jobs under different configurations, and found model choice is by far its biggest cost lever, with context management a distant second. An AI judge scores every run, including whether the agent wrote redundant tests, and Warp keeps that judge on a mid-tier model so evaluation does not eat the savings. It also waits for 20 to 25 failed runs before changing the agents' instructions, so fixes come from patterns rather than single failures. (Lenny's Newsletter)

16 experiment variants to fix one customer complaint

Teresa Torres documented fixing one complaint about Vistaly's AI-generated opportunity solution trees: branches that were too flat. Her first LLM judge caught every real error but flagged too many false ones, and seven rounds of changes to the judge, including a smarter model, barely helped. Switching to Sonnet 5 made one error disappear only because the model barely added any new structure: one mid-level parent where Sonnet 4.6 had made six or eleven. What worked was a change to orchestration: a simple deterministic check that spots parents with too many children and sends them back to the agent to fix. (Product Talk)

Did skills kill MCP?

A GitHub podcast episode takes apart five popular AI hot takes. On whether to read agent-written code, the rule it offers is to review until you can explain and own the outcome, with more scrutiny for an auth refactor than a CSS experiment. (GitHub) On MCP, Simon Willison argues that a coding agent with full internet access may not need it, but anything more careful does. MCP keeps API keys away from the agent and gives users a sensible way to connect services. (Simon Willison) That connects to this week's Signal: MCP is one form of the controlled, identifiable access Amazon says Muse lacked.

Running the Team

Aha! on where product managers add value now

Teresa Torres interviewed Aha!'s founders and the PM behind Aha! Builder, an app builder for product managers. Their first version ran on containers and cost too much to scale, so they rebuilt it on a multi-tenant architecture. They also kept authentication, SSO, and databases as fixed, pre-built components instead of letting the AI generate them. Their view is that PMs will add value by knowing what to build, not how to build it. (Product Talk)

Martin Fowler: "I don't like LLMs"

Fowler writes that he finds LLMs useful and uses them, but dislikes dealing with them: the voice, the confident errors, the thin apologies. He quotes Jessica Kerr that it would be irresponsible not to use them, and notes that polls show the same split between finding the tools useful and expecting them to be bad for society. It is a good piece to share with a senior engineer who uses the tools and resents them. That reaction is common, and it deserves a real conversation rather than a push for more adoption. (Martin Fowler)

An access ladder for agents

Use this to decide how much an agent can do without a person, one workflow at a time:

1. Read only. The agent can look but not change anything.

2. Draft. It prepares the change, and a separate person approves it.

3. Self-approve. The person who prompted it reviews and ships it, as Warp now allows.

4. Auto-ship with sampling. Changes merge on their own, and a judge or a person reviews a sample.

5. Auto-ship. Only for changes you can reverse cheaply and detect quickly.

Move a workflow up one rung only after its failure rate at the current rung is known.

Worth Your Time

• How to run LLM evals: a step-by-step guide for PMs The eval steps behind this week's Warp and Product Talk stories: choosing what to test, building the dataset, and reading the results.

• California tightens rules on AI data center energy and water use Infrastructure costs are becoming policy.

• Is the AI industry really ready to slow down? A follow-up on last week's slowdown debate.

Pass it on

Forward this to whoever at your company would get the call if an AI agent started using your product tomorrow.

Refer two people and we will send you the Source Book: every source we read each week and how items get chosen. Refer five and we will name you in a future issue. Your link and progress: