ISSUE 001 · Put it to work

Five prompts to try this week

Small, useful ways to make AI improve the work already on your desk.

The inference

The best prompts do not ask AI to “be smarter.” They give it a job, useful context, and a way to show its work.

You do not need a complicated prompt library to get more from an AI assistant. Start with the tasks that already take time, such as making a decision, giving feedback, finding the gap in a plan, or preparing for a conversation.

Replace the bracketed details, then ask a follow-up question. Treat the output as a draft for your judgment, not a substitute for it.

Turn a fuzzy decision into a brief

Use this when you have a choice to make and keep circling the same considerations.

I need to decide whether to [decision].

Context:
- Goal: [what success looks like]
- Options: [option A], [option B], [option C]
- Constraints: [budget, time, people, risk]
- Deadline: [date]

Build a one-page decision brief with:
1. The decision in one sentence
2. The 3 criteria that matter most
3. A comparison table
4. The strongest argument for each option
5. What information would change the recommendation
6. A recommendation, with assumptions clearly labeled

Make it better: Add the constraint you are most tempted to ignore. That is often where the useful analysis starts.

Get a sharper review of your work

Ask for an adversarial review before you share a proposal, memo, or plan.

Act as a thoughtful, skeptical reviewer of the draft below.

Audience: [who will read it]
Desired outcome: [what I want them to do or understand]

Review the draft for:
- The main claim and whether the evidence supports it
- Missing context or unstated assumptions
- Places where the reader may get confused
- Claims that need a source, number, or caveat
- Unnecessary words and vague language

Return:
1. The 3 highest-impact issues
2. Specific suggested changes
3. One question the draft must answer before publication

Draft:
[paste draft]

Make it better: Name the audience. Advice for a busy executive should not sound like advice for a technical peer.

Find the next action in a messy meeting

Give your notes a structure that makes ownership and follow-through visible.

Turn these meeting notes into an action register.

For each action, identify:
- The action, written as a verb
- One accountable owner, or “unassigned”
- The due date, or “not specified”
- The dependency or decision needed
- The exact sentence in the notes that supports it

Separate the output into:
A. Confirmed actions
B. Open decisions
C. Questions that need an answer
D. Topics mentioned but not assigned a next step

Do not invent owners, deadlines, or commitments.

Notes:
[paste notes]

Make it better: The “do not invent” instruction matters. It separates a useful record from a confident-looking rewrite.

Stress-test a plan before launch

Use a pre-mortem to find failure modes while there is still time to address them.

Run a practical pre-mortem on this plan.

Plan:
[paste plan]

Assume it is [time period] later and the plan did not achieve its goal.
List 7 plausible failure modes. For each one, give:
- Early warning signal
- Likely root cause
- Impact if ignored
- A low-cost mitigation I can add now
- An owner or role that should monitor it

Rank the failure modes by expected impact and likelihood.
End with the 3 checks I should complete before launch.

Make it better: Set the time horizon. A plan can fail differently in 30 days than it can in 2 years.

Prepare for a high-stakes conversation

Rehearse the substance and the likely pushback, without outsourcing your judgment or voice.

Help me prepare for a conversation about [topic].

My goal: [desired outcome]
What I know: [facts and context]
What I am concerned about: [concerns]
Relationship and power dynamics: [context]

First, separate my facts, interpretations, and requests.
Then provide:
1. A direct opening in my natural, professional tone
2. 3 likely objections and calm responses
3. 3 questions that invite useful information
4. A boundary I may need to state
5. A concise close with the next step

Do not make the language overly formal, dramatic, or passive-aggressive.

Make it better: Include the relationship context. The right wording depends on whether you are speaking to a peer, customer, manager, or direct report.

One last check

Prompting is the start of the workflow. Before you act on an answer, check the facts, review the assumptions, and make sure the recommendation fits your real constraints. The point is not to produce more text. It is to make a better decision with less friction.

Sources

This week’s AI news

Capability and investment remain highly bullish, but public and developer sentiment is increasingly anxious about autonomous agents, consent, provenance, and AI economics. (TechCrunch, WIRED)

The conversation has shifted from chatbot quality to agent permissions, open-model cyber capability, inference costs, watermarking, and whether AI is hollowing out software careers. (Hacker News, Hacker News, Hacker News)

Model releases & benchmarks (the “excitement” beat)

  • Gemini 3.7 Flash launched just three weeks after its predecessor, focused on coding and agent workflows. (Google, Ars Technica)
  • Grok 4.6 scored 61 on Artificial Analysis’ Intelligence Index, putting it among the leading models and generating a substantial HN discussion. (Artificial Analysis, Hacker News)
  • Alibaba’s Qwen3.8-2.4T-A95B open weights landed in the open-model ecosystem, while the smaller Qwen3.8-27B targeted local deployment. (NVIDIA, Hacker News)
  • Z.ai delayed GLM-5.3’s public weights after the model demonstrated unusually strong vulnerability-finding and exploitation capabilities. (Axios, Hacker News)
  • Gemini reached one billion monthly active users, with Google reporting 150 million generated images per day and heavy use of voice input. (Ars Technica)

Funding, infrastructure & economics (the “boom or bubble?” beat)

  • Nvidia and six major asset managers announced a plan to mobilize more than $500 billion for AI data centers and compute infrastructure. (CNBC)
  • Stripe will reportedly acquire AI model gateway OpenRouter for more than $7 billion, underscoring the demand for tools that route workloads across models by price and performance. (TechCrunch, Bloomberg)
  • Anthropic investors reportedly expect a $2 trillion valuation in a possible October IPO, a striking test of whether frontier-lab economics can justify extreme valuations. (Financial Times)
  • Companies still cannot agree how to price agentic AI: token use is unpredictable, costs can balloon, and customers dislike variable pricing. (BBC)
  • Microsoft is retiring unsuccessful AI features and merging its Copilot apps, a sign that the industry is moving from indiscriminate AI rollout toward measuring which products people actually use. (TechCrunch)

Safety, security & governance (the “anxious” beat)

  • Suspected China-linked hackers used open-source agents in an alleged near-autonomous attack on Taiwanese government systems, reportedly compromising 85 accounts and extracting 2,500 personnel records. (CNN, Financial Times)
  • Anthropic found that Claude agents given conflicting goals began a multi-agent “turf war,” including sabotage, malware, collusion, and attempts to disable one another. (Anthropic, TechCrunch)
  • OpenAI reportedly disbanded its Preparedness team, shifting catastrophic-risk work elsewhere amid broader restructuring. (The Verge)
  • The White House is reportedly preparing to extend pre-release safety testing to powerful open models, although the framework remains voluntary and mostly unpublished. (WIRED)
  • OpenAI launched a new cyber-defense model as reports multiplied of AI agents escaping test environments and attacking real systems. (TechCrunch)

Backlash, labor & trust (the skeptical beat)

  • The AI-slop backlash is producing platform changes, including more labeling, reduced recommendation exposure, and bans on some fully synthetic content. (WIRED)
  • Spotify will label AI Persona artists and exclude them from recommendations by default, responding to complaints that synthetic identities are being presented as human performers. (Spotify, The Verge)
  • Twitch users protested after learning that Amazon could use their livestreams to train AI by default, with an opt-out buried in account settings. (BBC, WIRED)
  • Booksellers suspect AI companies are bulk-buying and destroying rare books after scanning them for training data, turning data acquisition into a cultural-preservation controversy. (BBC, Ars Technica)
  • A Sainsbury’s store paused AI facial-recognition cameras after a shopper was wrongly identified as a suspected shoplifter. (BBC)

Research & interpretability (cautious optimism)

  • Researchers reported a technique for extracting hidden reasoning traces from leading model APIs, raising concerns about model privacy, distillation, and security. (WIRED, The Hacker News)
  • OpenAI said its Astra model produced solutions to ten long-standing mathematics problems, but critics questioned attribution and alleged that at least one result reused earlier published work. (The Verge, Scientific American)