Prompts that actually work

OpenAI published six misalignment cases and a framework it runs itself. Anthropic disclosed four breaches and gave an outside body access to millions of transcripts. Both are more than a policy document and they are not the same thing. These six prompts are for reading any vendor governance claim and working out which kind it is.

Grok is the only assistant with live access to X, so the prompts worth using are the ones that need something that happened in the last hour. These six lean on that rather than treating Grok as a general assistant. Note that paid plans now run on one shared weekly pool across Chat, Imagine, Voice and Build, so a heavy research session eats the same budget as everything else.

Grok is the only assistant with live access to X, so the prompts worth using are the ones that need something from the last hour. These six lean on that rather than treating it as a general chatbot. One thing to budget for: paid Grok plans run on a single shared weekly pool across Chat, Imagine, Voice and Build, so a heavy research session reduces what is left for everything else.

All three write good code. What differs is how you run out — Codex resets every five hours, Claude shares one weekly pool with chat and Cowork, Cursor bills credits with a free mode. Claude Code limits settled lower on 14 September and OpenAI models leave Cursor on 12 November. These six prompts are for choosing on how you actually work rather than on a benchmark.

K3 is available from Moderato at 19 dollars a month, and K2.6 stays free and unlimited in chat without drawing on credits. These six prompts are built around the things K3 handles differently from a general assistant: long-context repository work, agent planning, and structured extraction. Test them on the free tier first, then decide which tier you need.

GPT-6 Astra launched with an executive calling it a generational leap, no published API pricing and no independent benchmarks. That combination is normal and it is also the point. These six prompts are for reading any model launch and working out which parts are measurements, which are projections, and which are marketing. They apply to every vendor equally.

A judge ordered Musk to produce Apple agreements today, publishers are suing Anthropic and OpenAI, and nearly every AI company is a defendant somewhere. Most of it changes nothing for anyone using the tools. These six prompts are for telling the difference between a headline and something that reaches the product you depend on. None of this is legal advice.

Mistral Large 4 shipped on 6 October 2026 with a 1-million-token context window, a 1.6B-parameter vision encoder, training across 160+ languages, and output tokens at $4.18 per million -- under half what Gemini 4 Argon and GPT-6.1 Sol charge. These six prompts are built around those four facts specifically. They are not generic prompts with the model name swapped in: each one does something that is either cheaper or only possible on Large 4. Paste them as-is, replace the bracketed parts, and read the note under each on where the model is weak.

Fable 5.1 reports roughly double its predecessor on a science benchmark at twice the price of Opus 5. A benchmark that doubles and a workload that does not improve is a common outcome, and vendor numbers cannot tell you which you are looking at. These six prompts are for testing a model on your own work before you pay more for it.

Suno just launched label-backed models, publishers are suing over training data, and a court has already ruled on how that data was acquired. If you sell work an AI helped produce, the question is what you can establish about it. These six prompts are for finding that out before a client asks. None of this is legal advice and none of it replaces a lawyer.

Perplexity cites a source for every claim, which is the whole point and the thing most people underuse. These six are built around checking rather than asking — tracing where a claim originated, finding informed disagreement, and separating what a source actually says from what it is being quoted as saying. They work on the free tier.

Apple spent its whole event arguing that on-device processing is the point, and for personal data it has a case. But most people still paste client documents, salary data and medical details into a cloud assistant without thinking. These six prompts are about getting the help without handing over the data — redacting properly, working from structure rather than content, and knowing when to stop.

ZCode packaged 42,411 files and full Git history into encrypted uploads at startup, and two settings that looked like privacy controls governed something else entirely. Nobody noticed for months. These six prompts walk through auditing whatever tool you currently run, using the machine in front of you rather than trusting a privacy page.

K2 Horizon shipped with its training data published. Kimi, GLM and Qwen did not. A federal court has already held that acquiring training material through piracy is unlawful, and Sony and Warner sued Anthropic on that question last week. These six prompts are for establishing what you can and cannot say about the models in your stack, before someone asks. None of this is legal advice.

ChatGPT is adding Sponsored Agents, where clicking an ad opens a labelled conversation with a business. Labelling solves the obvious problem and not the harder one — whether the unsponsored answer changes because sponsorship exists. These six prompts are for getting a recommendation you can trust, from any assistant, regardless of how it is funded.

October 2026 produced a whole category of small single-purpose models while most teams kept running frontier models on narrow, repetitive tasks. These prompts work out whether the model you are paying for is the one the job needs. Paste in your task description, a few real examples and your monthly volume.

Salesforce in Claude went to open beta on every paid plan this week, which means an administrator somewhere is about to enable it for a sales floor. The failure modes in a team rollout are different from individual use — shared budgets run out, permissions surprise people, and habits form before anyone checks the output. These six prompts are for the hour before you flip the switch.

Salesforce in Claude ships 37 prebuilt sales skills to paying enterprise customers. Most of what they do is available to anyone willing to paste in their own context — the plugin buys live data and consistency across a team, not capability. These six cover the recurring work: call prep, deal review, forecast narrative and the follow-up nobody does.

OpenAI put water systems and local government defenders first in its Daybreak pilot, and paired the access with training — because the constraint on a one-person security team is never detection, it is capacity to act on what gets found. These six prompts are built for that: triage over enumeration, and a defensible order rather than a longer list.

OpenAI's dots went live on 29 September: always-on agents with their own cloud computer and browser, able to sign into websites using saved passwords. The safety limits that matter are that background research runs read-only and password changes stay user-controlled. Everything else is a decision you make once and live with. These six prompts turn that decision into something you have actually thought about rather than clicked through.

arXiv capped every submitter at two papers a month on 1 October, after submissions went from 20,569 to 40,363 in two years and September alone generated nearly 9,000 support tickets. The underlying break is not about preprints. Every review process - code review, hiring, support, moderation, submissions - was calibrated on producing a plausible artefact being expensive. It is not any more, and rate limits are what you reach for when the real fix takes a year. These six prompts find the better fix.

System76's COSMIC desktop banned LLM-generated pull requests on 2 October - source code, comments and descriptions - and the stated reason was review workload rather than code quality. arXiv capped submissions at two a month the day before. The rules are now changing project by project, and the checklist COSMIC published asks for something tools cannot supply: that you understand your change and can answer questions about it. These six prompts get AI-assisted work to that standard, and tell you when to disclose.

Nvidia published OpenShell under Apache 2.0 on 28 September at version 0.1.0, with more than 100 partners named. A hundred logos and a permissive licence are not evidence that it works, and a security dependency is the worst place to learn that. These six prompts pull apart what a new safety tool actually guarantees, what it only appears to guarantee, and what it costs you to find out.

Anthropic is spending $100 million to train 10,000 engineers in one specific skill, and it is not prompting. The Claude Frontier Academy curriculum runs use case selection, through security review, to handover - and it states that prior AI agent experience is not required. That tells you where the bottleneck moved: not what the models can do, but getting them into production inside an organisation that has rules. The academy is nomination-only. The skill is not. These six prompts work through the same sequence.