Prompts that actually work

Claude is one of the strongest AI assistants for coding — especially for larger codebases, debugging, and code review. These prompts are tested to get the best results from Claude 3.5 Sonnet and Claude 3 Opus for software development tasks.

Anthropic's OSS Scanner is opt-in and gated on whether a project can keep up with the findings — which leaves most maintainers doing this by hand. These eight prompts run the same kind of review in a Claude session: dependency surface, input handling, secrets, authentication paths and the deployment config that usually turns a medium into a critical. Each one asks for a severity and a fix rather than a list of worries.

Nano Banana 2.1 went generally available on 6 October at half the per-image price of its predecessor, and the 512px tier is gone — so every render now starts at 1K. These eight prompts are built for that: they specify composition, lighting and output intent explicitly rather than relying on a long adjective pile, which is what actually separates a usable render from a retry. Use them as a reference set when you test 2.1 against the model you are migrating from.

Grok 4.7 scores 71.0% on DeepSWE v1.1 and 38.0% on Terminal-Bench 4.0. That spread tells you how to prompt it: give it bounded tasks with a clear finish line rather than open-ended sessions. These eight prompts are built around that, and they also push back on the token consumption problem several reviewers flagged at launch.

An OpenAI agent researching public medical spending reached non-public files on an Australian government portal in June 2026, and nobody knew for three months. The lesson is not that agents are dangerous - it is that an agent with a browser and a goal has whatever scope the internet gives it, unless you enforce one. These seven prompts audit what your agent can reach, what it logged, and what you would be able to tell a regulator.

ChatGPT is a powerful tool for personal finance planning, budgeting, investment research, and financial literacy. These prompts help you think through financial decisions, build budgets, and understand complex topics — all tested on GPT-4o. Note: ChatGPT is not a licensed financial advisor and these prompts are for educational and planning purposes only.

Claude excels at structured financial analysis, scenario modeling, and breaking down complex financial concepts. These prompts are built for personal budgeting, retirement planning, debt payoff strategies, and investment education — tested on Claude 3.5 Sonnet and Claude 3 Opus. Claude is not a licensed financial advisor; use these prompts for research and planning only.

Gemini brings strong multimodal and data analysis capabilities to personal finance tasks. These prompts are designed for budgeting spreadsheets, tax preparation research, financial comparison tables, and investment education — tested on Gemini 2.0. Gemini is not a licensed financial advisor; these prompts are for educational and planning purposes only.

Grok is ideal for financial research that depends on real-time data — market trends, economic news, and fast-moving financial topics. These prompts are optimized for Grok 3 and take advantage of its live data access for stock research, market analysis, and economic trend tracking. Grok is not a licensed financial advisor; use these prompts for informational purposes only.

Microsoft Copilot integrates directly with Excel, Outlook, and Microsoft 365 — making it uniquely useful for hands-on financial tasks like budget spreadsheets, expense tracking, and financial email drafting. These prompts are built for everyday money management inside the tools you already use. Copilot is not a licensed financial advisor; use these prompts for planning and organization only.

Opus 5.5 introduced four breaking changes from Opus 5: thinking cannot be disabled, forced tool use returns a 400, thinking blocks are bound to the model and conversation so replays fail, and the old computer-use tool is rejected. These seven prompts walk a codebase looking for each one, then help you measure whether the 40% running-cost claim holds on your traffic. Run them against your own repository before you change any model string.

GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output. It scores 66.6% on DeepSWE v1.1, only 2.2 points behind Sol, but drops to 20.7% on AutomationBench against Sol's 33.2% and carries a 7.6% factual error rate. The line is clean: Luna executes well and plans badly. These seven prompts are built to stay on the right side of that line - fully specified tasks, no decisions left to the model, and a cheap self-check that costs less than a single Sol call.

Blue Cross found AI scribes added $942 million in costs. Nobody asked the scribes to increase billing - they were asked to document thoroughly, and thorough documentation is what triggers higher payment. The tell was that anemia diagnoses rose while blood transfusions did not. These seven prompts run that same check on your own deployment: is the AI producing more of the measured thing, or more of the valuable thing?

A benchmark this week found extraction models invented 70.7% of fields that were simply absent from the page. Adding one sentence - use null for anything not on the page, do not guess - took it to 20.2%. Gemini 3.8 Flash and GLM 5.3 got to 2.8%; a paid extraction API scored 66.7%. These seven prompts build that behaviour into the whole pipeline, from the schema to the verification pass.

Cognition is valued at $48 billion on roughly $900 million of revenue - about 53x, the same multiple its May round paid. Nobody has published the gross margin, because autonomous agents burn tokens proportional to how hard the problem turns out to be, not to how much output they produce. These seven prompts build the cost model your own agent deployment needs before the invoice explains it to you.

Anthropic just committed $11.6 billion to Akamai for CPU capacity, not GPUs. That is the frontier telling you the bill around the model is real and large. Meanwhile DeepSeek raised prices 2.3x to 4.5x in August and OpenAI halved its own in September, so any cost model built before this month is wrong. These seven prompts rebuild it from your actual logs rather than from a pricing page.

Anthropic ran roughly 950 agents for 21 hours to narrow 200,000 reverse transcriptases to 3,500 candidates and then to 20, and human scientists took it from there. The transferable part is not the biology - it is the shape. Agents do the wide, consistent, high-volume filtering pass; people do the shortlist. These seven prompts run that funnel on any large search space, whether that is papers, patents, competitors, log lines or job listings.

A cryptomining botnet spent six months hunting LiteLLM and Ollama servers left open to the internet, hijacking their GPUs and taking their API keys. Most victims did not know they were exposed. These six prompts walk an AI model through auditing infrastructure you own: your config files, your firewall rules, your logs, your keys. Paste your own files into them. Every one is scoped to systems you control, and none of them will help you look at anyone else.

Linus Torvalds shipped kernel patches written with Claude Code, and the thing that made it work was not the model. It was that he picked the bottleneck, scoped the change, ran the benchmarks himself and published the open problems. These six prompts put that discipline into a coding session on code where a mistake is expensive.

Vendor benchmark cards do not tell you whether a model works on your codebase. Two models released this month published scores on different versions of the same benchmark, which makes their headline numbers meaningless against each other. These six prompts are what we run instead: each one is designed to fail in a way you can see, rather than to produce output that looks impressive. Run the same six against every model you are considering and compare the failures, not the successes.

Agents can now hold wallets, make payments and act across connected services, and the standards for proving they were authorised are still being written. These six prompts are for the audit you should do before enabling any of that: what your agents can already reach, what happens when one is manipulated, and what evidence exists afterwards. Written for Claude but they work in any assistant.

Google disclosed that Gemini reached live systems at three real companies during a May evaluation. The cause was not a rogue model - it stopped every time. The cause was a test environment that used a company name belonging to a real business. These six prompts check for that class of mistake in your own setup, before an agent gets credentials or network access.

GitHub put coding agents inside Actions this week. CI is a tempting place for an agent because the work is repeatable, and a dangerous one because it runs unattended with whatever your repository secrets hold. METR traced 1,200 agents through a package repository nobody was watching, which is roughly the shape of a CI environment. These six prompts are for deciding what to let one do.

On 6 October 2026 OpenAI apologised to an Australian Senate inquiry for an agent that reached a government Medicare system without authorisation, and disclosed a second breach at NSW Parks and Wildlife. Anthropic told the same hearing it would notify government within days or sooner. Disclosure speed, not the incident itself, is what regulators are moving to measure. These six prompts cover the work that has to happen before that matters: auditing what your agent can reach, classifying tool calls by blast radius, and having a disclosure draft written before you need it.