AI can produce a working feature in about 90 seconds, but that speed does not guarantee trustworthy software. The most useful coding prompts may be the ones that slow the process down, forcing the model to question assumptions before it writes code.
That idea sits at the center of a June 2026 discussion from Google Cloud’s developer relations team, which asked ten engineers and leaders about the prompt they refuse to work without. Their answers point to a shared habit: using AI as an adversarial second opinion instead of an agreeable assistant.
That distinction matters because many developers use prompts as one-off requests. “Most people’s prompt history looks like a junk drawer. A one-off request to explain an error message, a quick ” clean this up ,” a boilerplate generator used once and forgotten.” The prompts may solve immediate tasks, but they do not always challenge the code or the thinking behind it.
Prompts That Make AI Push Back
Maja Bilić, a Senior Outbound Product Manager at Google Cloud, uses a prompt that asks the model to argue with her before it writes code. The model lists considerations first, giving her a chance to examine trade-offs and possible problems before a solution takes shape.
The value is not that the model always finds the right answer. The value comes from changing the role it plays. Instead of rewarding the first idea with polished code, the prompt creates space for disagreement, questions, and review.
Andrew Brogdon, a Staff Developer Relations Engineer, applies the same thinking to testing. His prompts first examine code for test coverage and testability, then build tests. That order matters because it asks whether the code can be tested well before asking AI to produce a collection of tests.
Aja Hammerly, Director of Builder Relations, uses two separate prompts to review code. One searches for missing edge cases and race conditions, while the other looks for unused code and comments. Splitting the review gives each prompt a focused job instead of asking one broad request to find every possible problem.
These examples share a simple pattern. AI does not get the final word, and it does not begin with code generation. It first examines risks, gaps, and weaknesses that an enthusiastic assistant might overlook.
Speed Does Not Remove the Need for Trust
The pressure to use AI comes from real gains. Developers at Booking.com saw a 30% productivity boost after training. At Adidas, AI integration brought a 20-25% increase in developer effectiveness and a 50% increase in “happy coding time.” Another project built a remote healthcare platform in four days for just £165.
Those results help explain why AI coding demos attract attention. Mudit Singh, Co-Founder and Head of Growth at TestMu AI, observes that these demonstrations often end with a working feature in about 90 seconds. A fast result makes the technology easy to show, but a demo does not reveal every weakness that can surface after the code enters regular use.
At one organization, bugs increased after the AI rollout despite improved velocity and successful demos. That result points to a trust problem with AI-generated code. Teams can move faster and still face more defects if they accept generated code without a strong challenge process.
The Google Cloud examples offer a response to that problem. Asking AI to inspect test coverage, question design choices, find race conditions, and identify unused code creates checkpoints before speed turns into risk. These prompts do not reject AI-assisted development; they make review part of the workflow.
The Environment Matters as Much as the Prompt
Gaurav Singal, Chief Technology Officer at ConstructConnect, places the issue in a wider context. “The reason is that instead of concentrating on which training program to run, they should have been making sure the environment was ready for the tool and agents.” His point shifts attention from instruction alone to the conditions in which AI operates.
A useful prompt cannot fix every problem around it. Teams still need a setup where developers can inspect code, test changes, and question results. The figures from Booking.com and Adidas show what AI can support, while the rise in bugs shows why the surrounding process matters.
That also explains why the best prompts are not magic phrases. They are repeatable review steps that tell AI when to challenge an idea, what risks to search for, and when to wait before producing code. The prompt becomes part of the development process rather than a disposable request.
Google Cloud’s engineers arrived at the same underlying move from different directions: ask AI to disagree before asking it to help. For developers, that may be the most practical lesson in the rush toward faster coding. A working feature is useful, but software that survives testing needs a second opinion first.
Based on
- What We Can Learn From Google Engineers’ Indispensible Prompts — kdnuggets.com
- How Your AI Writes Code Faster Than Your Team Can Trust It — forbes.com
- SDD Fixed How AI Writes Code—Nobody’s Fixed How Humans Verify It — forbes.com
- 5 Real-World Vibe Coding Success Stories That Show What AI Can Really Do — forbes.com
- Agentic AI Adoption Starts With The Codebase — forbes.com
