Ask ChatGPT what it thinks of your plan and you get a cheerful "ABSOLUTELY" back. That feels nice, and it tells you almost nothing you can act on. The original poster behind this thread hit that exact wall: a draft architecture, zero alternatives in sight, and a model too eager to please.
So this engineer changed the job, not the model, with one line: play devil's advocate and tear my idea apart. The objections came back sharp enough that the whole move got packaged as an installable skill, so the same process runs every time without retyping a word. The thread carries just four upvotes, which makes it feel like a small, quiet drop.
Here's why I find it worth your time: the real win isn't a clever prompt you can copy. It's turning a one-off prompt that worked into something repeatable you actually keep. Most of us find a good prompt, use it once, and lose it in chat history forever.
In partnership with
B2B support leaders who made AI work
AI agents are reshaping B2B support, and a few leaders are already ahead. Join Victoria Lowney of Inside Real Estate and Alex Rivas of HoneyBook as they share how they launched, scaled, and what's next. Hear how they built the business case, connected AI to data and systems, and reshaped team roles. Live Q&A. Register now to save your spot and get the recording.
*Ad
Why the cheerful yes is the problem
Most people treat a language model like a colleague who reads tone, context, and unspoken intent. It doesn't do any of that. Ask it "what do you think?" and the default behavior kicks in: agree, smooth the edges, hand you a warm serving of validation.
That's fine when you want a morale boost before lunch. It's poison when you're about to commit real time and money to a plan. The fix is almost insulting in its simplicity: give the model a different role and the entire character of the output flips.
You're not making the AI smarter. You're pointing its energy at a better target. Same model, opposite job.
The move: change the job, not the model
The workflow runs on four beats. Write your draft idea, plan, or architecture in plain words, no formatting required. Then tell the AI to play devil's advocate and tear it apart, not to review it with kid gloves.
Then you read the objections and sort them into two piles: real risks, or noise you can drop. When a run earns its keep, you save the pattern as a skill and fire it again next time. That final step is the quiet genius of the whole thing.
Think about how much good prompting evaporates. You craft the perfect instruction at 11pm, it nails the answer, and by next week it's buried under a hundred other chats. Packaging kills that waste.
Stop typing AI prompts. Start talking.
You think 4x faster than you type. So why are you typing prompts? Wispr Flow turns your voice into ready-to-paste text inside any AI tool. Speak naturally, tangents and all, and Flow cleans it up. Available on Mac, Windows, iPhone, and Android.
*Ad
Why 'installable' is the real upgrade
Sit with that word installable for a second. A prompt lives in your head or a scratch file, and it quietly rots. A skill lives in your tooling: versioned, named, and ready, so the same input triggers the same ritual every single time.
This matters more than any wording trick. Your sharpest thinking stops being a lucky accident you have to remember at the right moment. It becomes infrastructure you reach for on reflex.
And it travels. Hand the skill to a teammate and they get your reviewer, not a vague instruction to be skeptical. One person's hard-won prompt becomes the whole team's default.
A skill can store the prompt, but knowing when to hand a model the critic's job is a habit you build by doing it. 3 Minute AI runs on that idea: three-minute lessons that each end with a task you try in a built-in chat lab, including chatgpt and claude tracks where role swaps like this one get real reps. Daily lessons and two full courses are free.
Context is the whole game
One detail from the discussion stuck with me: context decides everything here. A contributor ran this on a live design doc and it surfaced three failure modes they had walked right past.
Strip the context out and the model just picks at edge cases that will never fire in the real world. Feed it your constraints, your goals, and what you already tried, and the objections sharpen fast. Stakes in, signal out.
The lesson underneath: a critic with no information is just a complainer. Arm it, and it becomes an auditor.
Give it a job, not a vibe
Another reader skipped the vague role entirely. Instead of be critical, they ask for the three most likely ways a design fails in six months, ranked by probability. A concrete job with a ranking beats an attitude every time.
I'd bolt one more thing onto that: for each failure, ask what you'd notice first. That turns a fuzzy worry into a warning sign you can watch for. Vague dread becomes a metric you can track.
State of Product reveals what AI still hasn’t solved
80% of product professionals say AI helps them ship faster, but customers aren’t seeing value any sooner. Atlassian’s State of Product 2027 explores the gains AI is delivering and the challenges that remain, from decision-making that hasn’t kept pace to gut instinct overriding customer evidence.
*Ad
The move that saves you from your own blind spot
Here's the part that sold me: ask the model to argue for the option you already rejected. Tell it to defend that discarded choice to your team, and defend it hard. It drags the tradeoff you skipped past back into the daylight.
If the rejected option still loses, good. Now you can explain the call in one sentence and trust it. If it wins, you just dodged a mistake nobody on your team was going to flag.
That single reframe is worth more than the fancy prompt that started all this.
Your one decision this week
Pick one decision you're about to lock in this week, hand it to your AI with the devil's advocate role, and ask for the three likeliest ways it falls apart in six months, ranked by probability.
If you want it to run the same way every time instead of rewriting the prompt from scratch, grab the installable skill and the exact failure-ranking prompt.
Worth 10 minutes if you'd rather hear the ugly objections from an AI today than from production six months out.
Credits to the original creator.
Five reads on hiring a professional skeptic
Anthropic's research on sycophancy digs into why language models drift toward telling you what you want to hear.
Gary Klein's project premortem is the classic exercise of imagining a plan already failed, then working backward to why.
The original devil's advocate was an actual Church office, paid to argue against a candidate for sainthood.
Red teaming covers how militaries and security teams formalize attacking their own plans before someone else does.
Claude's Agent Skills overview shows how a prompt you trust gets packaged into something installable and reusable.
