I built a Devil's Advocates skill

W

When you decide alone, your strongest critic is missing. Friends are busy or too polite, and the AI you ask just agrees with whatever you already lean toward, or the push back is very soft. So the brain settles into "this feels right" the moment you have half-committed to a direction. That is how I end up shipping things I should have killed and killing things I should have shipped.

So I built a Claude Code skill that provides this honest feedback. It is called Devil's Advocates. You bring it a real decision and before it argues with you, it ask you few claryfing questions. One of them always forces you to say out loud what could go wrong, what you would lose if this call is the wrong one. Half the time that round alone is the value, because most bad decisions are just badly framed ones. A vague question gets you vague answer.

Once the question is sharp, it dispatches five adversarial sub-agents in parallel. Parallel matters: each one answers without seeing the others, so no devil softens its attack to fit what someone else already said. Each comes at the decision from one fixed angle:

• A Skeptic hunts the fatal flaw.
• An Architect tells you that you are solving the wrong problem.
• A Scout digs for the upside everyone else is missing.
• A Stranger reacts with zero context, like a first-time reader.
• An Operator only cares about what you do Monday morning.

After all five answer, their responses get shuffled, anonymized, and each devil peer-reviews the others blind. Then a Warden running on Opus reads all five arguments plus all five reviews and writes the verdict. Anonymizing the peer review kills positional bias and the curse of knowing whose argument is whose.

The shape of it (parallel advisors, blind peer review, one chairman who synthesizes at the end) I borrowed from Andrej Karpathy's LLM Council. The five named roles, the sharpening round, the low-signal guard and the file output are my own additions on top.

The output is one markdown file: TLDR at the top, the concrete first step right under it (one thing, not a list), where the devils agreed, where they clashed, what only surfaced through peer review, and the full transcript at the bottom. The Warden is also told to refuse confident answers when the inputs are weak – if the five devils came back with generic, interchangeable arguments, it returns a "low-signal" note and asks you to re-frame the question. Better than slop with confidence.

So when do you actually reach for it? Not for simple questions. It is great for something when you could lose money or time (opportunity to do something else instead). You reach for it when a question is bothering you for days. A few examples:

• Should I learn APEXLang now and be the pioneer, spending months hitting every rough edge that a more patient developer could just wait out?
• Should I sit the APEX certification this year, when clients may not care about the badge and I could lose the fee, the study time and my nerve to a failed exam?
• Should I start a blog and chase ACE status to build a name in the community, knowing I might pour evenings into writing for an audience that never arrives?
• Should I run a local AI model to keep client code private, taking on the hardware cost, the setup time, the ongoing maintenance and the security burden of guarding it myself?
• Should I start a SaaS project with a friend, knowing we could pour months of effort into it before we ever land a single paying customer?

One warning, it is not cheap. 11+ parallel sub-agents per run, around 50–100k tokens, 2-5 minutes wall time.

That is all for today. Pass it to a friend if you like this.