I need a board of grounded advisors for my big decisions.

Why a costume isn’t a consultant

There’s a tempting shortcut everyone tries: tell the model to “act as a world-class strategist” and ask your hard question. It feels like consulting a board of experts. It isn’t. A persona prompt doesn’t add knowledge — it relabels the model’s own opinion in a more authoritative voice. The substance is identical; only the costume changed. Asked to “be skeptical” or “are you sure?”, a model will often just cave and agree with whatever you imply, because agreeableness reads as helpfulness. A self-critic grading its own work scores far better than the same model checked by something it can’t see around. The lesson underneath all of it: a model reviewing itself launders its own bias. It can’t be the independent check because it’s the same mind twice.

What actually moves the answer is grounding — pointing the advisor at a real, named body of work and making it reason over that, citing the specific rule each call rests on. The expertise lives in the canon, not in the prompt. Swap the source and the advice genuinely changes; that’s the tell that something real is happening. This is the one rule the whole system obeys: an advisor’s edge is the body of knowledge it reads, never the persona it wears.

One advisor, then a board

Start with a single advisor on the one decision that matters most. Curate the canon, write a fixed pass that forces citations, and run your decision through it. That alone beats the costume, because now the read is anchored to something outside the model’s guesswork.

A board is the same engine, multiplied where the decision needs it. A real call — change the offer, reposition, raise the price — touches several fields at once, and a single advisor reading a single canon only sees its own slice. So you convene the lenses that apply: offer-and-economics, sales, demand-and-positioning, copy, and the architecture lens that judges how the pieces fit. Five at most, and never all five by reflex — only the ones whose canon the decision actually touches. Each reads the decision through its own field. Then a chair synthesizes.

The chair is where most “AI council” setups quietly cheat. They average the votes, which buries the one objection that mattered under four that didn’t. This board doesn’t vote. On disagreement, the chair asks which constraint is binding — the single gate that, if you get it wrong, makes the rest moot — and lets that lens govern. When two lenses genuinely conflict and neither clearly binds, the chair stops and hands the call back to you instead of inventing a resolution. That hand-back is the honest move. The board’s job is to surface the tightest constraint and the blind spot you’d have missed; the decision stays yours.

The trap to design around

The danger with a board isn’t too few opinions — it’s false consensus. Five grounded lenses can all agree and still be wrong together, because they share a model and its blind spots underneath the different canons. Agreement among them is reassurance, not verification. The real check is an outside signal: a source the board didn’t generate, and ultimately whether the call worked. Treat unanimity as a prompt to look harder, not a green light.

The advanced version adds a factory — a workflow that stamps new advisors to the same grounded shape, each shipping with a held-out check that proves it reasons over its canon rather than parroting it — and an auto-fire that runs the relevant lenses on every substantial decision as a logged, non-binding read. The factory lets the board grow with the business. The log lets each advisor earn trust the slow way, on a track record you can audit, instead of by sounding confident. Keep two things wired in from day one: a refresh schedule so no canon silently goes stale, and yourself as the final committer. The board pressure-tests the decision from every angle that matters. You still make it.