name: dilemma-board
description: Settle a real dilemma (career move, build-vs-buy, quit-vs-stay, which offer to take, which strategy to pick) with a verdict that's actually backed by cited, checkable evidence instead of confident-sounding opinion. Three independent lenses — Evidence Skeptic, Steelman Arbiter, Precedent Realist — each build and stress-test the case for every option, blind to each other, then a board verdict names a winner with the evidence that decided it. Use when asked to help decide between options, resolve a dilemma, or settle a debate.
trigger: The user describes two or more options they're torn between (a career move, a technical choice, an offer, a strategic direction) and asks for help deciding, a verdict, or which one is actually right — especially when they want more than a gut-feel answer.
version: 1.0.0
compatible_agents: [claude-code, codex, openclaw]

Dilemma Board — independent review, then cross-examination, then a cited verdict

Most "help me decide" answers are just confident-sounding opinion. This is a three-phase board
process built to force every argument to earn its place with real, checkable evidence — a fact,
a study, a named precedent, a documented outcome — not just "it seems like." Phase 1 builds the
strongest evidence-backed case for each option independently; Phase 2 makes the lenses attack each
other's evidence; Phase 3 delivers one verdict naming the option and the evidence that decided it.

Anti-sycophancy rule (applies throughout)

Do not default to validating whichever option the user seems to prefer. Your job is to find the
strongest real case for every option on the table, then let the evidence — not the user's
framing — decide. Every claim used as evidence must be something checkable: a fact the user gave
you, something you can verify with a tool available to you (web search/fetch), or a named,
identifiable precedent — never an invented statistic, study, or quote. If you cannot verify a claim
and have no tool to check it, label it "unverified" and weight it accordingly rather than presenting
it as settled fact.

Phase 1 — Three independent passes (blind)

Run each lens as an independent pass over the same dilemma and options. Each lens must reach its
own conclusion without seeing the others' output yet.

🔍 Evidence Skeptic

Lens: for each option, list every claim being used to support it (by the user or by common
wisdom) — then rate each claim's evidence quality: verified (checkable fact, cited source, or
tool-confirmed), plausible-but-unverified (reasonable but no real source), or unfounded
(asserted with no basis, i.e. "everyone knows"). Ask "if I stripped out every unfounded claim, what
case is actually left for each option?"

⚖️ Steelman Arbiter

Lens: build the strongest possible case for each option using only claims rated verified or
plausible by the Evidence Skeptic's standard — actively looking for the best argument, not the
user's stated argument. Ask "what would a smart, well-informed person who genuinely preferred this
option say in its defense?"

🌍 Precedent Realist

Lens: has this exact dilemma (or a close structural analog) played out before, for the user or
for anyone identifiable — a real company, a real person, a documented case? Cite the precedent and
its actual outcome, not a hypothetical. If you have web access, search for real cases; if not, draw
only on precedents you can name concretely (not "many people who..."), and say so if none exist. Ask
"what actually happened last time someone faced this choice, and does that transfer here?"

If a lens finds nothing after genuinely looking, note that plainly.

Phase 2 — Cross-examination (the board argues)

Now show each lens the other two lenses' Phase 1 findings (never its own). For every finding from
another lens, each reviewer answers two questions in one line each:

1. Challenge: "Is this evidence actually as strong as claimed, or does it not really apply
   here?" If a reviewer can name a concrete reason a piece of evidence is weaker than presented
   (e.g. the precedent doesn't structurally match, or the "verified" claim was actually only
   plausible), the finding gets challenged.
2. Corroborate: "Does this evidence also strengthen my lens's case for the same option?"
   (e.g. a precedent the Precedent Realist found might also be the exact case the Steelman
   Arbiter's argument needs — corroboration from a second, independent lens is a stronger signal
   than one lens's claim alone.)

Do not let a lens skip this step or rubber-stamp everything as settled.

Phase 3 — Board verdict

Merge everything into one report, organized by option (not by lens). For each option, list its
surviving evidence tagged:

• CONFIRMED — no lens challenged it, or a challenge was raised but not substantiated.
• CORROBORATED — flagged or strengthened independently by more than one lens. Weight these
  most heavily in the verdict.
• DISPUTED — a lens raised a concrete, specific reason to doubt it. Include the evidence and
  the challenge side by side.
• Drop nothing silently — a challenged piece of evidence still gets listed, marked DISPUTED.

End with one Verdict line naming the option the evidence actually supports and why: e.g.
"Option B — 2 CORROBORATED precedents outweigh Option A's mostly-unverified claims" or "Too close
to call — both options rest on similarly-weighted CONFIRMED evidence, the real deciding factor is
[X], which only you can weigh" (say this plainly when it's true rather than forcing a false
confidence).

Output format

## Phase 1 — Independent findings
### 🔍 Evidence Skeptic
- ...
### ⚖️ Steelman Arbiter
- ...
### 🌍 Precedent Realist
- ...

## Phase 2 — Cross-examination
- [Evidence Skeptic #1] challenged by Steelman Arbiter: "..." / corroborated by Precedent Realist: "..."
- ...

## Phase 3 — Board verdict
### Option A
1. **[CORROBORATED]** ...
2. **[DISPUTED]** ... — challenge: "..." (your call)
### Option B
1. **[CONFIRMED]** ...

**Verdict:** ...

Handling sensitive info in the dilemma

If the dilemma involves real names, employers, or financial details, treat that as context only —
do not echo it back beyond what's needed to explain a finding. This Skill sends whatever you paste
as-is to the model; redact anything you'd rather not share before pasting. This is a
decision-support exercise, not professional advice — for high-stakes legal, medical, or financial
dilemmas, get an actual advisor.