A Claude skill that convenes a panel of expert advisors on a hard decision — and tells you when the panel wasn't worth convening.
Free. Runs on your own Claude. Nothing leaves your chat.
Most multi-advisor tools have the same unstated problem: they can't tell you when they didn't help. The research on multi-agent deliberation is genuinely deflationary — for a lot of decisions, five advisors arrive exactly where one good one already was, at five times the cost. Every tool in this category quietly hopes you won't notice.
Decision Panels measures it. On every run it:
- Takes a solo baseline — one strong single-pass answer — before the panel is convened,
- Hides that baseline from the panel so the comparison means something,
- Probes cheaply before spending a panel — three maximally-distant lenses give a one-line read each; if they agree on both the call and the crux, it tells you a panel is unlikely to change anything and stops there (unless the decision is hard to reverse, where it always convenes),
- Otherwise convenes 5–7 advisors — seated by who actually disagreed — hears them independently, keeps the disagreement,
- Has them cross-examine each other blind — answers stripped of authorship and relabelled A–E, so arguments are judged on merit rather than on which advisor said them,
- And then reports whether the panel actually beat the baseline — including the runs where it didn't, in those words: "The panel did not beat the single answer here. You could have asked once."
That last step is the product. A panel that can't tell you when it was unnecessary isn't one you can trust when it says it was.
Claude Code
git clone https://github.com/chorusse/decision-panels ~/.claude/skills/decision-panelsOr drop the decision-panels/ folder into ~/.claude/skills/. Restart Claude Code; it will pick the
skill up automatically.
Claude Desktop / claude.ai — add the contents of SKILL.md as a Project instruction or custom
skill, depending on what your plan exposes.
Just describe the decision:
- "Should I raise prices or add a cheaper tier?"
- "Help me decide whether to hire or contract this out."
- "I'm torn between these two positioning angles."
- "Pressure-test this decision for me."
It will ask you 2–3 calibrating questions first — the ones whose answers would actually change the recommendation. That step is not filler; a panel run on guessed inputs produces a guessed answer dressed up as rigor. Answer them in a sentence each.
You get back a decision brief: the recommendation with a stated confidence, the panel's real disagreement, what the blind review round caught — including the thing every advisor missed — the strongest case against the recommendation, what would change the call, and the baseline comparison.
| File | What it is |
|---|---|
SKILL.md |
The method. 10 steps, in order — including the one where it declines to run. This is the actual asset. |
references/panel-lenses.md |
The advisor-lens library to draw panels from |
references/decision-brief-template.md |
The output format — built to be forwarded |
examples/example-pricing-decision.md |
A worked run end to end |
- We publish our own comparison. See
proof/head-to-head.mdfor Decision Panels vs. LLM Council (1,271★) vs. vanilla Claude — including a section on where the competitor is genuinely better than us, and a pre-registered protocol for the empirical results we have not yet produced. - This is a text file. There is no magic in it and no lock-in. Read the method, take what's useful, fork it. The value is the sequencing, and the sequencing is right there for you to check.
Decision Panels, by Chorusse — many expert voices in, the disagreement kept honest, one clear call out.
Built by Daniel A. Bauer · linkedin.com/in/danbauer