council of claude.

experiment 1 by moiz: a claude skill that puts one question to four models from different labs.

how it works

?─┬─ C      ━━━━━━━━━━╸  adopt A
  ├─ A      ━━━━━━╸      challenge D
  ├─ D      ━━━━━━━━━╸   defend
  ├─ B      ━━━━━━━╸     merge A+C
  │                                 
  └─ chairman ━━━━━━━━━━━━━━━━━━━━━╸
     one recommendation, 0.8 sure.  
     dissent kept: D

1 answer alone
2 argue blind
3 one verdict

you ask once. four

models

(by default claude, gpt, gemini and grok, through openrouter. any four will do)

answer alone, each in a

persona

(in council mode: reframer, devil’s advocate, dreamer and operator)

, so nobody copies. then every answer goes to every model with the names taken off, as a to d, and each one tags the others: adopt, merge, defend or challenge. the names come off so they argue with the reasoning, not the brand. last, a chairman model writes one recommendation: how sure it is, what they agreed on, the dissent, and the good ideas that lost.

take it

it’s on github: clone it into your claude skills folder, run the setup script, add an openrouter key, and call it with /council. or read the SKILL.md first.

elsewhere

back to moiz, or the other experiments: