← ClaudeAtlas

councillisted

Fan out a question to Claude and GPT-5.6-sol in parallel, run up to 3 rounds of adversarial debate where each model challenges the other's weakest argument and updates its position, assess convergence after each round with a dedicated subagent, then synthesize and deliver a final answer weighted by argument quality — not consensus.
steven112163/dotharness · ★ 0 · AI & Automation · score 65
Install: claude install-skill steven112163/dotharness
# Council Coordinates an adversarial two-model debate: Claude and GPT-5.6-sol generate independent positions, challenge each other directly for up to 3 rounds, and converge on the strongest answer by argument quality — not agreement. ## When to use - User asks a hard technical or factual question where a second perspective reduces error - User says "council", "ask GPT", "get a second opinion", "debate this" - On your own initiative: question is high-stakes, has conflicting priors, or your confidence is low ## When NOT to use - Simple factual lookups (just answer directly) - Creative or subjective tasks (debate adds noise, not signal) - Question already well-resolved in context ## Protocol ### Phase 1 — Independent answers (parallel, round 0) Create a temp dir under the repo's `tmp/` (gitignored, not in `/tmp`): ```bash _repo_root=$(git rev-parse --show-toplevel 2>/dev/null || pwd) mkdir -p "$_repo_root/tmp" COUNCIL_DIR=$(mktemp -d "$_repo_root/tmp/council-XXXXXX") mkdir -p "$COUNCIL_DIR/round-0" ``` Write the question to a file first (avoids shell injection and handles multiline questions): ```bash printf '%s' "<question>" > "$COUNCIL_DIR/question.txt" ``` **Write your own independent answer to `$COUNCIL_DIR/round-0/claude.txt` first.** Form and commit your position before spawning the GPT subagent — this is the anti-anchoring guarantee. Then spawn a `general-purpose` subagent to get GPT's response. Provide the literal expanded path values in the prompt (subagent