Claude Code that invents more and talks above the user's level is a settings job (--install-claude: pin opus 4.6 + analogize). Character count is not the metric. Do not paste "be concise" into CLAUDE.md. Before orchestration, critic the execution document (the plan and any child prompts) with read-only Fable. The session agent does not switch. The consult costs tokens and is not ground truth.
English (main). Korean: branch ko.
npx skills add dalsoop/orchestrator-consultant-gate -g -y
python3 scripts/resolve-consult.py --install-claude --json
python3 scripts/resolve-consult.py --list --json
python3 scripts/resolve-consult.py --exec-spawn --briefing templates/fable-briefing.md
python3 scripts/resolve-consult.py --print-spawn
python3 scripts/resolve-consult.py --report --json
python3 scripts/resolve-consult.py --json
python3 eval/run.py
EVAL_LIVE=1 python3 eval/run.py # Cursor/Claude spawn; Grok uses claude -p --max-turns 1. Cursor may skip on workspace trust.npx copies files. A second opinion needs the host to spawn a child. That is not E2E from install.
Korean: npx skills add dalsoop/orchestrator-consultant-gate@ko -g -y
One AI writes the execution document from the work order. GATE before orchestration and before audit-then-change. Fable rebuts it read-only: one more missed bottleneck, a hole in a child prompt, an omitted category. Notes go back. Same session. The payoff is before dispatch, in that document, not at merge. Claims are probabilistic. Digest them. Do not treat them as answers.
Claude Code mistakes and harder language: --install-claude (opus 4.6 + analogize, keep-coding-instructions). --install-hook is PreToolUse. Pick the critic from --list. Default is Fable (agent-model-registry get fable). Spawn the slug from --json. If Fable is blocked, pick opus 4.6 from the list, not grok. --report is usage history, not critic quality. Grok: --exec-spawn --briefing after --list; Grok has no PreToolUse. Default eval is static; EVAL_LIVE=1 only when asked. Hosts: Cursor, Claude Code, Codex, Grok TUI (GROK_AGENT=1). The session agent fills templates/ at the gate.
SKILL.md · templates/ · scripts/ · eval/ · MIT
Score a SKILL.md: npx skills add dalsoop/skill-audit -g -y
Vendored in host-skills-mono as orchestrator-consultant-gate (deploy: personal-mac). Public install is still npx skills add. Optional: agent-model-registry set fable <id>.