orchestrate

name: orchestrate description: "Full multi-LLM pipeline orchestration. Runs the cruising pipeline that coordinates GPT, Gemini, and Grok as consultants while Claude remains lead developer. Auto-detects cruising mode from this repository."

Orchestrate — Multi-LLM Pipeline

Claude leads. External models consult. Claude decides what survives.

Usage

/orchestrate "task description"
/orchestrate cruising "Build a new ship page for Norwegian Prima"

Mode is cruising by default in this repository. You can override by specifying the mode explicitly.

How It Works

The orchestrator runs a multi-step pipeline defined in /home/user/ken/orchestrator/modes/cruising.yaml:

1. Read Standards (Claude) — Load ITW-Lite v3.010 standards, hero + compass pattern, accessibility, canonical URLs
2. Generate (Claude) — Generate standards-compliant page content with full codebase context
3. Content Review (GPT) — Suggest content additions, structural alternatives, SEO improvements
4. Completeness Check (Gemini) — Check for missing sections, travel data, port information
5. UX Challenge (Grok) — Challenge layout assumptions, suggest UX innovations
6. Integrate (Claude) — Integrate feedback, enforce compliance, produce final output

External steps (3-5) are optional — the pipeline continues gracefully if any fail.

Backend Invocation

IMPORTANT: Execute these commands directly using the Bash tool. Do NOT check if files exist first — just run them.

bash /home/user/ken/orchestrator/bootstrap-env.sh 2>/dev/null; pip3 install -q -r /home/user/ken/orchestrator/requirements.txt 2>/dev/null && python3 /home/user/ken/orchestrator/orchestrate.py cruising "task description"

Only if the command fails with No such file or directory or ModuleNotFoundError, tell the user:

"The orchestrator backend isn't available. Make sure the ken repo is cloned to /home/user/ken/ and run pip3 install -r /home/user/ken/orchestrator/requirements.txt."

Integration with Claude Code

After the orchestrator returns its JSON output:

1. Read the consultations — Each external model's feedback is in consultations[]
2. Check claims — Review unverified_claims and failed_claims before trusting anything
3. Claude integrates — Use the feedback to inform your work, but Claude makes all final decisions
4. Enforce standards — All output must comply with ITW-Lite v3.010: hero + compass, right-side rail, accessibility, canonical URLs

Context Boundaries

SEND to external models

• Page requirements and structure
• Content topics and outlines
• SEO targets
• General cruise/travel data

NEVER SEND to external models

• Full codebase or source files
• Internal standards documents (send summaries only)
• Analytics data

Constraints

• One hero + one compass per page
• Right-side rail required
• Accessibility script included
• Canonical URLs on all pages
• Versioning system followed
• All ITW-Lite v3.010 standards enforced