メニュー
Round-robin multi-LLM consultation tuned for InTheWake PWA development. GPT proposes, Gemini refines, Grok challenges — full-context debate with wheat/chaff justifications, low-hanging fruit, and PWA checklist.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | orchestra |
| description | Round-robin multi-LLM consultation tuned for InTheWake PWA development. GPT proposes, Gemini refines, Grok challenges — full-context debate with wheat/chaff justifications, low-hanging fruit, and PWA checklist. |
| version | 1.0.0 |
Ask the orchestra. Let them debate. Build what survives.
cd /home/user/ken/orchestrator && python3 orchestra.py cruising "task description"
Round-robin debate where each model sees EVERYTHING from prior rounds — proposals, verdicts, justifications, rejections. Nothing filtered. What one model calls chaff, the next might rescue.
Round 1: GPT — Proposes structure, approach, innovations. Each proposal has CONFIDENCE + JUSTIFICATION.
Round 2: Gemini — Sees GPT's full response. For each proposal: WHEAT | CHAFF | WHEAT_WITH_REFINEMENT (with justification). Adds own proposals. Identifies low-hanging fruit.
Round 3: Grok — Sees both GPT + Gemini. Verdicts on EVERY prior proposal. Specifically rescues dismissed ideas worth saving. Adds own proposals. Identifies blind spots + low-hanging fruit.
Blind Spot Check — Grok reviews the emerging synthesis: "What are we all still missing?"
Claude Synthesizes — Full debate chain → attributed plan → encode to memory.
Each model provides justifications for wheat/chaff verdicts. The next model can disagree with the reasoning, not just the conclusion. An idea rejected in Round 2 can be rescued in Round 3 because the justification was visible.
After every orchestra run, verify these PWA concerns:
| Concern | Check | File |
|---|---|---|
| Service Worker | Does sw.js need updating for new pages/assets? | sw.js |
| Cache Strategy | Static assets: cache-first. Dynamic data: network-first. | sw.js |
| Manifest | App name, icons, theme color, start URL current? | manifest.json, manifest.webmanifest |
| Offline | New pages fallback to offline.html? | offline.html |
| Precache | New files added to precache manifest? | precache-manifest.json |
| Prefetch Images | New images in prefetch list? | prefetch-images.json |
| Performance | First load <200KB? LCP <2.5s? | Lighthouse |
| Sitemap | New pages listed? | sitemap.xml |
| AI Discovery | llms.txt current with new content? | llms.txt |
| Schema | POI/port data validates against schema? | schema/*.json |
Typical: $0.03–$0.08 per full orchestra run. Report shows per-round breakdown.
Soli Deo Gloria — Let the orchestra play. Build what serves the traveler.
Use when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions always
The dedication under which all work in this household is done — every task offered to God, excellence as worship. Automatic posture skill, present and byte-identical in every repository and sibling workspace. Preserve the invocation; getting it right is the worship, not getting it fast.
Hard ban on any reference to the external domain getbets.us on cruisinginthewake.com. Enforced via PreToolUse hook (.claude/hooks/no-getbets-guard.js) and git pre-commit hook (.githooks/pre-commit). Fires automatically on every Write/Edit/MultiEdit and every commit; no manual invocation.
Voice & Presence — fires during cruise content writing. Guards the sound of the prose so it reads as written by someone who has actually sailed the route, not by a brochure or a model. Shapes voice markers, rhythm, vocabulary, and pastoral honesty. For post-draft diagnostics, see voice-audit.
Post-draft diagnostic for InTheWake content. Scans for cruise-marketing tells, AI authorship-cluster signals, and absence-of-authenticity gaps. Assesses authenticity risk by clustering signals across layers (Layer 1 strong, Layer 2 supporting, Layer 3 counter-signals). Required: every voyage-pack ships with a voice-audit attestation in its .factcheck.json sidecar. Fires before committing content edits or before publishing a new page. For during-writing standards, see like-a-human. For corpus measurement, see voice-dna.
External red-team critique pass. Sends a commit range or proposed change to a second model (defaulting to a different LLM) with the explicit charter to find every flaw "as though you're trying to take the engineer's job." Grok-native usage notes.