Skip to content

multi-tool-orchestrator-setup.md

Latest commit

 

History

History
176 lines (125 loc) · 7.37 KB

multi-tool-orchestrator-setup.md

File metadata and controls

176 lines (125 loc) · 7.37 KB
Raw

Multi-tool orchestrator setup

How to run the SuperApp five-specialist team across Claude, ChatGPT, Gemini, NotebookLM, and Cursor. You stay in the orchestrator and architect seat; each tool handles a focused lane.

Full team contracts: team/agent-index.md · Paste team/ADAPTER-PROMPT.md into your primary Orchestrator (Claude) on day one.

Tool × role map

SuperApp role Primary tool What it does
Orchestrator Claude Routes work, owns team/STATE.md and team/HANDOFF.md, session plans
Architect Claude Blueprints, data contracts, file trees, STOP gates
Assistant Claude Day-to-day follow-ups, clarifications, doc polish
Document designer Claude Specs, handbook pages, README/runbooks, blueprint prose
Engineer Cursor Code, tests, migrations, git in the repo
Assistant 2 ChatGPT Shorter coding passes — sanity checks, bugs, “does this diff look right?”
Researcher Gemini or NotebookLM Web search, source grounding, competitive scans, digest for Orchestrator
Creative Claude, ChatGPT, or Gemini Graphics, UI direction, copy, tokens — use whichever is strongest that day
Quality ChatGPT + Cursor Pre-merge review; ChatGPT for fast review, Cursor for repo-aware fixes

Graphics and design are intentionally shared: all three LLMs can draft layouts, copy, and visual direction. Pick one per task; send outputs to Orchestrator for a single HANDOFF to Cursor.

Process stages

1. Intake      → You + Claude (Orchestrator)
2. Research    → Gemini / NotebookLM (Researcher) → brief back to Orchestrator
3. Blueprint   → Claude (Architect) → §DATA CONTRACT if DB changes
4. Implement   → Cursor (Engineer) → db:validate → tests
5. Quick check → ChatGPT (Assistant 2) → short bug/review pass
6. Design      → Claude / ChatGPT / Gemini (Creative) → optional UI/copy pass
7. Quality     → ChatGPT or Quality specialist → acceptance vs blueprint
8. Ship        → HANDOFF → staging checklist → gated production

Stage 1 — Intake (you + Claude Orchestrator)

  1. Open Claude (Projects recommended for long context).
  2. Paste team/ADAPTER-PROMPT.md.
  3. Ask: “Where are we?” — Claude reads team/STATE.md and team/HANDOFF.md.
  4. Describe the feature or fix in plain language; Orchestrator assigns a wave.

Exit: A one-paragraph HANDOFF or blueprint outline exists.

Stage 2 — Research (Gemini or NotebookLM)

Use when you need external context (competitors, APIs, pricing, docs).

Gemini

  1. New chat with a focused question and URLs if you have them.
  2. Ask for a digest: bullets, sources, risks, open questions.
  3. Copy the digest into Claude Orchestrator — do not ask Cursor to “research the web” unless you have browse tools enabled.

NotebookLM

  1. Create a notebook; upload PDFs, links, or notes for the topic.
  2. Ask for summaries, comparisons, or “what contradicts X?”
  3. Export the summary to Orchestrator for blueprint input.

Exit: Research digest attached to the Orchestrator thread or saved under team/Researcher/ (optional).

Stage 3 — Blueprint (Claude Architect)

  1. Orchestrator (or a dedicated Claude thread) uses docs/agents/templates/blueprint-template.md.
  2. Include §EXECUTION CONSTRAINTS and §DATA CONTRACT when touching the database.
  3. Save the plan under docs/plans/ in your app repo (not required in this curriculum repo).

Exit: Engineer has a numbered, unambiguous implementation list.

Stage 4 — Implement (Cursor Engineer)

  1. Install Cursor; open your project folder (projects/<your-app>/ or copied template/).

  2. Point Cursor at team/Engineer/AGENTS.md and the blueprint section.

  3. Run from project root:

    pnpm run db:validate
pnpm dev

  4. Engineer implements one blueprint section at a time; you commit on a feature branch.

Exit: Tests pass locally; diff matches blueprint scope.

Stage 5 — Quick check (ChatGPT Assistant 2)

Short session — keep context small.

  1. Paste the diff summary or error message (not the whole repo).
  2. Ask: “What did I miss? Any obvious bugs or security issues?”
  3. Send actionable bullets back to Cursor or to Orchestrator if scope changed.

Exit: No blocking issues, or a tiny fix list for Cursor.

Stage 6 — Design (Creative — any of three)

  1. Choose Claude, ChatGPT, or Gemini for graphics, layout, or copy.
  2. Provide: audience, brand adjectives, screenshot or wireframe if you have one.
  3. Creative output goes to Orchestrator; Orchestrator writes HANDOFF for Engineer (token names, copy strings, Tailwind direction — not vague “make it pretty”).

Exit: Implementable design notes in HANDOFF (hex/Tailwind, component names, copy).

Stage 7 — Quality (pre-merge)

  1. Open a PR on [github].
  2. ChatGPT: quick pass on description + risk areas.
  3. Cursor or Orchestrator: run pnpm test, Playwright if configured, db:validate.
  4. Update team/STATE.md — Orchestrator only.

Exit: PR ready for human merge; staging deploy only with --confirm scripts.

Stage 8 — Ship (gated)

  1. Follow docs/00-start-here.md Track 3.
  2. Use playbooks under playbooks/hosting/ (GCP, AWS, or Oracle).
  3. Run deploy-staging.sh --confirm before any production gate.

Exit: Staging verified; production only after explicit owner sign-off.

How to add each tool (first-time setup)

Tool Setup steps
Claude Account + Project; paste ADAPTER-PROMPT.md; pin team/STATE.md uploads or repo access
ChatGPT Account; separate chat per “Assistant 2” / Quality pass; avoid pasting secrets
Gemini Google account; use for research chats; link sources in digest
NotebookLM Google account; create notebook per product/research topic
Cursor Install from referral link; open app repo; enable project rules from .cursorrules / CURSOR.md

After init wizard:

pnpm run init

Your project copy includes .superapp/ai/ with tool-specific starter files when you select them in the wizard.

Daily loop (minimum viable)

Claude (plan/HANDOFF) → Cursor (build) → ChatGPT (quick check) → PR → staging gate

Add Gemini/NotebookLM when the task needs external research. Add Creative when the task needs UX/copy/visuals.

Build your own models (coming soon)

We'll Do It Solutions is preparing a public track on developing and training your own models (when to fine-tune, when to use hosted APIs, data prep, and evaluation). Watch welldoit.solutions and this repo for announcements.

Until then, use the hosted stack above — it is enough to ship production apps without running your own training cluster.

Related