You are the entry point. Your only job is to pick which path's SKILL.md to hand off to. Do not scaffold anything yourself.

Step 0 — Read these references

• build-paths/README.md — what each path is, in plain English.
• build-paths/skills/README.md — the building-block skills the higher tiers compose.
• build-paths/PLAN.md — only if the user asks "what's the architecture?"

You do not need to load the per-path references at this stage. Each path skill loads its own.

Step 1 — Ask one question

Print:

build-paths — pick a difficulty:

  1. beginner       — RAG chatbot scaffolded around Oracle 26ai + langchain-oracledb,
                       three "X-to-chat" flavors (PDFs / Markdown notes / Web pages),
                       Open WebUI frontend, Grok 4 via OCI Generative AI.
                       (~1 afternoon · 3 ideas to pick from)

  2. intermediate   — Tool-calling agent over a live Oracle schema via
                       oracle-database-mcp-server, with embeddings happening
                       *inside* the database (registered ONNX model, no external
                       embedding API). Three flavors — NL2SQL data explorer,
                       schema-doc generator+Q&A, hybrid retrieval (vector + SQL).
                       Open WebUI + Grok 4.
                       (~1-2 days · 3 ideas to pick from)

  3. advanced       — Agent system where Oracle is the *only* state store,
                       composed from the build-paths/skills/ building blocks
                       (oracle-aidb-docker-setup + langchain-oracledb-helper +
                       oracle-mcp-server-helper). Three flavors — production
                       NL2SQL+RAG hybrid analyst, self-improving research agent,
                       conversational schema designer.
                       (~3-5 days · 3 ideas to pick from)

Which one?

Wait for a number. If the user describes their experience instead of picking a number ("I've built RAG before"), suggest the matching path and ask them to confirm.

Step 2 — Dispatch

Pass through target_dir and topic if the user supplied them.

After handoff, the path skill takes over completely. Don't second-guess it.

Step 3 — One-time prerequisites check (before handoff)

The new (post-restructure) skill set requires:

1. Docker — verify docker --version works.
2. Python 3.11+ — verify python --version.
3. OCI Generative AI API key — OCI_GENAI_API_KEY (a sk-... bearer token, generated in the OCI GenAI service console). All three tiers require this; auth is bearer-token over OpenAI-compatible HTTP, NOT SigV1, so the user does NOT need ~/.oci/config or a compartment OCID. Default endpoint is https://inference.generativeai.us-phoenix-1.oci.oraclecloud.com. If the user has no API key, ask them to generate one and stop.
4. OCI_GENAI_API_KEY must NEVER be committed — .env is in .gitignore. Confirm before writing any project files that reference it.

These prerequisites apply to every tier. Don't re-ask them inside each path.

Stop conditions

• User wants something not on this list ("can I do this in TypeScript?", "I want a Postgres backend", "what about Mongo?"). Print: "v1 of build-paths is Python + Oracle only. See PLAN.md for the rationale and the v2 backlog." Stop.
• User can't decide between two paths after 2 nudges. Default to intermediate and tell them so — it's the most representative shape.
• User has no OCI Generative AI API key (OCI_GENAI_API_KEY, a sk-... value from the OCI GenAI service console). Stop. Point them at the OCI GenAI console to generate one (no full OCI tenancy needed), or at archive/ for Ollama-flavored older ideas.

What you must NOT do

• Don't try to scaffold from this file. Hand off.
• Don't load all three path skills "to compare." Read README.md once; pick.
• Don't invent a fourth path.
• Don't suggest Ollama as a fallback. The post-restructure tiers are OCI-GenAI-only on purpose.