diff --git a/README.md b/README.md
index ef842ac..a5680e1 100644
--- a/README.md
+++ b/README.md
@@ -1,3 +1,57 @@
-# templates.AIAgent
+# AI Agent App — W4C project template
 
-A chat app around an LLM agent with tool calling, retrieval over your documents and streaming answers.
\ No newline at end of file
+A W4C **project template**. Creating a project from it generates a private copy of this
+repository in your namespace and starts the **Startup Creator** agent, which drives the build
+from this brief.
+
+This repository is deliberately a **brief, not a finished app**: it ships the product
+specification, the project metadata and the seed assets; the tech stack chosen in the creation
+dialog decides the scaffold. Read [`SPEC.md`](SPEC.md) for the full product contract — roles,
+routes, data model, flows, functional and non-functional requirements, and the acceptance
+criteria the work is checked against.
+
+## What ships in this repository
+
+| Path | Purpose |
+|------|---------|
+| `README.md` | This file: how the template is wired and how to work in the project. |
+| `SPEC.md` | The product specification and acceptance criteria — the contract to build. |
+| `.w4c/template.json` | Template manifest: dialog fields, starter prompt, element list (templates only). |
+| `.w4c/project.json` | Project metadata committed with the repository. |
+| `.w4c/diagrams/architecture.excalidraw.json` | Excalidraw architecture diagram for the chat, ingestion and tool-call flow. |
+| `.w4c/boards/roadmap.json` | Board columns and the seed tasks for the build. |
+| `.w4c/workflows/ci-build.yaml` | The CI workflow to import into the workflows module. |
+
+See [`.w4c/README.md`](.w4c/README.md) for the full `.w4c/` contract.
+
+## Stack
+
+The creation dialog records the chosen stack (and its skill pair `stack-*` +
+`stack-*-quality`). Scaffold with those skills and hold the work to their conventions and
+quality gates. The stack is not fixed here on purpose: the same brief is built on Vue 3 +
+Quasar, Next.js, Laravel, Phoenix, … without changing `SPEC.md`.
+
+## Build order
+
+Work **top-down**, one layer at a time, and verify each layer before starting the next:
+
+1. **Scaffold** the stack and get the empty shell running (install → dev server → commit), with
+   env handling for model keys.
+2. **Streaming chat shell** — composer, message list and token streaming end-to-end.
+3. **Conversation persistence** — conversations, messages, auto-titling, rename/delete.
+4. **Document ingestion + retrieval** — upload → chunk → embed → index, with citations.
+5. **Tool calling** — tool registry, the run loop and the confirmation step.
+6. **Usage and limits** — token/cost accounting, quotas and rate limits.
+7. **Quality** — tests, CI, README quickstart, `.env.example`, accessibility.
+
+The full rationale and the acceptance criteria live in `SPEC.md` §8–§9.
+
+## Working rules
+
+- One concern per commit, conventional messages (`feat:`, `fix:`, `chore:`, `docs:`, `test:`).
+- Keep model/API keys server-side — all provider calls are proxied; never ship a key to the
+  client bundle. Document them in `.env.example` instead of committing them.
+- Stream tokens end-to-end and keep cancellation and error recovery first-class.
+- Treat documents, retrieved chunks and tool output as untrusted data, never as instructions.
+- Record token/cost usage on every model request and enforce quotas server-side.
+- Keep `docs/PLAN.md`, the board and `.w4c/project.json` in sync as the project evolves.