All three hyperscalers now publish AI reference architectures faster than any architect can memorize them. This playbook loads the vendor architecture-center corpus into a retrieval tool, shortlists the references that actually fit your constraints, and adapts one — with every inherited design decision cited back to its source page.
Tier: Core · Hands-on time: ~25 minutes
Retrieval beats memory at 2026 publication pace. AWS shipped Well-Architected lenses for generative and agentic AI, Azure publishes baseline chat and landing-zone reference architectures, and Google maintains a family of RAG reference architectures spanning managed to fully custom. An architect who "remembers" the right pattern is recalling last year's guidance; one who retrieves it cites this quarter's. The skill this playbook trains is adaptation with provenance: knowing which parts of a reference are load-bearing guidance, which are vendor defaults, and which simply do not survive contact with your constraints.
Skip AI when you already know the exact reference document — read the primary source directly; retrieval adds a lossy middleman to a known destination.
Build the notebook for this decision, not forever. Load the corpus and title it for the engagement question — "Meridian: agent architecture references, 2026-Q3". Freshness is the point: note the retrieval date on each source. Vendor guidance revises quietly, and a notebook built last spring answers with last spring's canon.
Retrieve against constraints, not keywords.
From the loaded documents only: I am designing a customer-support
copilot with an order-status agent for a US retailer. Constraints:
client is committed to one hyperscaler; agent must call an on-prem
order API through private connectivity; retrieval corpus contains
PII under state privacy laws; client ops team is small (managed
services preferred).
1. List every reference architecture in the corpus relevant to this
shape. For each: source document + section, what it covers, and
which of my constraints it addresses or violates.
2. Recommend the 2 best-fit references and say specifically why the
others fit worse.
3. Answer only from the documents. If the corpus does not cover
something, say NOT IN CORPUS — do not fill gaps from general
knowledge.
Read the shortlisted primary sources. Not negotiable, and not long — you are reading two documents, not twenty, which is exactly the leverage retrieval bought you. The tool found the right pages; only reading tells you the caveats, the "this baseline is not production-ready without X" warnings, and the architecture's actual reasoning.
Adapt with a delta table. Ask the tool to structure the adaptation: "For reference architecture [chosen], produce a table: component in the reference, its role, keep/replace/drop for my constraints, and justification. Flag every row where my constraint forces a departure from the reference." The delta table is the artifact: it shows reviewers you started from canon and departed deliberately, with reasons — the strongest position an architect can take into an ARB.
Separate guidance from defaults. Walk the keep rows and ask of each: is this a considered recommendation (the reference argues for it) or a demonstration default (it merely appears in the sample)? References embed illustrative choices — instance families, example regions, convenience services — that carry no authority. Inheriting a default as if it were guidance is this playbook's characteristic failure.
Cite forward into the design. Every element your design inherits gets a citation to the reference (document, section, retrieval date) in the HLD or ADR; every departure gets its delta-table justification. Six months later, when the vendor revises the reference, the citations tell you exactly which parts of your design to re-examine.
The corpus is public vendor documentation — the sensitive side is your constraint set and design context in the prompts. Under a restrictive pb-03 ruling, redact the queries (workload shape without client identity: "US retailer" not the client name), or run retrieval inside the approved boundary: a Claude Project on enterprise tenancy, a client-subscription endpoint, or local RAG (Ollama + the downloaded PDFs) for fully closed environments. The reading and delta-table steps are AI-optional; the primary sources are free to read anywhere.
The delta table — "started from the vendor reference, departed here, because of this constraint" — is the most persuasive design document in the ARB room. Retrieval buys you the reference; your departures are where the architecture actually lives.