The Architect

Keep the architecture ladder human-owned. Let the agent draft and diagram, but judgment and sign-off stay with the accountable SME.

This skill is for decisions that persist — the ones that are expensive to reverse, that constrain the system for months, or that affect multiple teams. Implementation details belong in code review. Architecture decisions belong here.

When to use

• Choosing between two or more structural approaches (database type, service boundary, communication pattern)
• Deciding how to handle a cross-cutting concern (auth, observability, multi-tenancy, data residency)
• Designing an integration boundary between systems
• Anything that warrants an ADR

Procedure

1. State the decision to be made — one sentence. If you cannot state it in one sentence, it is not one decision.

2. Capture the context — what forces, constraints, and requirements make this decision necessary? What is the status quo, and why is it insufficient?

3. Enumerate the options — at least two credible options. For each:

   • What are the benefits?
   • What are the costs and trade-offs?
   • What does it require from the surrounding system?
   • What is the failure mode?

4. Run the Adversarial Gate on the leading option — before committing, invoke adversarial-gate on the preferred choice. A clean gate result is a prerequisite for committing to the decision.

5. State the decision — which option is chosen and why. The reasoning must reference the context and trade-offs from step 2-3, not just assert a preference.

6. Record consequences — what does this decision make easier? What does it make harder? What does it prevent?

7. Author the ADR — format:

   ---
   status: draft
   id: ADR-NNNN
   date: <date>
   deciders: [<handle>]
   ---
   
   # ADR-NNNN: <short title>
   
   ## Status
   draft | approved | superseded
   
   ## Context
   <forces, constraints, why a decision is needed>
   
   ## Decision
   <the chosen option and reasoning>
   
   ## Consequences
   <what becomes easier, harder, or foreclosed>
   

   A decision's lifecycle is visible in both the filename postfix and the status: field: ADR/ADR-NNNN-<slug>-DRAFT.md (status: draft) while under review, renamed to ADR/ADR-NNNN-<slug>-approved.md (status: approved) on approval.

8. Link from the relevant spec, HLD, or PR — ADRs that are not referenced are not found.

Cloud vendor routing

When a decision touches a specific cloud, invoke the matching expert skill before finalising the ADR, and record which guardrail checklist ran:

Vendor	Skill
Google Cloud, AWS, Azure, Alibaba Cloud	cloud-expert (read the matching references/<vendor>.md)

Multi-cloud decisions run every relevant checklist; the ADR names them. A cloud decision without its guardrail checklist is an unreviewed decision.

Outputs

• ADR/ADR-NNNN-<slug>-DRAFT.md → ADR/ADR-NNNN-<slug>-approved.md on approval
• Trade-off summary (can be included in the ADR or a separate HLD section)
• Updated architecture/HLD-<feature>.md if the decision changes the HLD

Guardrails

• One ADR per decision. Bundling multiple decisions into one ADR obscures accountability.
• The ADR captures reasoning, not just the outcome. "We chose Postgres" is not an ADR. "We chose Postgres because our consistency requirements rule out eventual-consistency stores, and our team has no operational experience with Cassandra" is an ADR.
• Architecture is not implementation. An ADR about which ORM to use is implementation. An ADR about whether the service owns its own database is architecture.
• Supersede, do not edit. When a decision changes, create a new ADR that supersedes the old one. History is the point.