When do you stop using ChatGPT for architecture and actually draw boxes?
I love brainstorming with a model, but at some point I need a diagram the whole team will actually maintain.
Where do you draw the line between exploration and commitment?
15 replies
Once two teams depend on the same interface, it goes into Mermaid in the repo — everything before that stays in scratch notes.
We timebox exploration to two sessions; after that, someone owns a short RFC with explicit non-goals.
LLMs are great for enumerating options; humans still decide trade-offs the model cannot weigh without business context.
Our rule: no AI-generated diagram becomes canonical until a staff engineer signs the file in git history.
Boxes help when onboarding — narrative helps when aligning execs — we maintain both but link them from one index page.
I caught a subtle ordering bug because the model drew a synchronous sequence that our runtime actually executes async.
Sequence diagrams aged fastest; C4 context stayed useful — we invest more in the latter now.
Rubber ducking with a model is cheaper than booking a whiteboard room, but it does not replace stakeholder buy-in.
We archive exploratory chats as markdown in the ticket so future readers know why we rejected the fancy approach.
The model suggested microservices where a modular monolith would have been cheaper — glad we sanity-checked with SRE.
Architecture is partly social: who owns on-call for each box matters more than the prettiness of the drawing.
For data flows involving PII, legal wants a human-approved diagram — AI drafts do not count as documentation there.
We embed lightweight ADRs next to the code instead of giant slides nobody opens after launch week.
Good prompts include constraints — 'assume single region SQL Server' changed the recommendations completely.
Line between exploration and commitment is when money moves — contract signature means frozen diagrams for that release.
Join the conversation.
Log in to reply