Trying the IDENTITY.md / CONTEXT.md pattern in a system with real retrieval guardrails

Been experimenting with something adjacent to the root-file orientation pattern from the ICM template, but in a system that works differently than a local folder setup.

No folder hierarchy to lean on. Instead, everything routes through an orchestrator that gates every request through read-only guardrails before it touches any retrieval source. Multiple separate source types — filings, news, an internal knowledge base, a couple of specialized lookups — none of them organized in any kind of tree structure. Just distinct categories the router has to choose between intelligently. Write actions are completely disabled on a separate blocked path, so this is read-and-cite only, nothing gets modified.

The question I was testing: could I still apply that root-file orientation thinking here, even without folders to organize?

Turned out yes, but the implementation looks different. Instead of a file that sorts first and orients the AI at session start, it’s more about writing precise routing instructions so the orchestrator knows which source category to hit for which kind of question, and in what order to check them. Same underlying idea — give the system a map before it starts retrieving — just expressed through routing logic instead of a root file.

Confirmed something for me: the folder structure was never really the magic. The instructions about how to navigate what’s available is the magic. Folders just make that easier when you have them. When you don’t, you have to be more deliberate about writing that logic explicitly somewhere else.

Anyone else working in a system without folder hierarchy? Curious how you’ve approached this

7 comments

Trying the IDENTITY.md / CONTEXT.md pattern in a system with real retrieval guardrails