Your transport protocol is Slack
Somewhere between your first coffee and your first meeting, it arrives: “quick one — which auth pattern are we using for internal services?”
You answered that on Tuesday. Different asker. You will answer it again on Thursday. Alongside it, the rest of the daily rotation: is Zustand blessed or are we still on Redux? What’s the retry convention for queue consumers? Where does the tenant ID come from in service-to-service calls? Every one of them is answered in a document you wrote. Every one of them gets asked anyway.
Then the last two years happened, and the questions acquired a second audience. Every developer on your team now runs coding agents, and every agent’s picture of “how we build here” comes from whatever its developer put in a rules file — which is to say, from a private, slowly rotting copy of answers you gave at various points over the last year. When the agent gets it wrong, the developer asks you. When the developer doesn’t ask, the wrong version ships, and you meet it in code review instead.
Nobody changed your title. But you did acquire a new job: you are the organization’s context layer. Requests arrive over DM. You resolve them from memory and documents you wrote. The response closes the connection. Nothing is cached. The same request arrives tomorrow from someone else.
You are the MCP server.
You just have terrible uptime, no caching, and you don’t scale horizontally.
When you are in a meeting, the context layer is degraded. When you take leave, it is down. When you eventually change jobs, the organization loses its context layer entirely and books the outage as “attrition.”
Why the wiki didn’t fix this
You did the responsible thing years ago. You wrote the golden path down. The asking continued — and it is worth being precise about why, because it is not laziness.
A document is a pull system, addressed to someone who might come looking. A question is a push. For the asker, interrupting you costs seconds and returns a current, authoritative, situation-specific answer. Searching costs minutes and returns a page of unknown freshness that they must interpret themselves. Asking dominates, rationally, and it will keep dominating for as long as you keep answering. You are competing with your own docs and winning, which is the worst prize in engineering.
Agents shifted this calculus in exactly one direction: an agent cannot DM you mid-session. So instead it falls back on whatever stale context its developer configured, ships its best guess, and converts your interrupt load into review findings. The question volume didn’t drop. It moved downstream, where it costs more.
The interrupt ledger
One week, zero tooling. This is the diagnostic that turns a vague sense of “I answer a lot of questions” into a number you can act on.
Tally
For one working week, count every interruption that is a question about something already decided — standards, patterns, blessed libraries, conventions, how-we-do-X-here. Tally marks in a text file are enough. Do not count novel design questions; those are your actual job, and they are not the problem.
Classify
At the end of the week, sort the tally into three columns: (a) answered in a document that exists; (b) decided at some point, but never written down anywhere; (c) genuinely new. Most people who run this are unsettled by how small column (c) is.
The fork audit
Collect the agent rules files — CLAUDE.md, Cursor rules, whatever your tools read — from three teammates. Diff them against each other and against the golden path. Count the contradictions. Every one of them is a future interrupt, or a future review finding, with your name on it.
Columns (a) and (b) are your servable load — the portion of your interruptions that exists only because the golden path is not machine-readable. Column (a) is a distribution failure: the answer exists but doesn’t reach the moment of need. Column (b) is a capture failure: the decision happened but was never recorded. Neither is a you-failure. Both are infrastructure gaps being papered over by your calendar.
Run a server instead of being one
The fix is not another wiki reorganization, and it is not answering faster. It is moving the golden path into what we call the document plane: a versioned layer that whoever — and whatever — is writing code reads at the moment of writing. At your altitude, that means four specific changes:
Distribution stops being you. Standards, checks, project rules, and specs are served over MCP, so every teammate’s agent pulls the same current, authoritative answer at write time. The blessed answer arrives without you in the loop. That is column (a), served by an actual server.
Capture stops being optional. The same connection runs backwards: decisions get recorded as a by-product of the work, from inside the workflow. Column (b) stops accumulating, because recording is no longer a separate act that loses to delivery pressure.
Trust stops decaying. The reason people ask you instead of the doc is that you are provably current and the doc isn’t. A layer that continuously compares its own content against the repositories — and flags the drift — earns the one thing a wiki never had: a freshness guarantee worth redirecting the asking to.
Your artifacts start executing. Project rules and prompt templates live in the same layer, so "how we do X here" ships to agents as standing, structured guidance rather than prose someone must remember to paste.
This is what we are building with SystemDox: the golden path git-first with GitHub as the source of truth, served over MCP to Claude Code, Cursor, and any MCP-compatible tool, with a published knowledge base for the humans and structured context for the agents — one layer, both audiences.
Your job description shifts by one word: from answering to curating. You maintain the source once. The protocol does the distribution, at write time, to every session, including the ones that start while you are asleep.
What this does not fix
Interrupts will not go to zero, and they shouldn’t. Column (c) — the genuinely novel — remains yours, and it is the part of the job worth having. Adoption is real work too: a context server nobody’s agent is connected to serves nobody. The setup is a configuration block per tool, but getting a team to add it is a social act, not a technical one — budget for that. And the content is still yours. This layer distributes judgment; it does not produce it. If the golden path is wrong, it is now efficiently wrong everywhere — with, at least, much better feedback loops for noticing.
Start here
Run the interrupt ledger Monday to Friday. Tally, classify, fork-audit. You will end the week knowing your servable load as a number.
Then take your top five answers — the five you gave most often — and put them into one instrumented repository as standards and project rules. Connect one team’s agents. Watch next month’s tally. SystemDox’s free tier covers one repository, permanently — one repo and one team is exactly the right size for the experiment.
The golden path was never the problem. The problem was that you were its only server.