checking-for-undisclosed-context

Checking for Undisclosed Context

Overview

Text written by someone deep in a project accumulates vocabulary, acronyms, and structural assumptions from whatever they've been reading and discussing - a design doc, a ticket thread, an earlier stage of the same pipeline. None of that is visible to a reader who wasn't there. The author can't tell the difference between "this term is obvious" and "I only find this obvious because I wrote the doc it came from" without deliberately stepping outside their own head.

This skill is that step: read the text as a specific named reader, not as its author, and find the places where understanding it secretly depends on a document or conversation that reader was never given.

There's a second version of this specific to text that passes through multiple model turns, drafts, or agent hops: a model can coin a term on the spot to name something in the moment - a workaround, a state, a recurring shape - and that term gets carried into later output simply because it appeared once, not because anyone deliberately chose it as shared vocabulary. A few hops later, the turn that coined it is gone, and the term now reads as established shorthand to whoever encounters it next, even though nothing ever defined it and it may not exist anywhere else at all. This is the same leak as human-authored undisclosed context, but the "document the reader never saw" is a vanished conversation turn instead of a doc that at least still exists somewhere.

Different axis from writing-for-scannability (can the reader find the structure) and writing-voice (does it sound human). This is about access: could this specific reader actually follow it, or does it quietly assume a tab they don't have open.

When to use

• Reviewing a design doc, code-review prompt, PR/issue description, or subagent persona/prompt that's meant to stand alone
• Someone asks for a "cold read," or to check whether a doc is self-contained
• Writing something for a reader who explicitly hasn't seen the predecessor material - a follow-up proposal that shouldn't require the first one, a skill extracted out of a larger multi-stage protocol, a persona dispatched with a fresh context window
• Authoring a subagent prompt that has to work without inheriting the dispatching session's context
• Reviewing content that passed through multiple drafts, turns, or agent hops, where a term coined once earlier may have been carried forward as if it were established vocabulary