Continuous onboarding is the idea that a fast-moving tool has to teach its users continuously, inside the work itself, because any separately-maintained documentation goes stale faster than users can re-read it. The term comes from the Inline Cues essay on this site.
Why onboarding stopped being a one-time event
A developer used to read a tool's docs once, internalise the mental model, and trust it for months. The essay argues modern agentic tools break that assumption through two compounding sources of variance: model variance (the model behind the tool shifts week to week as providers update system prompts, alignment, or pricing) and harness variance (the tool ships features faster than docs can describe them, partly because the tools now build themselves). The user's mental model goes stale faster than they can update it, leaving them juggling official docs, community forums, and their own testing, each going out of date at its own rate.
The fix: put the teaching surface inside the work
A worked example from the shipped Claude Code pack: the user types undo in a prompt; the word dims; navigating to it surfaces a tip in the status line - use /rewind to roll back context and file changes together - and cycling offers /rewind as a drop-in replacement for the word. The user has now discovered functionality they did not know existed, while using the tool, without leaving the prompt. When the feature changes, the cue updates with it; when a new feature ships, a new cue ships alongside; when behaviour is deprecated, the cue points at the current alternative.
Docs are the cues
The sharpest formulation in the essay: there is no documentation to go stale, because there is no separate documentation. The docs are the cues, and the cues live in the prompt. That's continuous onboarding as a property of the interface rather than a content-team obligation, and OpenCues' cue system is its delivery mechanism.
Related: What is a cue? · What is inline AI? · What is a HAII? · What is a kata? · GitHub repo