Skip to content
Article Every Oct 2026

Every: why adding more context to your AI writing setup makes it worse

What the article is about

Katie Parrott, writing in Every’s Working Overtime newsletter, documents a common failure pattern in AI writing setups: the context file that starts as a useful style guide and gradually becomes a liability. Her case is specific — a 759-word guide that grew to 3,855 words over several months — but the dynamics she describes are recognizable to anyone who has spent time tuning an AI writing workflow.

How the problem develops

The growth itself is not the issue. Each addition felt justified: a rule to fix a recurring mistake, a clarification after an ambiguous output, a new example to reinforce a preference. The problem emerged in how the model responded to accumulated context. It began treating drafts and logs stored in the same directory as additional instructions rather than as reference material. Old examples shaped new outputs in ways that were hard to trace. The model appeared to be generalizing from the volume of material rather than the intent behind it.

The result was output that technically followed the rules but had lost the voice the guide was trying to preserve. More instructions did not produce better compliance — they produced a different kind of miscalibration.

What she recommends

Parrott’s core advice is to subtract before adding. When output quality drops and the instinct is to add another rule, the more productive intervention is often to look at what is already there and remove what is ambiguous or outdated.

Her specific recommendations: archive the existing context and rebuild from scratch using only the principles that actually mattered; configure the agent to ask before modifying files rather than acting on assumed intent; write style guides as principles rather than templates, because templates constrain structure while principles guide judgment; and treat the context file as something that needs maintenance, not just accumulation.

The rebuild she describes ended up shorter than the original 759-word guide. The model’s output improved not because it had more to work with, but because what it had was clearer.

Who it is useful for

Writers and editors who have been using AI tools for several months and have noticed a gradual drift in output quality without a clear cause. It is also useful for anyone building AI writing workflows for a team, since the same dynamics appear when multiple people contribute to a shared context file over time.