Content, not documents: separate the words from the way they look

A contract arrives as a Word file. It has a cover page, a running header that says “Commercial in confidence” on every page, numbering held in Word’s styles, comments in the margin, a table that breaks across two pages and a few hundred kilobytes of metadata nobody has looked at. A person reading it sees past all of that without thinking. A model doesn’t. It has to work through it every time it reads.

A document is a presentation artefact

A document is designed for human eyes. Its fonts, spacing, page breaks and layout exist to help a reader, and they do that well. To a model they are a distraction from the words that matter. The content is what we want to work on. The document is just the way that content has been presented.

So the model never works directly with the document. Code extracts the content once, separated from its presentation, and every later step works from that.

Many reads, one source

Big-document work is rarely one read. A contract review checks each clause against the master agreement, then cross-checks definitions, then looks at the schedules, then checks the result. Each of those passes would otherwise have to navigate the document again, and each might read it slightly differently.

Working from content extracted once is more efficient, because the noise is paid for one time rather than on every pass. It is also more effective, because every check, cross-check and analysis is looking at exactly the same words. When two steps disagree, you know it’s a real disagreement and not a difference in how they read the file.

The same clean content is what goes into the Wordhouse, so it can be reused without anyone going back to the original. The Agentic Wordhouse

Where it meets structure

Content and structure go hand in glove. Structure is the way the content is organised, and the boundaries in the content, where one clause ends and the next begins, are what reveal it. Sometimes we care deeply about that structure, and sometimes we just want the content. Either way, separating the content from the document comes first. Structure first

A side benefit: unmasking hidden instructions

We don’t separate content for security reasons, but it helps. There have been cases of documents carrying hidden instructions aimed at the agent reading them: white text, text in metadata, or a sentence tucked at the end of a very long paragraph, asking the agent to rate or respond favourably. When the content is extracted, that text either drops away or, more usefully, becomes visible. Hidden text stops being hidden, and the adversary is unmasked.

That makes it a simple, testable control against a real risk. It costs nothing extra, because the extraction was already worth doing.

The document is for the reader. The content is for the work.