BigDocs: guiding principles

The BigDocs series is about one kind of work: big documents, often several at once, where the business transaction rests on getting the answer exactly right. Contracts against your own paper. Employment agreements against the award. Tenders and security questionnaires against everything you have said before. Help articles against the product.

More of this work now sits with the teams who own it, using tools like Claude without an engineering team behind them. The patterns in this series are how those teams can get answers they can stand behind. Here they are in a page.

Compartmentalise. One big prompt holds every clause against every other clause at once, and the result is a roll of the dice. Give each task only what it needs and nothing that competes with it. BigDocs: rolling the dice on big documents

Keep your words by topic. The Agentic Wordhouse is a dimensional model for your words: one document per topic, with every variant inside it, so each decision has something fixed to point at. Wordhouse

Get the structure right first. Documents are written for human eyes. Let code split them using the structure they already carry, prove nothing was dropped, and have a person confirm it. Structure first

Write the job down before it runs. The instruction file is a fully mapped task plan, approved by a person: every part, what it is checked against, and what it must return. The instruction file

Never retype. Point at existing text and let code copy it. Express edits as exact operations. If the model can’t retype it, it can’t mistype it. Never retype

Put people where judgement is needed. Show every decision and its basis, and set aside what the system doesn’t know. Fewer touches, each one with more effect. Output built for oversight

Keep the Wordhouse true. Maintain it with the same patterns, in reverse, and account for every section of every source. Maintaining the Wordhouse

Build the tests in. Check completeness inside the work, let it loop until the checks pass, and keep a test set that becomes your monitoring. Build the tests in

Steer with some strictness, for now. Models are improving quickly, and some of this will become unnecessary. Today, on work that matters, agents still do best with a confirmed structure, a library to point at, small clean tasks and tests that don’t take their word for it.

Where these came from

These patterns are mostly borrowed: from data warehousing, from software engineering, and from years of making automated decisions defensible to the people they affect. I’ve spent a long time on both sides of that, in the boardroom signing off on results and in the codebase building what produces them. The benchmark is my attempt to show they hold up when measured. BigDocs: the benchmark

If your team has a document job that comes back from AI looking right and turns out not to be, bring it to a BigDocs session.