TL;DR: A recurring pattern for turning structured content (Markdown, CMS entries) into finished, branded documents (Google Docs, Word) using an AI coding agent plus a dedicated export script. Four components, one common failure mode, and a decision point on whether it's worth building at all.
The problem, stated precisely
Content gets authored in a clean format. It has to leave as a formatted document with a cover page, consistent typography, and tables that don't fall apart. The step in between is usually manual, and usually falls on whoever wrote the content.
This is not the same problem as "intelligent document processing" (IDP), which is about extracting structured data out of existing documents (think OCR on invoices). This is the reverse direction: generating a document from content that's already structured. Worth separating clearly, because the tools and failure modes don't overlap.
The four-part pattern
structured content → agent (workflow definition) → formatting rules (encoded in editor context) → reusable content blocks → export script (native API) → branded document
Agent with a defined workflow, not a single prompt.
The agent isn't told "format this document" from scratch each time. It's given a workflow: what to validate in the source, what to assemble, when to pause for human review, what to flag for manual cleanup. This is what makes the setup reusable across document types instead of a one-off script.Formatting rules that live where the work happens.
Heading-to-style mappings, table structure conventions, which Markdown patterns export cleanly: encoded as context the agent picks up automatically, not documentation someone has to read first. New team member, new agent session, same correct output.Reusable content blocks.
Recurring sections (disclaimers, methodology, standard intros) are parametrized templates, not copy-paste targets. Consistency stops depending on who's writing that week.A dedicated export script against the native API.
Not a generic converter. Working directly against the target platform's API (Google Docs, Word's OOXML) is what gives full control over branding, and it's also where most of the interesting bugs live (see below).
Why not just a script
A script handles the happy path fine. It falls over on anything that doesn't match the expected shape: an unusual table, a heading pattern nobody anticipated, a section that needs a human call rather than an automatic export. An agent, working from intent rather than a fixed rule set, can reason through those cases instead of failing silently.
The trade-off: agent output needs the same review discipline as any automated pipeline. Correctness and speed are separate properties. Don't assume one from the other.
Where this actually breaks
Document APIs weren't built for this use case, and their documented behavior doesn't always match reality once you're in the weeds. A few patterns worth watching for if you're building something similar: inserting content in one place can shift positioning of everything downstream of it, and global style changes (spacing, margins) can silently affect elements you wouldn't think of as "text," like headers and footers. None of this shows up in the docs. It shows up when you actually run the pipeline against real content.
Because of this, a realistic pipeline still budgets for a short human pass before anything ships. The win isn't zero manual work, it's manual work measured in minutes of polish instead of hours of rebuilding.
Is it worth building
Quick gut check: recurring document type, produced weekly or more, structured and version-controlled source content, specific branding requirements a generic tool can't hit? Probably worth it, pays back within a few weeks once you count revisions. One-off documents, source content scattered across files with no consistent shape? Not worth the engineering time. Use an existing tool.
Curious whether others have hit the same API quirks building something similar (Google Docs, Notion API, Confluence, wherever). What broke for you that the documentation didn't warn about?
Top comments (0)