Guardrailed AI for technical editing

When each author edits with their own chatbot, you get inconsistent style, untracked changes and text leaving the company. Start small: use an offline LLM for language checks only, track every change in Git or Word, and keep a human responsible for the technical content.

I've had a couple of recent interview processes where the companies are keen to start using AI in their editing workflows. Both of these processes highlighted that although word is coming down to use AI -- they are unsure how to proceed.

The core problem is the same. Many engineers writing many documents, terminology and style differ, content overlaps and contradicts from one document to the next. It's a waste of engineers' time to clean it up, but someone has to do it.

Teams are already using AI, possibly not in a coordinated way. When several authors each draft with a different chatbot, the result is a mix of styles that no one chose (same as the original problem). My job does not get easier as the problems are just supercharged. Proprietary text may also be leaving the company without anyone noticing. In regulated and industrial settings, that is not acceptable, and raw LLM output can't be trusted with facts, numbers or meaning anyway.

Start small

I changed the approach during a recent contract in a way that gave us tangible and controlled results using an offline LLM. The goal was to use AI to do the initial legwork when receiving new documents -- check grammar, terminology, and natural rewording. The changes were recorded either via Git commits or MS Word tracked changes and were therefore controlled and reversible. Using a version-controlled workflow we could ensure consistency in the editing voice.

My first role then was to check and confirm these machine-led changes. Feeding back into the workflow scripts where editing errors or misses occurred. What would have previously taken me hours was now done in less than one with (now) shocking accuracy.

My primary role stayed the same -- an experienced human eye looking for inconsistencies in technical content and formatting. Numbers, specifications, procedures and structure are off limits to the AI, and anything uncertain is flagged. The AI is never the only control, and every document is brought to the same standard whatever tool it started from.

The job is done quicker and with higher quality. I can use a machine to do the heavy lifting where I just need to accept, reject, or rewrite chunks of content. I can concentrate on making sure the content is technically correct and compliant.

Technical details

I want the document clean and consistent with the company style and terminology before starting to read and understand it. Otherwise, instead of trying to find logic errors, I'm stuck on rewording strange sounding sentences or looking up terms.

To do this, I used Ollama and the local LLM phi4:14b. I passed the document either as plain-text (markdown / RST) or as an MS Word file and applied the prompts to edit it. SOURCE.

Who's this for?

This approach suits growing companies with many authors and no style guide yet, and established companies whose MS Word workflows have become a bottleneck. For Word users, files come back with tracked changes, so nothing about your tools or process has to change.

You also don't need a full-time hire. Work can be project-based, such as a batch before an audit or a style guide and termbase build, periodic every 6 to 12 months to catch drift, or a retainer for steady volume. The style guide and termbase carry over, and they leave your content better prepared for whatever you do with AI next.

Where next?

Where this goes next: once language is consistent, the next step is technical consistency. When a document changes, the wider suite can be checked for sections that may be affected, and flagged for a person to confirm. Flag only, never edit. I see this as the natural second stage of the same service.