Proof checking against a fixed checklist

For a healthcare agency’s medical writers: every finished document checked against the same written rules, every time.

Healthcare

AI skill

Pilot

An illustrated leaflet for an invented medicine with numbered lines and eight points highlighted. Margin notes give the line, a High, Medium or Low rating and the issue, with a rule code on all but the Low point, beside a rules checklist.

Client

A healthcare agency

My role

Design and build

Timeline

2 Weeks

Year

2026

A healthcare agency asked me to help their team work out where AI could help. I interviewed the team about what they expected and what worried them, turned that into a plan, and proposed a pilot with the medical writers.

The challenge

The job. Before sign-off, the medical writers read every finished PDF against the client’s house style: the spelling, punctuation and terms every document has to follow. It’s careful, repetitive work, and a good fit for AI.

What went wrong first. The first version gave the AI the style documents and asked it to work out the rules itself, on every run. Each time it read them, it came up with a different set of rules. The same PDF could get a different check each time.

The approach

The fix. I had an AI agent turn the style documents into a numbered checklist, one rule per line, and checked it myself against the originals. The checklist is loaded in full on every run, never searched, so no rule can quietly drop out.

The proof checker is a skill: a set of instructions and tools the AI loads for this one job. It works through every rule in turn, names the rule behind every point it raises, and re-reads each flag against the text it quotes. It finishes with a table showing every rule and what it found. Points it isn’t sure about go in a separate section. It suggests no corrections. A medical writer decides on every point and makes every edit.

I designed and built the skill and its Lambda tools. It runs on the platform my co-founder built.

The outcome

Where it stands. The agency’s team is using it in a pilot. It checks house style only, not claims or evidence. It reads finished, text-based PDFs one at a time, and it can’t see text a PDF doesn’t expose, such as lettering converted to outlines in a design file. A clean report means it went through every rule, not that the document is right. When early notes described it as checking claims, we corrected them, and its known gaps are tracked in writing.

What’s next. Today, when the house style changes, I update the checklist by hand with an AI agent. The next version will let the team keep the house style in their own document, with each change turned into checklist rules and reviewed by a person before it’s used.

An illustrated passage of house style guidance written for humans beside a checklist for agents that turns its key phrases into coded rules, one per line.

An illustrated proof check report with totals, one judgment note, and numbered items rated High or Medium that quote highlighted text, with an audit table of rules and results layered over it.

An illustrated flow. A finished PDF, a rules checklist and approved terms feed an AI step, which leads to a Word report and then a medical writer who decides every point.

Next project

A knowledge base built for regulated work

The Vaultia Knowledge Base in dark mode: the document library in front, with a document’s status panel and the review and expiry counts layered over it.

Proof checking against a fixed checklist

The challenge

The approach

The outcome

If you’re working out where AI fits in your team, I’d like to hear about it.