How my training catalogue stays current without me

Every product has a help centre that is six months behind the product. The feature changed, the screenshot did not, and the article now describes a button that is not there. Everyone knows this and nobody fixes it, because fixing it is nobody's job.

In my business it is several agents' job, and this post is how.

The teaching skill

Most of what I publish for governors is teaching. How the assurance model works. How to use a tool. What to ask at the next meeting. I got tired of writing those by hand and getting them slightly wrong, so I turned the process into a skill that an agent runs, with me as director and editor.

It works through six stages, and the order is the point.

Brief. Who the resource is for, the one outcome it produces, the part of the model it teaches, and the angle. Written down before any drafting.

Ground. Read the canon: the philosophy, the model, and the framework language for what secure and developing look like. The resource teaches our way of doing things, and it has to be anchored to something real.

Source and quality-assure. Find current authoritative sources, Ofsted, the National Governance Association, the Department for Education, the safeguarding guidance, and check that every method the resource asserts holds up against them. Record the citations. Note anything that could not be externally backed. This stage exists because a model that is only asserted is not defensible in front of a board.

Draft. In the house voice, with the brand rules as a hard gate. Not a style guide to keep in mind. A check that fails.

Render and review. Render the piece on desktop and on a phone. Then an independent reviewer agent verifies it with evidence: does each claim about the method have a cited source, and does the page actually look the way the author thinks. It iterates until the reviewer approves. The author does not get a vote.

Humanise and deliver. A final pass to strip anything that reads like a machine wrote it, then a short checklist of what was produced and where.

A side effect I did not plan for: the skill teaches me the topic as it goes. Authoring doubles as learning. I understand the model better for having had to source it.

Help that matches the product

The product help articles had drifted. Flows had been rebuilt and the articles still described the old ones.

So every article was rewritten from the current source code, not from memory. Five tools that had no article got one. And twenty-two one-minute explainer films now play inside the articles, one per tool, so a governor who does not want to read can watch.

The part that keeps it current is not the rewrite. It is the tests. Every film must have its video file and its poster, or the build fails. Every article must render with the real components, or the build fails. When a flow changes, the article that describes it is in the same repository as the code, and the same review that checks the code checks the help.

The films themselves are small web encodes, loaded only when someone presses play, so the help centre is not slow because it is thorough.

One catalogue, generated

The social plan, the help centre and the training hub all need to know what content exists. For a while each had its own list, and the lists disagreed.

Now there is one catalogue, and it is generated by a single command from the site itself. Every training article, every resource, every film, with a human name, the tool it belongs to, its length and format, and where it lives. Add a page and run the command, and the page is in the catalogue. The posting plan picks it up. The help centre can link to it. Nothing has to be remembered.

The catalogue also records what must not be used. Two old promotional videos carry retired names and counting claims. They are in the catalogue, marked do not post, because a thing that is hidden is a thing that comes back.

The freshness sweep

Some content goes stale without anyone touching it. SGAF relies on public education datasets, and those are republished on their own schedule.

A freshness agent checks weekly whether each dataset has a newer version than the one the product uses. If nothing has changed, it says nothing. If something has, I get one line with what slipped and what depends on it. Silence means all is well, and I have learned to trust the silence because the check is deterministic and I have watched it fire.

The brand gate, again

Every piece of copy that goes out passes a deterministic check for the hard rules: no em dashes, no retired names, no counting claims about the product. That gate matters here for a specific reason. Drafted copy must keep passing as the product changes. A post or article written in October that mentions a product count is wrong by February. So drafts are written close to their date, and the gate runs again on the day.

The principle

Current is a property you test, not a promise you make.

A help centre is current if the build fails when it is not. A catalogue is current if it is generated rather than maintained. A dataset is current if something checks and tells you when it is not. A claim is current if it had to pass a gate today, not when it was written.

None of this requires a content team. It requires deciding that "is this still true" is a question a machine should ask, every time, and that a person should only hear about it when the answer is no.