About the service
A help centre only earns its keep when it answers the questions people actually ask, so I start with the support inbox and not with a page tree. I read several months of conversations, group them by the job the person was trying to finish, and count how often each group comes back. That gives an honest map: which topics deserve a full walkthrough, which are a two-sentence answer, and which are not documentation problems at all but rough edges in the product that no article will paper over.
The structure follows those jobs rather than your internal org chart — a visitor looks for what they are doing, not for the team that owns the feature. Search sits on top of it: suggestions as you type, tolerance for typos and for a wrong keyboard layout, and synonyms for the words customers use instead of yours. Every article follows one template — a short line on what the reader will get, numbered steps, screenshots where a sentence is not enough, related links, and a closing "this did not help" block that hands the reader to support with the article context already attached.
The part most help centres skip is measurement. I set up a report of searches that returned nothing useful and of articles people open and leave within seconds; that list becomes the writing queue for the next quarter instead of guesswork. Staff get editing rights scoped to their area with a simple review step, answers buried in old correspondence are moved into proper articles, and a revision is scheduled once a quarter so the centre does not quietly drift away from the product it describes.
What the work includes:
— reading and grouping support requests to find the real top of the list;
— a section structure built around user tasks, named the way customers speak;
— search with suggestions, typo tolerance and a synonym dictionary;
— a single article template: steps, screenshots, related links, a "did not help" block leading into support;
— a report on unanswered queries and weak articles, with a rewriting queue;
— editing rights for employees and a lightweight review flow;
— migration of old answers from correspondence into articles;
— a quarterly revision with a checklist and a named owner.
Ground rules I keep to without exception: an article does not go live until it has been checked step by step against the working product; examples never carry real keys, contract numbers or customer data — everything is replaced with obviously invented values; visitor queries are collected anonymously, and the content of a search is never tied to a particular person.
The structure follows those jobs rather than your internal org chart — a visitor looks for what they are doing, not for the team that owns the feature. Search sits on top of it: suggestions as you type, tolerance for typos and for a wrong keyboard layout, and synonyms for the words customers use instead of yours. Every article follows one template — a short line on what the reader will get, numbered steps, screenshots where a sentence is not enough, related links, and a closing "this did not help" block that hands the reader to support with the article context already attached.
The part most help centres skip is measurement. I set up a report of searches that returned nothing useful and of articles people open and leave within seconds; that list becomes the writing queue for the next quarter instead of guesswork. Staff get editing rights scoped to their area with a simple review step, answers buried in old correspondence are moved into proper articles, and a revision is scheduled once a quarter so the centre does not quietly drift away from the product it describes.
What the work includes:
— reading and grouping support requests to find the real top of the list;
— a section structure built around user tasks, named the way customers speak;
— search with suggestions, typo tolerance and a synonym dictionary;
— a single article template: steps, screenshots, related links, a "did not help" block leading into support;
— a report on unanswered queries and weak articles, with a rewriting queue;
— editing rights for employees and a lightweight review flow;
— migration of old answers from correspondence into articles;
— a quarterly revision with a checklist and a named owner.
Ground rules I keep to without exception: an article does not go live until it has been checked step by step against the working product; examples never carry real keys, contract numbers or customer data — everything is replaced with obviously invented values; visitor queries are collected anonymously, and the content of a search is never tied to a particular person.
