Skip to content
Sabin Mainali

Building a 293-topic help center with a quality pipeline

How I planned a full knowledge base for a cloud phone system and built the writing pipeline that keeps every article accurate, consistent and searchable.

Outcome at a glance

A help center planned as 293 topics and shipped in waves, with more than 190 articles live and a pipeline that checks each one before it goes out.

Context

The product is a cloud phone system for businesses. It covers virtual numbers, calls, SMS, voicemail, IVR menus, call tagging, a power dialer, WhatsApp and shared team inboxes. Every one of those features creates support questions.

Problem

Support was answering the same questions by hand. Help content existed in pieces, written at different times in different styles. Some steps no longer matched the product. There was no plan for what the full help center should cover, and no way to check an article before it went live.

My role

I owned the help center end to end: the topic plan, the article standard, the writing, and the tooling that checks the work.

Approach

I ran the help center like a product, with its own backlog, priorities and release waves.

  • I mapped every feature and every common question into 293 topics.
  • I grouped the topics into build waves, so the most-asked questions shipped first.
  • I wrote house rules once: US English, the serial comma, one task per article, and a fixed escalation block at the end of each article.
  • I made "click-proof" the standard. Every step in an article is checked against the live product before it is published.
  1. PlanPick the next topic from the wave plan and confirm the exact user question.
  2. WriteDraft into a structured JSON store using the house template and voice rules.
  3. CheckValidators and pre-publish checks catch missing sections, style breaks and stray characters.
  4. Verify and shipWalk every step in the live product, mask personal data in screenshots, then publish.
The article pipeline, from topic to published page

What I built

  • A JSON article store. Articles live as structured data, so they can be validated, reused and exported.
  • Validators and pre-publish hooks. Automated checks block an article that breaks the template or the voice rules.
  • Screenshot marking with masking. Screenshots are annotated, and names, numbers and emails are masked before publishing.
  • Search-ready output. Articles answer the question in the first lines, use real headings, and are exported as a full-text file for AI search tools.
  • A custom AI skill. The writing rules are packaged as a Claude skill, so drafts start inside the standard instead of being fixed afterwards.

Outcome

More than 190 articles are live against the 293-topic plan. New articles follow one standard no matter who drafts them. Because every article is checked against the product, support can link to articles with confidence instead of rewriting answers by hand.

What I learned

A help center stays accurate only when checking is part of the pipeline. Style guides get ignored. A validator that blocks publishing does not.