Contextkeeping: a toolkit built the way it tells you to build

KCS was built on two assumptions. Humans write the articles, and humans read them. In the last three years both quietly died.

AI can draft an article from a resolved case in seconds, so the capture tax that sank a thousand KCS programs is nearly gone. And the reader of your knowledge base is, more and more often, a machine assembling an answer for someone who will never see the article page. What survives is the part everyone treated as vegetables: governance. In this new world it’s the load-bearing wall.

I wanted to package what I’ve learned running knowledge programs into something a team could pick up on a Monday and use, without hiring me, without buying a platform, and without pretending the AI era hasn’t happened. The toolkit’s own field guide puts it in six words: keep the soul, replace the machinery.

It’s KCS-informed, not KCS-branded, and it runs on whatever a team already owns: Zendesk, Salesforce Knowledge, Notion, plain git, or any other tooling you may have. Underneath is one operating model, the Governed Knowledge Loop.

That last idea has a name in the toolkit, the Precision Paradox: the more fluent AI answers get, the more each remaining error costs, because nothing in a polished answer signals doubt.

The toolkit lives in a git repository as Markdown and YAML, because a knowledge toolkit that wasn’t itself governed knowledge would be a very expensive joke. A build script turns the source into branded PDFs and per-product zip files, and a validator stands at the gate. It checks every document against the metadata schema, sweeps for leftover placeholders and stale branding, verifies the cross-references between modules, refuses a workbook whose formulas haven’t been recalculated, scans for terminology adherence, and then re-reads every built PDF to make sure the disclaimer and branding actually made it onto the page.

The first day, August 9, did most of the heavy lifting: the scaffold and the field guide, the curation gate kit, the governance pack, the measurement workbook, the prompt library, the assessment, and the rollout playbook, then the brand, the PDF pipeline, and the store copy. By evening it had a name, Contextkeeping, and a version number, 1.0. Claude did the typing. I did the arguing, and the commit log shows where I won: a reordered module list, renamed loops, a refund policy, and terminology corrected more than once.

Version 1.3 is built and validated: a field guide and nine modules, fifty branded PDFs, and every working file behind them. That means article templates with gold-standard examples, eight AI prompts (four of them for drafting) with an evaluation protocol that catches silent regressions when models change, three gate checklists, governance policies ready to ratify, a sixteen-tab measurement workbook built on cost per answer, a ninety-day rollout playbook, a terminology pack, and the content graduation model. All built from my own learning, words, and guidance. The store listings are up and delivering their files.

My favourite piece is the smallest. One Issue, End to End follows a single support ticket through every module: a resolved case, a drafted article, a gate decision, an update when the world changes, and the economics of what one governed article is worth. Total human time across the issue’s entire life is about seven minutes. If you only read one document in the box, read that one.

The lesson of building it is the one I’d want buyers to take away. A toolkit that tells you to govern your knowledge had better be governed knowledge itself: in version control, validated before every release, every claim traced, and nothing shipped that a gate hasn’t seen. Otherwise it’s just another binder on a shelf, and those have never saved a support organization.

So, a question for anyone who owns a knowledge base: who is its next reader? If your answer isn’t “a machine,” look at your logs again.

Transparency note: drafted with Claude from the toolkit repository, its changelog, and its field guide, then read, cut, edited heavily, rewritten, and finally approved by me. As always: AI drafts, humans ship.