A
ATÖLYE wiki
ATÖLYE wiki / Strategy

Decision: replacing Notion with a Git-backed wiki

Owner Aatish Patel Reviewed 4 days ago Next review 2027-09-17 Current

Decided 17 September 2026. Records are superseded, not edited — if this changes, write a new page and mark this one status: superseded with a link.

What was actually wrong

The ATÖLYE Playbook, read directly from Notion on 17 September 2026:

Pages in the Playbook wiki database 707
Owned by someone no longer in the workspace 413 (58%)
Not edited in over 12 months 436 (62%)
Edited in the last 90 days 18 (2.5%)
Pages ever marked verified ~50
Of those, expired 39
Pages carrying no tags 689 of 707

Two findings mattered more than the rest.

The owner field was full and meaningless. Not one page had an empty Owner, which looks like perfect coverage until you check the names. 413 of those people have left. Strip out the departed and the 188 owned by one person, and fewer than 110 pages were owned by a current colleague other than Aatish. A field nobody prunes at offboarding is worse than no field, because it looks answered.

Verification was attempted once and abandoned. Someone ran a real cleanup in late November 2025 — the whole Strategy branch was verified on a 90-day cycle. It expired as a block on 24 February 2026 and nobody renewed any of it. The tool guides went the same way between April and July 2026. Of the eleven verifications still valid, four were monthly Comms Reports, which never needed verifying.

Notion was not the problem. Nobody owned the recurring act of checking. Any replacement inherits that failure unless the checking is automated and addressed to a named person.

What we considered

Outline, self-hosted. The strongest wiki option. Google Workspace SSO already exists here, which removes its main setup obstacle, and its API is mature enough to build the review job against. Rejected not on merit but on fit: it is a real-time collaboration engine, and at a 2.5% quarterly edit rate we do not have a collaboration problem. It also means owning Postgres, Redis, backups and monthly migrations forever — a second recurring maintenance job, when an unowned recurring job is the thing that put us here.

Docmost. Genuinely open source and lighter, but the same shape of commitment with a less mature API.

Existing Cloudways servers. Managed PHP stack, no Docker, no root. Their newer Node product deploys from Git with framework presets rather than a prebuilt image, and does not appear to offer Redis. Not a fit.

Cloudflare Workers. V8 isolates, not a Node process. Cannot host a long-running server holding websockets. Cloudflare still earns a place in the stack as the front door.

Keeping Notion and fixing the process. Honest option, and cheaper than any migration. Set aside because the renewal decision was already in motion, but worth recording: most of the diagnosis above would have been fixed by one person owning the verification sweep.

What we chose

Markdown files in a Git repository, built to a static site by a small Node script, deployed to Cloudflare Pages on the free tier.

The metadata model is the one Notion already had — owner, review date, cycle, status — moved into YAML front-matter where it can be enforced:

Each of those is a few lines of configuration. The combination is the thing Notion, Outline and Docmost all lack.

Where everything else goes

Not one destination. Exporting 707 pages into one new wiki would recreate the problem with different software.

Content Destination
Policies, strategy, tools, Ai HQ, delivery process This wiki
Newsletters, Comms Reports, Community_Archived, IST handbook Read-only Google Drive folder
Per-project artefacts, project cards, closure checklists Productive
People directory (currently two overlapping databases) Leapsome, as one list
Notion AI meeting notes Nowhere. Not migrated.

The Drive bucket matters more than it looks — it is the pressure valve that lets people say "do not migrate" without feeling they are deleting history.

What we knowingly gave up

The assumption this rests on

Five or six editors, everyone else reads. If that number reaches roughly twenty, the pull-request workflow becomes a bottleneck and Outline becomes the right answer after all. Revisit this page if the editor count changes materially, not on a schedule.

There is also a reading of the 2.5% edit rate we should stay honest about: it might mean the content is genuinely stable, or it might mean nobody was maintaining it and should have been. If this system works, that number should go up. If it goes up a lot, Git will chafe.

Sequence

  1. Reassign owners in Notion for all 413 orphaned pages, or decide to drop them. Cheapest filter available, and it shrinks everything downstream.
  2. Export and import per branch, starting with Policies — smallest, highest consequence, clearest owner.
  3. Run both systems until Policies is trusted here.
  4. Notion read-only for one quarter, then closed. Not indefinitely: an archive nobody closes is exactly how Community_Archived happened.
Edit this page at content/strategy/decision-replacing-notion.md. Change the content and bump reviewed in the same commit — that is the whole review process.