Cleanup
Cleanup finds the pages in your docs that describe things you removed or never built. It checks the pages you pick against today's code and the full history of the repos you pick, and tells you which pages to keep and which to probably remove, with the proof. It never changes a page. Cleanup has its own screen in the sidebar, under Generate docs.
Who can use it?
In early access: turned on per team. Write to hello@amendary.com to try it. Once it is on, Cleanup has its own entry in the sidebar, and the Pages tab links to it with “Find pages to clean up”.
Which pages can a cleanup check?
- Notion: pages you shared with Amendary, watched or not. Ticking a page picks it and every page under it. “Leave out only this page” drops one page and keeps the pages under it; “Only this page” picks a page without them. A page that only lists its subpages is skipped and costs nothing.
- GitHub: Markdown files in a docs folder or wiki, watched or not.
- Zendesk and Intercom: the articles you watch.
Translations are skipped. A cleanup reads the source-language article only, because a translation follows its source. Product-update entries are skipped too.
What does a cleanup do?
- From the Cleanup screen, start a new cleanup and pick pages from any destination (next question). A cleanup checks up to 50 pages.
- You pick the repos to check them against. Every connected repo is ticked by default.
- You see the estimate in credits, with a line per repo, before anything runs.
- Amendary reads each page for the names it uses, then searches each repo in turn: the code as it is today and every commit in its history.
- Each page gets an outcome, Keep or Probably remove, with a one-sentence reason, a confidence (high, medium or low), the proof links for Probably remove, and 1 to 5 short guidelines on what to do with it.
- Every cleanup is listed on the Cleanup screen. Open one to see its report: Probably remove first, then Keep. Filter it by To do, Done or All, and by destination (Notion, GitHub, Zendesk or Intercom).
What does each outcome mean?
- Probably remove. What the page is about is gone from the code, or was never built. The card links the proof: the commit that removed it, or its absence from the product code history of every repo checked. Code checks that the proof is real.
- Keep. The page still describes the product. Without a real proof of removal, a page stays Keep. A dated record (changelog, decision record, incident review) and a page about process, policy or people are Keep too.
What do the guidelines say?
For a kept page: which statements to fix and what the code has instead, what to merge or split, what is missing, and what could not be verified.
For a page to probably remove: what to check before you delete it, and where the content moved if it moved to another repo.
Does an absence mean it was never built?
Not always. An absence means the name never showed up in the product code history of the repos checked. If code was squashed into one commit, imported from another repo, or lives in a repo you did not tick, it can be real and still not found. That is why a page to probably remove comes with guidelines on what to check before you delete it.
Why check several repos?
A page describes the product, not one repo. If a name is gone from your API repo but lives in your web app, the page is not wrong. So a page is Probably remove only when what it is about is gone from every repo checked. When the content moved to another repo, a guideline says where.
A repo that could not be searched (too large, or could not be copied) blocks Probably remove for every page, and a name left unsearched when time ran out cannot prove an absence. The report says why.
Why are some pages Keep and Not fully verified?
A name that reads as code can prove something is gone: a name in backticks, a file path, a setting key or a URL. So can the name of another product you connect to, like “Slack”, but only when the history shows the commit that added it and the commit that removed it, or when no file of any repo ever mentioned it. A feature name or a button label, like “CSV export”, only shows that something is still there, because the code may call the feature something else. When too little evidence covers a page, it stays Keep, marked Not fully verified, and the reason says so. That is an honest answer, not a failure.
Which pages are not checked?
These are not part of the report. The run’s status line counts them: pages that could not be read, pages that failed on our side, and contents pages that only list their subpages. A contents page is skipped and never charged. A run paused at the daily usage limit says when it resumes.
What does it never do?
Cleanup never edits, archives, deletes or maps a page or article on its own, in any destination. It only reports. You decide what happens to each page.
From a page card you can Mark done, with what you did (for Probably remove: deleted it, or kept it because the verdict was wrong; for Keep: fixed it, or nothing to fix), or select Watch (with several repos, then Watch with the one it describes). That goes through the normal mapping step and your plan’s limits, like watching from the Pages tab. A help-center article maps to the repo its help center follows, and an article of a help center that follows a changelog URL cannot be mapped this way. A GitHub file maps only to the repo it lives in. Watching a page hands it to the daily checks.
What does it cost?
A cleanup spends credits, like generating docs from scratch. Before it starts you see an estimate: a line for the pages and a line for each repo. Credits are held when you start and the part a cleanup does not use is refunded when it ends. Searching the history happens on our side and costs no credits. See Billing & plans for credit packs.
A cleanup that reaches your team’s daily usage limit pauses and picks up after the daily reset. A restart on our side does not lose it either: it resumes where it stopped. The team owner or the person who started it can pause, resume or cancel it; a cancel refunds the unspent credits. Watching a page afterwards, and the daily checks that follow, cost no credits.
What is stored, and can I share the report?
Each repo is copied to a temporary folder on our server, one at a time, for the length of its search, and deleted right after. The report keeps references (repo, commit, file path, tag), not copies of your code. Each page’s reason and guidelines, which can quote the page and a value from the code, are stored encrypted.
Export gives what the report’s filters show, so a teammate without an Amendary seat can act on it. The CSV has one row per page with its outcome, confidence, reason, guidelines and proof links. The Markdown file lists Probably remove first, then Keep.
Try Amendary on your own help center
Connect Notion, GitHub, Zendesk or Intercom and pick the articles to watch. Your first check shows what Amendary would fix, with the change as proof.
Free for 14 days with no card, and every edit waits for your review. Or free forever with your own Anthropic API key. Code is never touched.