Repairing a weak wiki is an information architecture migration, not a cleanup pass. Inventory everything, classify each page by intent, choose one canonical URL per important concept, build semantic hubs and child pages, add metadata and typed links, then stage redirects and content patches for human review.
Use this page when
- The wiki is flat, duplicated, date-coded, title-inconsistent, or mostly one mega-index.
- AI agents cannot tell which page is canonical, current, stale, draft, evidence, or archive.
- Public SEO matters and old routes need redirects, canonical tags, sitemap cleanup, and updated internal links.
What repair mode produces
Structural health
Inventory coverage, duplicate canonicals, orphan pages, metadata completeness, owner coverage, stale review dates, and discovery leakage.
Route map
A reviewed old-to-new URL table with action, redirect type, canonical target, reason, and approval state.
Remediation JSON
A machine-readable queue of keep, move, merge, split, redirect, archive-noindex, retire, and needs-human-decision items.
Restructure packet
A human-readable plan that stages patches, source notes, lint checks, and stop conditions before publication.
Standards scorecard
| Objective | High-quality target | Repair check |
|---|---|---|
| Inventory coverage | Nearly every known page, file, route, and source record is accounted for. | Crawl, export, sitemap, llms.txt, and repository/file inventory agree. |
| Canonical integrity | No unresolved duplicate canonical pages for the same concept. | Duplicate-title and duplicate-intent clusters have one winner and redirect/archive decisions. |
| Navigation quality | Public pages have breadcrumbs, a parent hub, and useful local next hops. | Orphans, dead-end pages, and untyped related links are flagged. |
| Metadata completeness | Durable pages expose owner, status, source status, sensitivity, canonical URL, and last-reviewed data. | Missing fields go to deterministic backfill or reviewer queue. |
| Safety boundary | Restricted, private, archive-only, and generated-package pages stay out of public discovery. | sitemap.xml, llms.txt, robots.txt, and public route checks reject leaks. |
| Change control | Destructive moves, merges, retirements, and authority changes require human approval. | Only deterministic, reversible, semantics-preserving fixes are candidates for automation. |
Input and platform map
| Input | Why it matters | Typical source |
|---|---|---|
| Target registry | Defines whether this is one site, one repo, one space, or a multisite workspace. | Root URLs, repository paths, space keys, wiki roots, workspace.uai. |
| Discovery files | Shows what the public surface claims today. | sitemap.xml, llms.txt, robots.txt, export manifests, index files. |
| Content export | Preserves source material and revision context before any repair. | Git clone, XML dump, Markdown export, HTML export, media archive. |
| Governance data | Maps owners, trust states, permissions, and review cadences. | Owner directory, ACLs, space permissions, review policy, escalation rules. |
| Writeback target | Keeps repair staged instead of directly mutating production. | Pull request, merge request, shadow wiki, staging space, patch bundle. |
| Rollback snapshot | Makes route and content repair reversible. | Git tag, database/files backup, export bundle, redirect config snapshot. |
Repair sequence
- Inventory. List all pages, routes, files, indexes, source records, public URLs, sitemap entries, llms.txt entries, and internal links.
- Classify. Mark every page as keep, move, merge, split, redirect, archive-noindex, retire, or needs-human-decision.
- Find failure modes. Flag duplicate concepts, orphan pages, flat mega-indexes, date-coded reports used as concepts, vague names, repeated global nav, missing breadcrumbs, and missing meta descriptions.
- Choose canonical winners by intent. Pick the page that best answers the durable user or agent task, not the newest or longest page.
- Build hierarchy. Create section hubs, child pages, breadcrumbs, children lists, and typed related links.
- Add metadata. Add title, slug, canonical_url, description, page_type, section, audience, status, quality_class, source_status, owner, last_reviewed, sensitivity, agent_use, aliases, redirects_from, and related links.
- Create the migration map. For every old URL, record old_url, new_url, action, redirect_type, canonical_target, and reason.
- Patch links and discovery. Update internal links to final URLs, ensure moved routes redirect, refresh sitemap and llms.txt, and noindex archive-only evidence pages when needed.
- Validate. Run link, metadata, duplicate canonical, duplicate title, breadcrumb, sitemap, robots, llms.txt, and JSON output checks.
- Stage review. Save the wizard-generated restructure packet and open a pull request or review packet. Do not publish destructive changes automatically.
Page decision model
| Decision | Use when | Required record |
|---|---|---|
| Keep | The page already has the correct intent and canonical URL. | Metadata patch and related links. |
| Move | The content is useful but the URL or parent hub is wrong. | 301 redirect and internal-link update. |
| Merge | Two pages answer the same durable concept. | Canonical winner, merged source notes, redirects_from. |
| Split | A mega-page mixes concepts, guides, evidence, and reports. | Child page list and source trace. |
| Archive-noindex | The page is evidence or history, not the canonical answer. | Archive status, noindex policy, link to canonical. |
| Retire | The page is obsolete, unsupported, private, or unsafe. | Reason, reviewer, and final status. |
Automation boundary
| Fix class | Can be automated after approval | Requires human decision |
|---|---|---|
| Safe and reversible | Regenerate discovery files from a canonical registry, add obvious missing descriptions, fix internal links to reviewed final URLs, normalize harmless metadata formatting. | Only if the source registry itself is disputed. |
| Semantically risky | Draft patch bundles, split suggestions, merge candidates, and contradiction records. | Canonical winner selection, ownership assignment, trust downgrade, source interpretation, and claim rewriting. |
| High-risk or destructive | Never publish automatically. | Moves, merges, retirements, archive suppression, privacy classification, redirects with SEO impact, or any change touching secrets or regulated data. |
For AI agents
Read the LLMWikis setup wizard and the target wiki before editing. Preserve raw sources, classify pages by intent, choose one canonical URL per concept, build a semantic hierarchy, add breadcrumbs and typed links, add metadata, create redirects, and stage changes for human review. Stop on missing permission, secrets, private data, unsupported claims, destructive moves, or ambiguous authority.
Copy-ready repair prompt
Read the LLMWikis setup wizard and this restructure packet. Inventory the target wiki. Preserve raw sources. Classify pages by intent. Choose one canonical URL per concept. Build a semantic hierarchy. Add breadcrumbs and typed links. Add metadata. Create redirects. Stage changes for human review. Do not publish automatically. Stop on missing permission, secrets, private data, unsupported claims, or ambiguous authority.