source: <url or "direct input"> date: <YYYY-MM-DD> captured: <how> ``` Raw files are immutable. Never edit after saving. 2. **Surface takeaways.** Read the source. Tell the user the key points and which existing wiki pages it touches. Keep this brief; it is their curation moment. 3. **Write a source summary page** at `wiki/<slug>.md`: what it is, key claims, notable quotes with locations, `[[links]]` to entities and concepts it mentions. 4. **Update touched pages.** A single source typically touches 10-15 pages. For each relevant entity or concept page: strengthen, revise, or flag contradictions. Create new pages for important entities that don't have one yet. This integration step is the whole point. Do not skip it. 5. **Update index.md.** Add new pages (one line each, under the right section). Update changed entries. 6. **Append to log.md:** ``` ## [YYYY-MM-DD] ingest | <Source title> <one-line note of what changed, pages touched> ``` ## Operation: update User wants to add a page, revise an existing one, or record operational knowledge (runbooks, decisions, configurations). 1. Write or update the relevant `wiki/` page directly. 2. Update `index.md` if a new page was created. 3. Append to `log.md`. ## Operation: lint Health-check the wiki. Scan for and fix/report: - **Contradictions** between pages - **Stale claims** newer sources have superseded - **Orphan pages** with no inbound `[[links]]` - **Missing pages** for concepts mentioned often but lacking their own page - **Missing cross-references** between pages that should link - **Data gaps** worth a web search to fill Fix safe issues directly. Propose wholesale rewrites before acting. Append lint summary to `log.md`. ## Bootstrap: creating a new wiki ### Option 1: General Knowledge (path from the Wikis registry in `~/.claude/CLAUDE.md`) The GK wiki already exists. Its entry in the `## Wikis` section of `~/.claude/CLAUDE.md` must name the wiki's actual `index.md` path. If the registry has no General Knowledge entry yet, ask the user for that path once, then add: ``` - General Knowledge: `<absolute path to the GK wiki>\index.md` - reusable patterns across engagements ``` Then proceed with the ingest/update operation. ### Option 2: Project-specific (`./knowledgebase/`) 1. Create the folder structure: ``` ./knowledgebase/ CLAUDE.md <- vault config (sections, conventions) index.md <- empty catalog with section headers log.md <- bootstrap entry raw/ wiki/ ``` 2. Seed `knowledgebase/CLAUDE.md` with project name, owner, focus, and section headers. Default sections: Concepts, Entities, Topics, Sources, Implementation. 3. Make sure a fresh session reads the index. First check `~/.claude/settings.json` for a `SessionStart` hook that injects wiki indexes; if one is present, skip this step (harness-interview says not to duplicate it). Otherwise update the **project root** `CLAUDE.md` (create it if missing) with: ``` At session start, read ./knowledgebase/index.md for this project's accumulated knowledge. ``` 4. Add an entry to the `## Wikis` section in `~/.claude/CLAUDE.md`: ``` - <Project Name>: `<absolute path>\knowledgebase\index.md` - <one-line description> ``` 5. Append to `knowledgebase/log.md`: ``` ## [YYYY-MM-DD] bootstrap | Wiki created Project-specific wiki initialized. 0 pages. ``` The global `SessionStart` hook auto-injects this project's `index.md` from now on: it self-activates on any folder that has `./knowledgebase/index.md`, so no per-project hook wiring is needed. Then proceed with the operation that triggered the bootstrap. ## Principles (from Karpathy's design, honor them) - **Lazy, not eager.** Don't mass-summarize. Grow where the user actually works. - **User curates, you maintain.** They choose sources and ask questions. You do the bookkeeping. - **Cheap index, lazy drill-down.** index.md stays small (one line per page). - **Plain markdown, Obsidian-compatible.** `[[Wikilinks]]` everywhere. No vector DB. - **Auto-write.** Make edits directly. User's safety net is git diff. Confirm before deleting a page wholesale.

# Write Wiki > Add to, update, or health-check a knowledge base wiki. Use when the user wants to ingest a new source (article, PDF, URL, meeting notes, runbook, design doc), update existing pages, lint the wiki for stale or orphaned content, or save a synthesis back to the wiki. Also the entry point for creating a new wiki: if no wiki exists for the current project, write-wiki asks whether to create a General Knowledge wiki (path from the Wikis registry in ~/.claude/CLAUDE.md) or a project-specific one (inside the current folder). Triggers on /write-wiki, /wiki, "add this to the wiki", "ingest this", "save this to the wiki", "update the wiki", "lint the wiki", "file this", "remember this source". - Skill: `yogiraja/write-wiki` (Agent Skill) - Install (CLI): `npx skillmds@latest add yogiraja/write-wiki` - Raw SKILL.md: https://api.skillmd.com/api/skills/yogiraja/write-wiki/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Docs & Writing - Author: YogirajA (https://skillmd.com/u/yogiraja) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/yogiraja/write-wiki --- # write-wiki: Add to and Maintain the Knowledge Base ## Step 0: Orient Read `~/.claude/CLAUDE.md` and find the `## Wikis` section. This tells you every known wiki and its path. Also check for `./knowledgebase/index.md` in the current working directory. ## Step 1: Does a wiki exist for this context? **If yes:** proceed to Step 2 with the matching wiki path. **If no wiki exists for the current project:** Ask the user once: > No wiki found for this project. Create one? > [1] General Knowledge: adds to the GK wiki named in the Wikis registry in `~/.claude/CLAUDE.md` (reusable patterns across engagements) > [2] Project-specific: creates `./knowledgebase/` inside this project folder Then bootstrap based on their choice (see Bootstrap section below) and proceed. ## Step 2: Determine the operation Infer from context; don't ask unless truly unclear. | Intent | Operation | |---|---| | User drops a URL, file, or paste of content | **ingest** | | User says "update", "add a page", "save this" | **ingest** or **update** | | User says "lint", "clean up", "audit" | **lint** | | User says "save this answer" after a read-wiki session | **ingest** (save synthesis as a new wiki page) | ## Operation: ingest A new source arrives (URL, file, paste, meeting notes, runbook, design doc). 1. **Save to raw/.** Fetch the URL with WebFetch if needed; a YouTube URL goes through the `watch` skill for its transcript first. Save as `raw/.md` with a header block: ``` title: source: <url or "direct input"> date: <YYYY-MM-DD> captured: <how> ``` Raw files are immutable. Never edit after saving. 2. **Surface takeaways.** Read the source. Tell the user the key points and which existing wiki pages it touches. Keep this brief; it is their curation moment. 3. **Write a source summary page** at `wiki/<slug>.md`: what it is, key claims, notable quotes with locations, `[[links]]` to entities and concepts it mentions. 4. **Update touched pages.** A single source typically touches 10-15 pages. For each relevant entity or concept page: strengthen, revise, or flag contradictions. Create new pages for important entities that don't have one yet. This integration step is the whole point. Do not skip it. 5. **Update index.md.** Add new pages (one line each, under the right section). Update changed entries. 6. **Append to log.md:** ``` ## [YYYY-MM-DD] ingest | <Source title> <one-line note of what changed, pages touched> ``` ## Operation: update User wants to add a page, revise an existing one, or record operational knowledge (runbooks, decisions, configurations). 1. Write or update the relevant `wiki/` page directly. 2. Update `index.md` if a new page was created. 3. Append to `log.md`. ## Operation: lint Health-check the wiki. Scan for and fix/report: - **Contradictions** between pages - **Stale claims** newer sources have superseded - **Orphan pages** with no inbound `[[links]]` - **Missing pages** for concepts mentioned often but lacking their own page - **Missing cross-references** between pages that should link - **Data gaps** worth a web search to fill Fix safe issues directly. Propose wholesale rewrites before acting. Append lint summary to `log.md`. ## Bootstrap: creating a new wiki ### Option 1: General Knowledge (path from the Wikis registry in `~/.claude/CLAUDE.md`) The GK wiki already exists. Its entry in the `## Wikis` section of `~/.claude/CLAUDE.md` must name the wiki's actual `index.md` path. If the registry has no General Knowledge entry yet, ask the user for that path once, then add: ``` - General Knowledge: `<absolute path to the GK wiki>\index.md` - reusable patterns across engagements ``` Then proceed with the ingest/update operation. ### Option 2: Project-specific (`./knowledgebase/`) 1. Create the folder structure: ``` ./knowledgebase/ CLAUDE.md <- vault config (sections, conventions) index.md <- empty catalog with section headers log.md <- bootstrap entry raw/ wiki/ ``` 2. Seed `knowledgebase/CLAUDE.md` with project name, owner, focus, and section headers. Default sections: Concepts, Entities, Topics, Sources, Implementation. 3. Make sure a fresh session reads the index. First check `~/.claude/settings.json` for a `SessionStart` hook that injects wiki indexes; if one is present, skip this step (harness-interview says not to duplicate it). Otherwise update the **project root** `CLAUDE.md` (create it if missing) with: ``` At session start, read ./knowledgebase/index.md for this project's accumulated knowledge. ``` 4. Add an entry to the `## Wikis` section in `~/.claude/CLAUDE.md`: ``` - <Project Name>: `<absolute path>\knowledgebase\index.md` - <one-line description> ``` 5. Append to `knowledgebase/log.md`: ``` ## [YYYY-MM-DD] bootstrap | Wiki created Project-specific wiki initialized. 0 pages. ``` The global `SessionStart` hook auto-injects this project's `index.md` from now on: it self-activates on any folder that has `./knowledgebase/index.md`, so no per-project hook wiring is needed. Then proceed with the operation that triggered the bootstrap. ## Principles (from Karpathy's design, honor them) - **Lazy, not eager.** Don't mass-summarize. Grow where the user actually works. - **User curates, you maintain.** They choose sources and ask questions. You do the bookkeeping. - **Cheap index, lazy drill-down.** index.md stays small (one line per page). - **Plain markdown, Obsidian-compatible.** `[[Wikilinks]]` everywhere. No vector DB. - **Auto-write.** Make edits directly. User's safety net is git diff. Confirm before deleting a page wholesale.