[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M78F3ZZR0P038GFDMVP7F6C7":3,"skill-same:01M78F3ZZR0P038GFDMVP7F6C7":45},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":30,"install":31,"resources":44},"01M78F3ZZR0P038GFDMVP7F6C7","ghl-knowledge-base",null,"Build and manage GoHighLevel knowledge bases — the content that feeds both Conversation AI and Voice AI. Covers rich-text documents, FAQs, web crawling, files, tables, Google Drive\u002FSheets and Internal Data (products) sources, plus the trigger conditions that tell an agent WHEN to use each knowledge base. Use when the user says 'add a knowledge base', 'train the bot on this', 'the AI isn't using my KB', 'add FAQs', 'crawl my website', 'upload docs for the agent', or asks why an agent answers from the wrong source. Internal API — six of the nine source types have no public equivalent.","uxieee\u002Fuxie-ghl-factory","uxieee","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F83715980?v=4","01M78F1K57TCCXME78DK6XTK6S","uxie","01M78F1K5WT9N3WHVE4N532ZWM",12,"MIT","https:\u002F\u002Fgithub.com\u002Fuxieee\u002Fuxie-ghl-factory\u002Fblob\u002Fmain\u002Fplugins\u002Fuxie-ghl-factory\u002Fskills\u002Fghl-knowledge-base\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: ghl-knowledge-base\ndescription: \"Build and manage GoHighLevel knowledge bases — the content that feeds both Conversation AI and Voice AI. Covers rich-text documents, FAQs, web crawling, files, tables, Google Drive\u002FSheets and Internal Data (products) sources, plus the trigger conditions that tell an agent WHEN to use each knowledge base. Use when the user says 'add a knowledge base', 'train the bot on this', 'the AI isn't using my KB', 'add FAQs', 'crawl my website', 'upload docs for the agent', or asks why an agent answers from the wrong source. Internal API — six of the nine source types have no public equivalent.\"\n---\n\n# GHL Knowledge Base\n\nThe shared content layer. Both Voice AI (`knowledgeBaseIds`) and Conversation AI\n(`knowledgeBaseIds` + `knowledgeBaseTriggers`) point at the same knowledge bases, so a change\nhere affects every agent attached to it. Check `GET \u002Fknowledge-base\u002Fassociated-entities`\nbefore editing one.\n\nBase: `services.leadconnectorhq.com\u002Fknowledge-base`.\n\n## Why this is an internal-rail skill\n\nNine source types exist in the enum; the UI offers five tabs (Web crawler · FAQ · Tables · Rich text · Files,\nwith Google Drive\u002FSheets and Internal Data behind flags). **Web search has no screen** at build 683 (hard-coded\noff). The public API documents FAQs, the crawler, files and the knowledge-base record itself (the public file\nupload is documented, not yet proven live). **Five source types have no public equivalent**, and `web_search`\nhas neither a screen nor a public operation:\n\n```\nfaq  ·  web_crawler  ·  file                                        ← public reaches these too\nrich_text  ·  table  ·  google_drive  ·  google_sheet  ·  internal_data   ← internal only\nweb_search                                                          ← no screen, no public op\n```\n\n`rich_text` is the one that matters most in practice: it is how you give an agent authored\nprose rather than scraped pages, and there is no way to create one through the public API.\n\n## Which rail, per job\n\nThe singular internal `\u002Fknowledge-base\u002F*` paths still answer after GHL's 2026-08-18 rename; on the sandbox they returned\nthe same knowledge bases as the public plural `\u002Fknowledge-bases` (2026-10-29).\n\n\u003C!-- rail-ledger:knowledge-base -->\n_generated from the rail ledger (mcp-internal\u002Fcatalog\u002Frail-ledger.json) — do not edit; the whole ledger is `ghl-orientation\u002Freferences\u002Frail-ledger.md`._\n\n| Job | Decision | Routed today | Internal | Public |\n|---|---|---|---|---|\n| read | both | internal | `raw_request:GET \u002Fknowledge-base\u002Fall`, `raw_request:GET \u002Fknowledge-base\u002F{knowledgeBaseId}` | `knowledge-base-v3__listAllKnowledgeBasesPaginated`, `knowledge-base-v3__getKnowledgeBaseById` |\n| create | undecided | internal | `raw_request:POST \u002Fknowledge-base\u002F` | `knowledge-base-v3__createKnowledgeBase`, `knowledge-base-v3__create` |\n| edit | undecided | internal | `raw_request:PUT \u002Fknowledge-base\u002F{knowledgeBaseId}` | `knowledge-base-v3__updateKnowledgeBase`, `knowledge-base-v3__update` |\n| delete | undecided | internal | `raw_request:DELETE \u002Fknowledge-base\u002F{knowledgeBaseId}` | `knowledge-base-v3__deleteKnowledgeBase`, `knowledge-base-v3__delete` |\n\n- **read:** The internal rail reaches source types the public one lacks; whether its singular paths still answer after the rename is unverified.\n- **create:** The internal rail reaches source types the public one lacks; whether its singular paths still answer after the rename is unverified.\n- **edit:** The internal rail reaches source types the public one lacks; whether its singular paths still answer after the rename is unverified.\n- **delete:** The internal rail reaches source types the public one lacks; whether its singular paths still answer after the rename is unverified.\n\nInternal only: rich_text, table, google and internal_data sources; gaps.\nPublic only: FAQs; crawler sitemap preview; files.\n\u003C!-- \u002Frail-ledger:knowledge-base -->\n\n## The endpoints\n\nAll 25 are in the endpoint catalogue — `search_endpoints`, then `describe_endpoint`. The corpus\npage `ai-agents\u002F20-api\u002Fknowledge-base.md` carries the source-type enum, the analytics vocabulary\nthat reveals the per-plan caps, and the training routes.\n\nThe shape worth holding in your head: a knowledge base is a record with **sources attached to it**.\n`POST \u002Fknowledge-base\u002F` makes the record (body `{locationId, name, description?}`, name ≤ 50). **Every source\nkind has its own sub-resource** — `\u002Ffaqs`, `\u002Frich-text\u002F`, `\u002Fcrawler`, `\u002Ffiles`, `\u002Ftable\u002F…` (a multi-step\nupload → schema → select-columns → status pipeline), `\u002Fgoogle-sheets\u002F*`, `\u002Fgoogle-drive\u002F*`, `\u002Finternal-data\u002F*`.\nThere is **no generic \"attach a source\" call**: the POST to a bare KB id and its `\u002Fbulk` variant that older\nnotes listed are the translation client's calls, not the KB's. Creation is **asynchronous** — poll each source's\nstatus. 15 knowledge bases per sub-account.\n\n### The record (all live-proven 2026-09-26 on a test sub-account, each read back on a separate request)\n\n| Operation | Call | Notes |\n|---|---|---|\n| Create | `POST \u002Fknowledge-base\u002F` `{locationId, name, description?}` | **keep the trailing slash**. → 201 `{data:{id, name, nameLowerCase, kbMetadata{faqs, urls, richText, webSearches, files, tables, firstQueryAt, lastQueryAt, lastSourceAddedAt}}}`. The editor caps the name at 50 characters and the description at 175; the server stored a 56-character name, so the name cap is client-only |\n| Read | `GET \u002Fknowledge-base\u002F{id}?locationId=` | carries `description` and the `kbMetadata` source counts |\n| List | `GET \u002Fknowledge-base\u002Fall?locationId=` | → `{data:{knowledgeBases[{id, name, createdAt, updatedAt}], activeCount}}`. `activeCount` is what the editor's \"Knowledge base quota n \u002F 15\" reads; the 15 is editor-side (not pushed past on the test account) |\n| Rename \u002F describe | `PUT \u002Fknowledge-base\u002F{id}` `{locationId, knowledgeBaseId, name, description}` | → 200; the editor's header pencil opens \"Edit knowledge base\" |\n| Who uses it | `GET \u002Fknowledge-base\u002Fassociated-entities?locationId=&knowledgeBaseId=` | → `{entities, blockingEntities, advisoryEntities, unverifiedSources, partial}`; an unused KB reads all empty with `partial:false` |\n| Delete | `DELETE \u002Fknowledge-base\u002F{id}?locationId=` | → `{success}`; afterwards `GET` answers **400 \"Knowledge base not found\"**. The editor refuses while agents use it: \"This knowledge base cannot be deleted … Agents using this knowledge base\" lists each agent (Conversation AI \u002F Voice AI) with Open and a **Re-check** button. Read `associated-entities` yourself before deleting through the API, which does not stop you |\n| Default KB | `POST \u002Fknowledge-base\u002Fdefault` `{locationId, migrateDocs:true}` | → 201 `{data:{id}}`, the same id every time: the location's default KB. The list page fires it on every load |\n\nRead-back list of a KB's rich-text docs: `GET \u002Fknowledge-base\u002Frich-text\u002Fknowledge-base\u002F{kbId}` →\n`{data: [{id, title, content, contentMarkdown, status, …}]}`.\n\n## The traps\n\n**1. Rich-text create is asynchronous.** `POST \u002Fknowledge-base\u002Frich-text\u002F` returns before the\ndocument is usable — there is a status poll. Do not attach the KB to an agent and report\nsuccess on the create response alone.\n\n**2. Opening the Knowledge Base screen writes.** `POST \u002Fknowledge-base\u002Fdefault` fires on page\nload and returns 201. If you are replaying captured traffic, that call is expected behaviour,\nnot something you triggered.\n\n**3. Attaching a KB is not enough — the agent needs to know WHEN to use it.** On Conversation\nAI:\n\n```jsonc\n\"knowledgeBaseTriggers\": [\n  { \"mode\": \"all\",    \"knowledgeBaseIds\": [\"…\"], \"triggerCondition\": \"\",  \"priority\": 2 },\n  { \"mode\": \"custom\", \"knowledgeBaseIds\": [\"…\"], \"triggerCondition\": \"\u003Cwhen to use this>\", \"priority\": 1 }\n]\n```\n\nThe editor labels the text **\"Instructions (Optional)\"**, and it is optional: with no text the agent decides by\nitself when to use the KB (`mode: \"all\"`, one such trigger, ≤ 7 KBs); with text the trigger is `mode: \"custom\"`\n(≤ 5 KBs). The editor caps the text at 1000 characters, but the server stored 1001. The **server** allows at most 4\ntriggers (`\"knowledgeBaseTriggers.4.priority must not be greater than 4\"`). A trial chat on a bot with a KB answered\nfrom that KB (live-proven 2026-09-26). An older version of this skill said the text was required, 10–500\ncharacters, labelled \"When to use this knowledge base\": that was wrong.\n\n**4. Gaps are a dated log of misses, not an inventory of what is missing now.**\n🔴 A row stays `open` after the answering content is added — proven by differential (documents\nadded, the questions re-asked live and answered from them, the list re-read unchanged, `lastAskedAt`\nunmoved) — and a question the KB answers is never logged, so every row looks current until you\nread its dates. List: `GET \u002Fknowledge-base\u002Fgaps?locationId=&knowledgeBaseId=&status=open` (AI rail).\n1. **Read the knowledge base first** — content present ⇒ the row is stale; do not change the agent.\n2. **Check `lastAskedAt`** against the window you are judging.\n3. **Write the knowledge to match `topQueryTexts`** (the customer's own wording).\n4. **Never filter by `categories`** — a genuine product question was filed under *Noise \u002F Gibberish \u002F Chitchat*.\nThe DISMISS write is **`PATCH \u002Fknowledge-base\u002Fgaps\u002F{gapId}\u002Fstate`** `{locationId, state:\"archived\"}` —\nlive-proven 2026-09-08, reversible with `state:\"active\"`. The write says `active`, the read says `open`. Fields and counts endpoint: `knowledge\u002Fcorpus\u002Fai-agents\u002F20-api\u002Fknowledge-base.md` → \"Gaps\".\n\n**5. Editing a rich-text doc is a PUT, not delete-and-recreate.** `PUT\n\u002Fknowledge-base\u002Frich-text\u002F:id` is a **live-verified full-replace** (2026-08-28, the designated\ntest sub-account): same body shape as create — `{locationId, knowledgeBaseId, title, content}` —\n200, response carries `status: \"training\"`, and a read-back of the sent `content` came back\n**byte-identical**. It re-chunks and re-embeds automatically, exactly like create: poll `GET\n\u002Fknowledge-base\u002Frich-text\u002F:id\u002Fstatus` until `\"trained\"` (a full retrain took ~4.6s in the proving\nrun) — there is no separate retrain call.\n\nDo not delete-and-recreate a doc to edit it, and do not add a second doc alongside the old one\nas a workaround: both orphan or duplicate content — a delete drops any id another object\nreferences, and a second doc leaves stale content the agent can still draw from, which is a\nworse failure than the one you were trying to fix. PUT the existing id instead.\n\n`content` is HTML; the server derives `contentMarkdown` from it. **A direct `contentMarkdown`\nwrite 200s and changes nothing** — measured in both a contentMarkdown-only body and an\nunchanged-content-plus-contentMarkdown body. `kb-compiler.mjs`'s `compileRichTextUpdate(id, doc,\n{locationId})` builds this PUT descriptor, and throws rather than silently no-opping if the\ncaller's IR carries a `contentMarkdown` key — a caller supplying that key believes something\nfalse about this API.\n\n**6. The retrieval test needs the editor's body.** `POST \u002Fknowledge-base\u002Fchunks` `{query, knowledgeBaseIds,\nlocationId, rerankChunks: true, threshold: 0, topK: 5, subCaller}` returns up to 5 chunks with `score` and\n`category` (`bulk` = crawled page, `finetuned` = FAQ, `rich_text`, `files`, `table`). Without those parameters it\nreturned one chunk, which looks like \"the KB only knows one thing\". The editor's ⌘K \"Test retrieval\" drawer keeps\na history: `POST \u002Fknowledge-base\u002Fretrieval-test\u002Fconversations\u002Fentries` `{locationId, knowledgeBaseId, query,\nagentLayer, chunks, totalChunks, latencyMs, idempotencyKey}` → `{conversationId, entryId, isNew, isDuplicate}`;\n`GET …\u002Fretrieval-test\u002Fconversations?locationId=&knowledgeBaseId=` lists them, `GET …\u002Fconversations\u002F{id}?locationId=`\nreads one, `DELETE` the same removes it (then 404 \"Conversation not found\").\n\n**7. The crawler judges what the CDN serves, refuses thin pages, and obeys robots.txt.** Status `Restricted` with `error:\n\"Insufficient content (174 chars)\"`. A page you just edited can be judged on its stale cached copy. 🔴 The crawler reads\nrobots.txt as user agent **`AITrainingBot`**. A blocked page still gets **201 `status: \"Processing\"`** on the POST, then\n`status: \"Restricted\"`, `error.code: \"ROBOTS_TXT_BLOCKED\"`, `retryable: false`, and no URL is stored. GHL's own\n`app.gohighlevel.com\u002Fv2\u002Fpreview\u002F{pageId}` URLs are blocked this way (proven live 2026-09-29). **A 201 is not a crawl:**\nread `GET \u002Fknowledge-base\u002Fcrawler\u002Fstatus?locationId=&knowledgeBaseId=` for `operationDetails.status` and `.error`, and\npoint the user to Rich Text when robots.txt blocks the page.\n\n## Sources — what ran live (2026-09-26 \u002F 28)\n\nEach was built in an own test KB, read back on separate requests, retrieved through `\u002Fchunks`, and deleted by exact\nid. Full shapes: corpus `ai-agents\u002F20-api\u002Fknowledge-base.md` → \"Sources executed live\".\n\n| Source | Create | Read | Remove | Traps |\n|---|---|---|---|---|\n| FAQ | `POST \u002Fknowledge-base\u002Ffaqs` `{locationId, knowledgeBaseId, question, answer, metadata?}` → `{faq{…, id}}` | `GET \u002Fknowledge-base\u002Ffaqs?locationId=&knowledgeBaseIds=&limit=&offset=` → `{faqs, count, hasMore}` | — | a duplicate question → **409** \"A FAQ with this question already exists in this knowledge base\"; the 1000-character answer cap is client-only (1001 stored) |\n| Rich text | `POST \u002Fknowledge-base\u002Frich-text\u002F` `{locationId, knowledgeBaseId, title, content (HTML)}` → `status:\"training\"` | `GET \u002Fknowledge-base\u002Frich-text\u002F{id}\u002Fstatus?locationId=` → `{sourceType, status, stages[CHUNKING, EMBEDDING]{state, durationMs}}` | `DELETE \u002Fknowledge-base\u002Frich-text\u002F{id}` | trained in ~3 s; the server adds generated questions to the chunk; edit with PUT (Trap 5); the editor also offers bulk delete (`\u002Frich-text\u002Fbulk`) and a per-document cancel (`\u002Frich-text\u002F{id}\u002Fcancel`), neither executed |\n| Web crawler | `POST \u002Fknowledge-base\u002Fcrawler` `{locationId, url, option: \"Exact\" \\| \"Path\" \\| \"Domain\", knowledgeBaseId}` → `{operationId, status:\"Processing\"}` | `GET \u002Fknowledge-base\u002Fcrawler?locationId=&knowledgeBaseId=` → `{count, urls[]}`; `GET …\u002Fcrawler\u002Fstatus` = the **latest** operation only; `GET …\u002Fcrawler\u002Fpages\u002Fcontent?locationId=&urlId=&knowledgeBaseId=` → the scraped text | `DELETE \u002Fknowledge-base\u002Fcrawler` `{locationId, urlIds, deleteAll:false, knowledgeBaseId}` → `{deletedCount}` | Exact mode **trains with no separate train call**; `sitemap-preview` refuses Exact (400); the URL refresh `PUT \u002Fknowledge-base\u002Fcrawler?locationId=&knowledgeBaseId=` `{urlIds, refreshAll, urlIdsToSkip}` answered 200 and changed nothing within 8 s, so its effect is unproven; Path \u002F Domain, cancel and retry are unexercised |\n| URL auto-refresh | `POST \u002Fknowledge-base\u002Fcrawler\u002Fautomation` `{knowledgeBaseId, refreshFrequency: DAILY \\| WEEKLY \\| MONTHLY}` → an ACTIVE `WEBSITE_RETRAIN` task, first run a week out for WEEKLY | `GET …\u002Fcrawler\u002Fautomation\u002Flatest?knowledgeBaseId=` | `DELETE …\u002Fcrawler\u002Fautomation\u002F{id}` (then `latest` → `data:null`) | the update `PUT …\u002Fautomation\u002F{id}` needs `taskId` = the automation's own id (422 \"taskId must be a string\" without it) |\n| File | **multipart** `POST \u002Fknowledge-base\u002Ffiles` | `GET \u002Fknowledge-base\u002Ffiles\u002F{id}`; `…\u002F{id}\u002Fstatus` stages CONVERSION → EXTRACTION → CHUNKING → EMBEDDING | `DELETE \u002Fknowledge-base\u002Ffiles\u002F{id}` (then 400 \"Invalid file ID\") | ≤ 10 MB, `.pdf .doc .docx .md`; trained in ~3 s |\n| Table (CSV) | **multipart** `POST \u002Fknowledge-base\u002Ftable\u002Flocation\u002F{loc}\u002Fkb\u002F{kb}\u002Fupload`, then `POST …\u002F{fileId}\u002Fselect-columns` `{selectedColumns[{name, originalType, selectedType, isSelected, sampleValue, displayName, searchable, required}]}` | `GET …\u002F{fileId}`, `…\u002Fstatus` (ANALYZING → PROCESSING → INDEXING), `…\u002Fdata?page=&limit=` | `DELETE …\u002F{fileId}` (then 404) | ≤ 50 MB; trained in ~48 s |\n\n🔴 **Files and tables are multipart uploads.** `raw_request` sends JSON only, so neither can be created through\nthis plugin today: point the user to the editor's Files \u002F Tables tab.\n\nThe scraped text of a crawled page is edited through View scraped content → Save:\n`PUT \u002Fknowledge-base\u002Fcrawler\u002Fpages\u002Fcontent?locationId=&urlId=&knowledgeBaseId=` `{content}` → 200 with an empty body. The\nrow retrains and reads back `status: \"Successful\"`, **`contentEditedByUser: true`**, and retrieval returns the edited text\n(proven live 2026-09-29). Retraining the URL from the website reverts the edit (the editor's warning). The URL list's\n`content` field is a storage URL to the text file, not the text itself. The `\u002Fconversations-ai\u002Ftrain\u002Fchunk\u002Fcontent`\nroute older notes gave is not called by the Knowledge Base app.\n\n**Gaps**: `GET \u002Fknowledge-base\u002Fgaps\u002Fcounts?locationId=` → `{data[{knowledgeBaseId, gapCount}]}` for the whole\nlocation; `GET \u002Fknowledge-base\u002Fgaps?locationId=&knowledgeBaseId=&status=open&page=&perPage=` →\n`{gaps[], stats{openGaps, queriesAffected, resolvedThisMonth, archivedGaps}, total}`. The editor's tabs are Needs\nanswers · Answered · Archived. `POST \u002Fknowledge-base\u002Fgaps\u002F{gapId}\u002Ffill {title, answer, locationId}` writes into the\nKB itself and has not been executed.\n\n## What GHL can do that this plugin cannot reach on the test account\n\nIf the user wants one of these, **GHL can do it** — say so and point to the screen.\n\n| Capability | Where in GHL | Why not here |\n|---|---|---|\n| Import files from Google Drive, daily Drive sync | KB → Files → \"Import from Google Drive\" (flag `knowledgeBase.googleSheets`) | needs a connected Google account (`GET \u002Fknowledge-base\u002Foauth?locationId=&type=google`); the only one on the test account is a person's own, so it was not used |\n| Google Sheet as a table, auto \u002F scheduled sync | KB → Tables → \"Import from Google Sheets\" | same |\n| Internal Data (sync the sub-account's Products) | KB source behind flag `knowledgeBase.internalData` | `GET \u002Fknowledge-base\u002Finternal-data\u002Fsources` answered 403 \"Forbidden resource\" on the test account |\n| Web search as a source | no screen: the editor hard-codes it off | the client ships the module, but `GET \u002Fknowledge-base\u002Fweb-search\u002Fkb\u002F{id}` answered 404 \"Cannot GET\" |\n| Create with AI (Beta) | KB list → \"Create with AI\" | it opens the Ask AI copilot (\"Create a knowledge base from brand voice\") and makes no KB call of its own |\n| Open the KB screen at all | AI Agents → Knowledge Base | the page sits behind the `conversation_AI` billing opt-in (`config.optIn`), not `enabled` |\n\n## Writing FAQs from real conversations\n\nWhen the job is to fill a knowledge base from what customers already asked, keep only the questions a customer asked to\nget facts (products, services, prices, hours, policies, how things work). Drop greetings, bookings, one-word replies,\nstaff-only notes and system lines. Take the question from the customer's side and the answer from the business's side.\nRemove anything personal (names, numbers, addresses, billing details) and anything only true on one day (\"closed today\",\na slot time). Rewrite colloquial questions so each one stands alone and names its product. Answers the bot gave that\nsaid it did not know are the most useful candidates: complete them where the facts are known. Where a value is unknown\nand changes (a price, hours, a phone number), write a clear placeholder such as [price] instead of inventing it or leaving\nthe answer empty. Merge questions that ask the same thing into one pair, then save them in one bulk call and read back\nhow many were created, updated or failed.\n\n## Limits\n\nFiles ≤ 10 MB (`.pdf .doc .docx .md`) · table CSV ≤ 50 MB · rich text ≤ 25,000 characters and 512 KiB of HTML\n(`richTextBlockedByLimit` fires on the character cap, not on a document count) · FAQ answer ≤ 1000 (client-only) ·\nKB name ≤ 50 (client-only) and description ≤ 175 · ≤ 7,000 crawled pages per KB · 15 KBs per sub-account · ≤ 4\nknowledge-base triggers per bot (server) · at least one KB must be selected where a bot requires one.\n\n## Proof status — read before trusting a write\n\n**Live-proven** (executed on a test sub-account and read back on a separate request): the KB record's create, read,\nlist, rename, delete and associated-entities read; FAQ, rich text (create, full-replace PUT, delete), web crawler\n(Exact mode, delete, auto-refresh create\u002Fupdate\u002Fdelete), file and table sources (uploaded from the editor, then\nread and deleted through the API); retrieval and its history; the gap list and the gap archive\u002Frestore write.\nAlso live-proven (2026-09-29): the scraped-text edit, and a robots.txt-blocked crawl's status and error. **Not proven:** the\ncrawl refresh's effect; Path \u002F Domain crawls; crawl cancel \u002F retry (an Exact crawl of one page finishes in about 2 s, and\nanything longer needs a multi-page crawl); gap fill; and everything in the table above. Treat a first write of an unproven operation as a throwaway validation run\non a test sub-account.\n\n## Scope\n\nKnowledge bases only. The agents that consume them are `ghl-voice-ai` (internal) and\nConversation AI (lists its bots and configures them on the internal rail — see `ghl-conversation-ai`).\n","# GHL Knowledge Base\n\nThe shared content layer. Both Voice AI (`knowledgeBaseIds`) and Conversation AI\n(`knowledgeBaseIds` + `knowledgeBaseTriggers`) point at the same knowledge bases, so a change\nhere affects every agent attached to it. Check `GET \u002Fknowledge-base\u002Fassociated-entities`\nbefore editing one.\n\nBase: `services.leadconnectorhq.com\u002Fknowledge-base`.\n\n## Why this is an internal-rail skill\n\nNine source types exist in the enum; the UI offers five tabs (Web crawler · FAQ · Tables · Rich text · Files,\nwith Google Drive\u002FSheets and Internal Data behind flags). **Web search has no screen** at build 683 (hard-coded\noff). The public API documents FAQs, the crawler, files and the knowledge-base record itself (the public file\nupload is documented, not yet proven live). **Five source types have no public equivalent**, and `web_search`\nhas neither a screen nor a public operation:\n\n```\nfaq  ·  web_crawler  ·  file                                        ← public reaches these too\nrich_text  ·  table  ·  google_drive  ·  google_sheet  ·  internal_data   ← internal only\nweb_search                                                          ← no screen, no public op\n```\n\n`rich_text` is the one that matters most in practice: it is how you give an agent authored\nprose rather than scraped pages, and there is no way to create one through the public API.\n\n## Which rail, per job\n\nThe singular internal `\u002Fknowledge-base\u002F*` paths still answer after GHL's 2026-08-18 rename; on the sandbox they returned\nthe same knowledge bases as the public plural `\u002Fknowledge-bases` (2026-10-29).\n\n\u003C!-- rail-ledger:knowledge-base -->\n_generated from the rail ledger (mcp-internal\u002Fcatalog\u002Frail-ledger.json) — do not edit; the whole ledger is `ghl-orientation\u002Freferences\u002Frail-ledger.md`._\n\n| Job | Decision | Routed today | Internal | Public |\n|---|---|---|---|---|\n| read | both | internal | `raw_request:GET \u002Fknowledge-base\u002Fall`, `raw_request:GET \u002Fknowledge-base\u002F{knowledgeBaseId}` | `knowledge-base-v3__listAllKnowledgeBasesPaginated`, `knowledge-base-v3__getKnowledgeBaseById` |\n| create | undecided | internal | `raw_request:POST \u002Fknowledge-base\u002F` | `knowledge-base-v3__createKnowledgeBase`, `knowledge-base-v3__create` |\n| edit | undecided | internal | `raw_request:PUT \u002Fknowledge-base\u002F{knowledgeBaseId}` | `knowledge-base-v3__updateKnowledgeBase`, `knowledge-base-v3__update` |\n| delete | undecided | internal | `raw_request:DELETE \u002Fknowledge-base\u002F{knowledgeBaseId}` | `knowledge-base-v3__deleteKnowledgeBase`, `knowledge-base-v3__delete` |\n\n- **read:** The internal rail reaches source types the public one lacks; whether its singular paths still answer after the rename is unverified.\n- **create:** The internal rail reaches source types the public one lacks; whether its singular paths still answer after the rename is unverified.\n- **edit:** The internal rail reaches source types the public one lacks; whether its singular paths still answer after the rename is unverified.\n- **delete:** The internal rail reaches source types the public one lacks; whether its singular paths still answer after the rename is unverified.\n\nInternal only: rich_text, table, google and internal_data sources; gaps.\nPublic only: FAQs; crawler sitemap preview; files.\n\u003C!-- \u002Frail-ledger:knowledge-base -->\n\n## The endpoints\n\nAll 25 are in the endpoint catalogue — `search_endpoints`, then `describe_endpoint`. The corpus\npage `ai-agents\u002F20-api\u002Fknowledge-base.md` carries the source-type enum, the analytics vocabulary\nthat reveals the per-plan caps, and the training routes.\n\nThe shape worth holding in your head: a knowledge base is a record with **sources attached to it**.\n`POST \u002Fknowledge-base\u002F` makes the record (body `{locationId, name, description?}`, name ≤ 50). **Every source\nkind has its own sub-resource** — `\u002Ffaqs`, `\u002Frich-text\u002F`, `\u002Fcrawler`, `\u002Ffiles`, `\u002Ftable\u002F…` (a multi-step\nupload → schema → select-columns → status pipeline), `\u002Fgoogle-sheets\u002F*`, `\u002Fgoogle-drive\u002F*`, `\u002Finternal-data\u002F*`.\nThere is **no generic \"attach a source\" call**: the POST to a bare KB id and its `\u002Fbulk` variant that older\nnotes listed are the translation client's calls, not the KB's. Creation is **asynchronous** — poll each source's\nstatus. 15 knowledge bases per sub-account.\n\n### The record (all live-proven 2026-09-26 on a test sub-account, each read back on a separate request)\n\n| Operation | Call | Notes |\n|---|---|---|\n| Create | `POST \u002Fknowledge-base\u002F` `{locationId, name, description?}` | **keep the trailing slash**. → 201 `{data:{id, name, nameLowerCase, kbMetadata{faqs, urls, richText, webSearches, files, tables, firstQueryAt, lastQueryAt, lastSourceAddedAt}}}`. The editor caps the name at 50 characters and the description at 175; the server stored a 56-character name, so the name cap is client-only |\n| Read | `GET \u002Fknowledge-base\u002F{id}?locationId=` | carries `description` and the `kbMetadata` source counts |\n| List | `GET \u002Fknowledge-base\u002Fall?locationId=` | → `{data:{knowledgeBases[{id, name, createdAt, updatedAt}], activeCount}}`. `activeCount` is what the editor's \"Knowledge base quota n \u002F 15\" reads; the 15 is editor-side (not pushed past on the test account) |\n| Rename \u002F describe | `PUT \u002Fknowledge-base\u002F{id}` `{locationId, knowledgeBaseId, name, description}` | → 200; the editor's header pencil opens \"Edit knowledge base\" |\n| Who uses it | `GET \u002Fknowledge-base\u002Fassociated-entities?locationId=&knowledgeBaseId=` | → `{entities, blockingEntities, advisoryEntities, unverifiedSources, partial}`; an unused KB reads all empty with `partial:false` |\n| Delete | `DELETE \u002Fknowledge-base\u002F{id}?locationId=` | → `{success}`; afterwards `GET` answers **400 \"Knowledge base not found\"**. The editor refuses while agents use it: \"This knowledge base cannot be deleted … Agents using this knowledge base\" lists each agent (Conversation AI \u002F Voice AI) with Open and a **Re-check** button. Read `associated-entities` yourself before deleting through the API, which does not stop you |\n| Default KB | `POST \u002Fknowledge-base\u002Fdefault` `{locationId, migrateDocs:true}` | → 201 `{data:{id}}`, the same id every time: the location's default KB. The list page fires it on every load |\n\nRead-back list of a KB's rich-text docs: `GET \u002Fknowledge-base\u002Frich-text\u002Fknowledge-base\u002F{kbId}` →\n`{data: [{id, title, content, contentMarkdown, status, …}]}`.\n\n## The traps\n\n**1. Rich-text create is asynchronous.** `POST \u002Fknowledge-base\u002Frich-text\u002F` returns before the\ndocument is usable — there is a status poll. Do not attach the KB to an agent and report\nsuccess on the create response alone.\n\n**2. Opening the Knowledge Base screen writes.** `POST \u002Fknowledge-base\u002Fdefault` fires on page\nload and returns 201. If you are replaying captured traffic, that call is expected behaviour,\nnot something you triggered.\n\n**3. Attaching a KB is not enough — the agent needs to know WHEN to use it.** On Conversation\nAI:\n\n```jsonc\n\"knowledgeBaseTriggers\": [\n  { \"mode\": \"all\",    \"knowledgeBaseIds\": [\"…\"], \"triggerCondition\": \"\",  \"priority\": 2 },\n  { \"mode\": \"custom\", \"knowledgeBaseIds\": [\"…\"], \"triggerCondition\": \"\u003Cwhen to use this>\", \"priority\": 1 }\n]\n```\n\nThe editor labels the text **\"Instructions (Optional)\"**, and it is optional: with no text the agent decides by\nitself when to use the KB (`mode: \"all\"`, one such trigger, ≤ 7 KBs); with text the trigger is `mode: \"custom\"`\n(≤ 5 KBs). The editor caps the text at 1000 characters, but the server stored 1001. The **server** allows at most 4\ntriggers (`\"knowledgeBaseTriggers.4.priority must not be greater than 4\"`). A trial chat on a bot with a KB answered\nfrom that KB (live-proven 2026-09-26). An older version of this skill said the text was required, 10–500\ncharacters, labelled \"When to use this knowledge base\": that was wrong.\n\n**4. Gaps are a dated log of misses, not an inventory of what is missing now.**\n🔴 A row stays `open` after the answering content is added — proven by differential (documents\nadded, the questions re-asked live and answered from them, the list re-read unchanged, `lastAskedAt`\nunmoved) — and a question the KB answers is never logged, so every row looks current until you\nread its dates. List: `GET \u002Fknowledge-base\u002Fgaps?locationId=&knowledgeBaseId=&status=open` (AI rail).\n1. **Read the knowledge base first** — content present ⇒ the row is stale; do not change the agent.\n2. **Check `lastAskedAt`** against the window you are judging.\n3. **Write the knowledge to match `topQueryTexts`** (the customer's own wording).\n4. **Never filter by `categories`** — a genuine product question was filed under *Noise \u002F Gibberish \u002F Chitchat*.\nThe DISMISS write is **`PATCH \u002Fknowledge-base\u002Fgaps\u002F{gapId}\u002Fstate`** `{locationId, state:\"archived\"}` —\nlive-proven 2026-09-08, reversible with `state:\"active\"`. The write says `active`, the read says `open`. Fields and counts endpoint: `knowledge\u002Fcorpus\u002Fai-agents\u002F20-api\u002Fknowledge-base.md` → \"Gaps\".\n\n**5. Editing a rich-text doc is a PUT, not delete-and-recreate.** `PUT\n\u002Fknowledge-base\u002Frich-text\u002F:id` is a **live-verified full-replace** (2026-08-28, the designated\ntest sub-account): same body shape as create — `{locationId, knowledgeBaseId, title, content}` —\n200, response carries `status: \"training\"`, and a read-back of the sent `content` came back\n**byte-identical**. It re-chunks and re-embeds automatically, exactly like create: poll `GET\n\u002Fknowledge-base\u002Frich-text\u002F:id\u002Fstatus` until `\"trained\"` (a full retrain took ~4.6s in the proving\nrun) — there is no separate retrain call.\n\nDo not delete-and-recreate a doc to edit it, and do not add a second doc alongside the old one\nas a workaround: both orphan or duplicate content — a delete drops any id another object\nreferences, and a second doc leaves stale content the agent can still draw from, which is a\nworse failure than the one you were trying to fix. PUT the existing id instead.\n\n`content` is HTML; the server derives `contentMarkdown` from it. **A direct `contentMarkdown`\nwrite 200s and changes nothing** — measured in both a contentMarkdown-only body and an\nunchanged-content-plus-contentMarkdown body. `kb-compiler.mjs`'s `compileRichTextUpdate(id, doc,\n{locationId})` builds this PUT descriptor, and throws rather than silently no-opping if the\ncaller's IR carries a `contentMarkdown` key — a caller supplying that key believes something\nfalse about this API.\n\n**6. The retrieval test needs the editor's body.** `POST \u002Fknowledge-base\u002Fchunks` `{query, knowledgeBaseIds,\nlocationId, rerankChunks: true, threshold: 0, topK: 5, subCaller}` returns up to 5 chunks with `score` and\n`category` (`bulk` = crawled page, `finetuned` = FAQ, `rich_text`, `files`, `table`). Without those parameters it\nreturned one chunk, which looks like \"the KB only knows one thing\". The editor's ⌘K \"Test retrieval\" drawer keeps\na history: `POST \u002Fknowledge-base\u002Fretrieval-test\u002Fconversations\u002Fentries` `{locationId, knowledgeBaseId, query,\nagentLayer, chunks, totalChunks, latencyMs, idempotencyKey}` → `{conversationId, entryId, isNew, isDuplicate}`;\n`GET …\u002Fretrieval-test\u002Fconversations?locationId=&knowledgeBaseId=` lists them, `GET …\u002Fconversations\u002F{id}?locationId=`\nreads one, `DELETE` the same removes it (then 404 \"Conversation not found\").\n\n**7. The crawler judges what the CDN serves, refuses thin pages, and obeys robots.txt.** Status `Restricted` with `error:\n\"Insufficient content (174 chars)\"`. A page you just edited can be judged on its stale cached copy. 🔴 The crawler reads\nrobots.txt as user agent **`AITrainingBot`**. A blocked page still gets **201 `status: \"Processing\"`** on the POST, then\n`status: \"Restricted\"`, `error.code: \"ROBOTS_TXT_BLOCKED\"`, `retryable: false`, and no URL is stored. GHL's own\n`app.gohighlevel.com\u002Fv2\u002Fpreview\u002F{pageId}` URLs are blocked this way (proven live 2026-09-29). **A 201 is not a crawl:**\nread `GET \u002Fknowledge-base\u002Fcrawler\u002Fstatus?locationId=&knowledgeBaseId=` for `operationDetails.status` and `.error`, and\npoint the user to Rich Text when robots.txt blocks the page.\n\n## Sources — what ran live (2026-09-26 \u002F 28)\n\nEach was built in an own test KB, read back on separate requests, retrieved through `\u002Fchunks`, and deleted by exact\nid. Full shapes: corpus `ai-agents\u002F20-api\u002Fknowledge-base.md` → \"Sources executed live\".\n\n| Source | Create | Read | Remove | Traps |\n|---|---|---|---|---|\n| FAQ | `POST \u002Fknowledge-base\u002Ffaqs` `{locationId, knowledgeBaseId, question, answer, metadata?}` → `{faq{…, id}}` | `GET \u002Fknowledge-base\u002Ffaqs?locationId=&knowledgeBaseIds=&limit=&offset=` → `{faqs, count, hasMore}` | — | a duplicate question → **409** \"A FAQ with this question already exists in this knowledge base\"; the 1000-character answer cap is client-only (1001 stored) |\n| Rich text | `POST \u002Fknowledge-base\u002Frich-text\u002F` `{locationId, knowledgeBaseId, title, content (HTML)}` → `status:\"training\"` | `GET \u002Fknowledge-base\u002Frich-text\u002F{id}\u002Fstatus?locationId=` → `{sourceType, status, stages[CHUNKING, EMBEDDING]{state, durationMs}}` | `DELETE \u002Fknowledge-base\u002Frich-text\u002F{id}` | trained in ~3 s; the server adds generated questions to the chunk; edit with PUT (Trap 5); the editor also offers bulk delete (`\u002Frich-text\u002Fbulk`) and a per-document cancel (`\u002Frich-text\u002F{id}\u002Fcancel`), neither executed |\n| Web crawler | `POST \u002Fknowledge-base\u002Fcrawler` `{locationId, url, option: \"Exact\" \\| \"Path\" \\| \"Domain\", knowledgeBaseId}` → `{operationId, status:\"Processing\"}` | `GET \u002Fknowledge-base\u002Fcrawler?locationId=&knowledgeBaseId=` → `{count, urls[]}`; `GET …\u002Fcrawler\u002Fstatus` = the **latest** operation only; `GET …\u002Fcrawler\u002Fpages\u002Fcontent?locationId=&urlId=&knowledgeBaseId=` → the scraped text | `DELETE \u002Fknowledge-base\u002Fcrawler` `{locationId, urlIds, deleteAll:false, knowledgeBaseId}` → `{deletedCount}` | Exact mode **trains with no separate train call**; `sitemap-preview` refuses Exact (400); the URL refresh `PUT \u002Fknowledge-base\u002Fcrawler?locationId=&knowledgeBaseId=` `{urlIds, refreshAll, urlIdsToSkip}` answered 200 and changed nothing within 8 s, so its effect is unproven; Path \u002F Domain, cancel and retry are unexercised |\n| URL auto-refresh | `POST \u002Fknowledge-base\u002Fcrawler\u002Fautomation` `{knowledgeBaseId, refreshFrequency: DAILY \\| WEEKLY \\| MONTHLY}` → an ACTIVE `WEBSITE_RETRAIN` task, first run a week out for WEEKLY | `GET …\u002Fcrawler\u002Fautomation\u002Flatest?knowledgeBaseId=` | `DELETE …\u002Fcrawler\u002Fautomation\u002F{id}` (then `latest` → `data:null`) | the update `PUT …\u002Fautomation\u002F{id}` needs `taskId` = the automation's own id (422 \"taskId must be a string\" without it) |\n| File | **multipart** `POST \u002Fknowledge-base\u002Ffiles` | `GET \u002Fknowledge-base\u002Ffiles\u002F{id}`; `…\u002F{id}\u002Fstatus` stages CONVERSION → EXTRACTION → CHUNKING → EMBEDDING | `DELETE \u002Fknowledge-base\u002Ffiles\u002F{id}` (then 400 \"Invalid file ID\") | ≤ 10 MB, `.pdf .doc .docx .md`; trained in ~3 s |\n| Table (CSV) | **multipart** `POST \u002Fknowledge-base\u002Ftable\u002Flocation\u002F{loc}\u002Fkb\u002F{kb}\u002Fupload`, then `POST …\u002F{fileId}\u002Fselect-columns` `{selectedColumns[{name, originalType, selectedType, isSelected, sampleValue, displayName, searchable, required}]}` | `GET …\u002F{fileId}`, `…\u002Fstatus` (ANALYZING → PROCESSING → INDEXING), `…\u002Fdata?page=&limit=` | `DELETE …\u002F{fileId}` (then 404) | ≤ 50 MB; trained in ~48 s |\n\n🔴 **Files and tables are multipart uploads.** `raw_request` sends JSON only, so neither can be created through\nthis plugin today: point the user to the editor's Files \u002F Tables tab.\n\nThe scraped text of a crawled page is edited through View scraped content → Save:\n`PUT \u002Fknowledge-base\u002Fcrawler\u002Fpages\u002Fcontent?locationId=&urlId=&knowledgeBaseId=` `{content}` → 200 with an empty body. The\nrow retrains and reads back `status: \"Successful\"`, **`contentEditedByUser: true`**, and retrieval returns the edited text\n(proven live 2026-09-29). Retraining the URL from the website reverts the edit (the editor's warning). The URL list's\n`content` field is a storage URL to the text file, not the text itself. The `\u002Fconversations-ai\u002Ftrain\u002Fchunk\u002Fcontent`\nroute older notes gave is not called by the Knowledge Base app.\n\n**Gaps**: `GET \u002Fknowledge-base\u002Fgaps\u002Fcounts?locationId=` → `{data[{knowledgeBaseId, gapCount}]}` for the whole\nlocation; `GET \u002Fknowledge-base\u002Fgaps?locationId=&knowledgeBaseId=&status=open&page=&perPage=` →\n`{gaps[], stats{openGaps, queriesAffected, resolvedThisMonth, archivedGaps}, total}`. The editor's tabs are Needs\nanswers · Answered · Archived. `POST \u002Fknowledge-base\u002Fgaps\u002F{gapId}\u002Ffill {title, answer, locationId}` writes into the\nKB itself and has not been executed.\n\n## What GHL can do that this plugin cannot reach on the test account\n\nIf the user wants one of these, **GHL can do it** — say so and point to the screen.\n\n| Capability | Where in GHL | Why not here |\n|---|---|---|\n| Import files from Google Drive, daily Drive sync | KB → Files → \"Import from Google Drive\" (flag `knowledgeBase.googleSheets`) | needs a connected Google account (`GET \u002Fknowledge-base\u002Foauth?locationId=&type=google`); the only one on the test account is a person's own, so it was not used |\n| Google Sheet as a table, auto \u002F scheduled sync | KB → Tables → \"Import from Google Sheets\" | same |\n| Internal Data (sync the sub-account's Products) | KB source behind flag `knowledgeBase.internalData` | `GET \u002Fknowledge-base\u002Finternal-data\u002Fsources` answered 403 \"Forbidden resource\" on the test account |\n| Web search as a source | no screen: the editor hard-codes it off | the client ships the module, but `GET \u002Fknowledge-base\u002Fweb-search\u002Fkb\u002F{id}` answered 404 \"Cannot GET\" |\n| Create with AI (Beta) | KB list → \"Create with AI\" | it opens the Ask AI copilot (\"Create a knowledge base from brand voice\") and makes no KB call of its own |\n| Open the KB screen at all | AI Agents → Knowledge Base | the page sits behind the `conversation_AI` billing opt-in (`config.optIn`), not `enabled` |\n\n## Writing FAQs from real conversations\n\nWhen the job is to fill a knowledge base from what customers already asked, keep only the questions a customer asked to\nget facts (products, services, prices, hours, policies, how things work). Drop greetings, bookings, one-word replies,\nstaff-only notes and system lines. Take the question from the customer's side and the answer from the business's side.\nRemove anything personal (names, numbers, addresses, billing details) and anything only true on one day (\"closed today\",\na slot time). Rewrite colloquial questions so each one stands alone and names its product. Answers the bot gave that\nsaid it did not know are the most useful candidates: complete them where the facts are known. Where a value is unknown\nand changes (a price, hours, a phone number), write a clear placeholder such as [price] instead of inventing it or leaving\nthe answer empty. Merge questions that ask the same thing into one pair, then save them in one bulk call and read back\nhow many were created, updated or failed.\n\n## Limits\n\nFiles ≤ 10 MB (`.pdf .doc .docx .md`) · table CSV ≤ 50 MB · rich text ≤ 25,000 characters and 512 KiB of HTML\n(`richTextBlockedByLimit` fires on the character cap, not on a document count) · FAQ answer ≤ 1000 (client-only) ·\nKB name ≤ 50 (client-only) and description ≤ 175 · ≤ 7,000 crawled pages per KB · 15 KBs per sub-account · ≤ 4\nknowledge-base triggers per bot (server) · at least one KB must be selected where a bot requires one.\n\n## Proof status — read before trusting a write\n\n**Live-proven** (executed on a test sub-account and read back on a separate request): the KB record's create, read,\nlist, rename, delete and associated-entities read; FAQ, rich text (create, full-replace PUT, delete), web crawler\n(Exact mode, delete, auto-refresh create\u002Fupdate\u002Fdelete), file and table sources (uploaded from the editor, then\nread and deleted through the API); retrieval and its history; the gap list and the gap archive\u002Frestore write.\nAlso live-proven (2026-09-29): the scraped-text edit, and a robots.txt-blocked crawl's status and error. **Not proven:** the\ncrawl refresh's effect; Path \u002F Domain crawls; crawl cancel \u002F retry (an Exact crawl of one page finishes in about 2 s, and\nanything longer needs a multi-page crawl); gap fill; and everything in the table above. Treat a first write of an unproven operation as a throwaway validation run\non a test sub-account.\n\n## Scope\n\nKnowledge bases only. The agents that consume them are `ghl-voice-ai` (internal) and\nConversation AI (lists its bots and configures them on the internal rail — see `ghl-conversation-ai`).\n",{"name":5,"description":7},"main","35d05921ecd17673183702dca9c1bb7d2c961f71","plugins\u002Fuxie-ghl-factory\u002Fskills\u002Fghl-knowledge-base",5,"https:\u002F\u002Fgithub.com\u002Fuxieee\u002Fuxie-ghl-factory","https:\u002F\u002Fgithub.com\u002Fuxieee","2026-11-13 05:11:02",{"source":8,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add uxieee\u002Fuxie-ghl-factory --skill ghl-knowledge-base",{"tool":37,"command":38},"bunx","bunx skills add uxieee\u002Fuxie-ghl-factory --skill ghl-knowledge-base",{"tool":40,"command":41},"pnpm","pnpm dlx skills add uxieee\u002Fuxie-ghl-factory --skill ghl-knowledge-base","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fuxieee\u002Fuxie-ghl-factory\u002Fblob\u002Fmain\u002Fplugins\u002Fuxie-ghl-factory\u002Fskills\u002Fghl-knowledge-base\u002FSKILL.md",true,[],[46,53,60,67,74,81],{"skillNo":47,"name":48,"nameZh":6,"description":49,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":50,"categories":51,"keywords":52,"publishedAt":19},"01M78F5XMBPT6K5KN5H7JWWYRH","ghl-workflow-specialist","Senior GoHighLevel workflow architect — designs automations (nurture, speed-to-lead, no-show recovery, internal ops, data hygiene, routing) grounded in the full trigger\u002Faction catalog and known anti-patterns, and reviews or fixes live workflows. Use when the user asks what automation to set up for X, wants a workflow designed, or asks why a live workflow is not getting through, is failing, or whether an A\u002FB test is done. Recons + reads the client brief before proposing anything. Once approved, the build itself goes to create-ghl-workflow.","https:\u002F\u002Fgithub.com\u002Fuxieee\u002Fuxie-ghl-factory\u002Fblob\u002Fmain\u002Fplugins\u002Fuxie-ghl-factory\u002Fskills\u002Fghl-workflow-specialist\u002FSKILL.md",[],[],{"skillNo":54,"name":55,"nameZh":6,"description":56,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":57,"categories":58,"keywords":59,"publishedAt":19},"01M78F5PXJXRMGCK95MAPVBHAJ","ghl-workflow-fast-forward","Fast-forward contacts parked at a GoHighLevel \u002F HighLevel workflow WAIT step to the next step — the builder's \"Action statistics → move to next step\" control, driven via the internal API. Use to accelerate end-to-end workflow testing (drive a multi-day wait ladder to completion in minutes instead of waiting real time), to unstick contacts stranded at a wait, or when the user says \"fast-forward this wait\", \"skip the wait\", \"move parked\u002Fstuck contacts to the next step\", \"requeue stuck contacts\", \"test this long delay without waiting days\", or gives an app.gohighlevel.com workflow URL and asks to push contacts past a wait. This MUTATES live enrollments (it fires whatever steps come next, i.e. real SMS\u002Femail to real people) — it is a WRITE skill and follows the plugin's write-rails. It is the mutating counterpart to the read-only get-ghl-workflow-logs.","https:\u002F\u002Fgithub.com\u002Fuxieee\u002Fuxie-ghl-factory\u002Fblob\u002Fmain\u002Fplugins\u002Fuxie-ghl-factory\u002Fskills\u002Fghl-workflow-fast-forward\u002FSKILL.md",[],[],{"skillNo":61,"name":62,"nameZh":6,"description":63,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":64,"categories":65,"keywords":66,"publishedAt":19},"01M78F5GZERF2S6CMHXWQYQEHZ","ghl-voice-ai","Build and configure GoHighLevel Voice AI phone agents — the prompt, the voice, the actions (appointment booking, call\u002Fagent transfer, contact-field extraction, SMS, workflows, custom API), knowledge bases, MCP servers, working hours, and outbound consent. Use when the user says 'set up a voice agent', 'build a phone agent', 'my AI should answer calls', 'book appointments over the phone', 'outbound calling', 'the agent isn't transferring calls', 'change the voice', or names a Voice AI agent. Internal API — the public rail exposes only how the agent sounds, not what it does.","https:\u002F\u002Fgithub.com\u002Fuxieee\u002Fuxie-ghl-factory\u002Fblob\u002Fmain\u002Fplugins\u002Fuxie-ghl-factory\u002Fskills\u002Fghl-voice-ai\u002FSKILL.md",[],[],{"skillNo":68,"name":69,"nameZh":6,"description":70,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":71,"categories":72,"keywords":73,"publishedAt":19},"01M78F5A7S9450G94K3Z7P0HWK","ghl-system-conventions","House conventions for building GoHighLevel systems — workflow naming and structure, tag\u002Ffield\u002Fcustom-value discipline, pipeline and opportunity decisions, copy rules, and the pre-build approval document the operator signs off before anything gets created. Account-agnostic, for ANY GHL sub-account. Use whenever designing, building, editing, reviewing, or discussing GHL workflows, automations, pipelines, stages, opportunities, tags, custom fields, custom values, calendars, or AI agents — even if the user never says \"conventions\" or \"standards\". Load it BEFORE proposing a design or creating any GHL object, so the proposal already looks the way the operator wants it.","https:\u002F\u002Fgithub.com\u002Fuxieee\u002Fuxie-ghl-factory\u002Fblob\u002Fmain\u002Fplugins\u002Fuxie-ghl-factory\u002Fskills\u002Fghl-system-conventions\u002FSKILL.md",[],[],{"skillNo":75,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":19},"01M78F54ZQZ21VS3JQJ21QVGWS","ghl-snapshots","GoHighLevel snapshots — capturing a sub-account's assets into a reusable snapshot, refreshing one, curating what it carries, and checking load conflicts. Use when a task involves creating or updating a snapshot, deciding what an account template should include, asking why a snapshot came out empty or is stuck processing, or planning a rollout onto other sub-accounts. Read it BEFORE calling any \u002Fsnapshots path by raw_request — two of the write endpoints silently do the wrong thing.","https:\u002F\u002Fgithub.com\u002Fuxieee\u002Fuxie-ghl-factory\u002Fblob\u002Fmain\u002Fplugins\u002Fuxie-ghl-factory\u002Fskills\u002Fghl-snapshots\u002FSKILL.md",[],[],{"skillNo":82,"name":83,"nameZh":6,"description":84,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":85,"categories":86,"keywords":87,"publishedAt":19},"01M78F4Y7ZHHJ1DMV11TC1CPVW","ghl-reverse-engineering","Map a GoHighLevel internal surface exhaustively — enumerate every screen, panel, control and save path a feature has, capture the real request behind each one, prove what each field does by differential, and write the result into the corpus. Use when the user wants to understand how a GHL feature works under the hood, extend an engine to an object the public API can't reach, re-sweep a surface that already has partial coverage, or asks to reverse-engineer \u002F sniff \u002F capture \u002F trace anything internal. GHL permits inspecting your own account's traffic. This is a MAPPING skill, not a lookup: it finishes when the surface is exhausted, not when the first endpoint is found.","https:\u002F\u002Fgithub.com\u002Fuxieee\u002Fuxie-ghl-factory\u002Fblob\u002Fmain\u002Fplugins\u002Fuxie-ghl-factory\u002Fskills\u002Fghl-reverse-engineering\u002FSKILL.md",[],[]]