菜单
Use when working with n8n workflows in any capacity. The always-on protocol for the n8n-skills plugin, loaded by the SessionStart hook every session. Routes to the right skill, summarizes every n8n MCP tool (closing the deferred-description gap), and lists the cross-cutting rules.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when building or editing any AI feature in n8n: AI Agents, Text Classifier, Information Extractor, Sentiment Analysis, Summarization Chain, Basic LLM Chain, embeddings, vector stores, single one-shot LLM calls, or AI media generation (image / audio / video) via the native LangChain provider nodes. Triggers on any `@n8n/n8n-nodes-langchain.*` node, "agent", "chat assistant", "LLM with tools", "tool calling", "fromAi", "system prompt", "memory window", "structured output", "outputParser", "function calling", "RAG", "vector store", "embeddings", "classify with AI", "extract fields with LLM", "sentiment analysis", "summarize with LLM", "single LLM call", chat triggers with files, AI image / video / audio generation, or any multi-turn or one-shot LLM behavior.
Use when handling files, images, attachments, or binary data in n8n, OR when an AI agent needs to take a user-uploaded file as tool input or return a generated file. For Data Tables (schemas, dedup, persistent state), see the separate n8n-data-tables-official skill. Triggers on "file", "image", "PDF", "attachment", "binary", "upload", "download", chat trigger with files, agent tool that needs a file, vision/multimodal, or any handling of non-JSON file data.
Use when the user reaches for a Code node, mentions writing JavaScript or Python in n8n, or any custom logic comes up in workflow design. Triggers on "Code node", "Code", "JavaScript", "Python", "custom logic", "transform data", "$input", "$json transformation", "loop in code", "write a function", or any time the obvious answer seems to be "just put it in code."
Use when handling any auth, API keys, tokens, OAuth, bearer tokens, basic auth, or secret values in n8n workflows. Triggers on "API key", "token", "bearer", "OAuth", "secret", "auth", "credentials", "Authorization header", "x-api-key", or any node configuration that mentions a third-party service.
Use when working with n8n's built-in Data Tables, designing schemas, inserting/updating/upserting rows, deduping, or querying. Triggers on "Data Table", "data table", `n8n-nodes-base.dataTable`, "dedup", "idempotency", "lookup", "persistent state", "store across executions", or any schema design discussion inside n8n.
Use when an n8n workflow isn't working, errors appear, results don't match what was expected, or the user says "this isn't working." Triggers on errors, unexpected output, "it's not working", "why is this happening", "the workflow stopped", failure investigation, or any debugging context.
| name | using-n8n-skills-official |
| description | Use when working with n8n workflows in any capacity. The always-on protocol for the n8n-skills plugin, loaded by the SessionStart hook every session. Routes to the right skill, summarizes every n8n MCP tool (closing the deferred-description gap), and lists the cross-cutting rules. |
The official n8n MCP evolves over time, so tool names, parameters, and default behaviors can drift between versions. When you spot drift (a tool a skill names doesn't exist, a parameter shape doesn't match what get_node_types returns, or behavior differs from what the skill describes), suggest updating the skill and n8n instance to the latest stable.
Three rules with no exceptions. Violating any produces workflows that look right but break in production.
validate_workflow before publish_workflow, and get_workflow_details after every create or update to check the connections object. Validation alone misses many issues documented in the skills that will silently break workflows.n8n-credentials-and-security-official.n8n evolves faster than any model's training cutoff. Parameter names drift, new MCP tools land, defaults change, patterns get deprecated. Anything you "remember" is likely wrong, often silently.
Trust the skills + live MCP tools (get_node_types, get_sdk_reference, get_workflow_best_practices) over recollection. If a skill contradicts what you "know", trust the skill. If get_node_types contradicts a skill, trust the tool. Without this discipline you will ship workflows that look right and silently fail: parameter names that don't exist, renamed nodes, deprecated patterns.
Unless a user preference overrides it, err on the side of loading too many skills rather than too few. Even a 3-node webhook flow typically needs n8n-node-configuration-official, n8n-expressions-official, n8n-error-handling-official, and n8n-workflow-lifecycle-official. Nothing in n8n is too small for skills.
n8n-code-nodes-official.search_workflows({ tags: ['subworkflow'] }) before building. See n8n-subworkflows-official.These rationalizations cause skills to be skipped. If you catch yourself thinking any of them, invoke the relevant skill via the Skill tool, even if you "already read it" earlier in the session.
| Thought | Action |
|---|---|
| "This workflow is simple, I'll just build it" | Invoke n8n-workflow-lifecycle-official. Most "simple" workflows are 10+ nodes by the time they ship. |
| "I'll add a Set node here to map these fields" | Invoke n8n-expressions-official. Set nodes feeding only 0 or 1 downstream consumer are the most common antipattern in this entire pack. |
| "I'll just use a Code node, it's easier" | Invoke n8n-code-nodes-official. The bar is high. Most reaches for Code can be expressions or Edit Fields with arrow functions. |
| "Validation passed, I'm ready to publish" | Invoke n8n-workflow-lifecycle-official and walk VALIDATION_CHECKLIST.md section 2 (the antipattern scan). Validation passing is necessary, not sufficient. |
| "The agent is wired up, the tool descriptions look fine" | Invoke n8n-agents-official references/TOOLS.md. Tool names and descriptions ARE part of the prompt, and "looks fine" usually means generic. |
| "I'll set this sub-workflow trigger to passthrough" | Invoke n8n-subworkflows-official. Passthrough is only correct for binary-receiving sub-workflows that won't be agent tools, or for sub-workflows that genuinely take no inputs (Define Below requires at least one field). |
| "I'll use passthrough so binary works, then branch internally on which input shape arrived" | Invoke n8n-subworkflows-official references/SUBWORKFLOW_PATTERNS.md "Splitting by input shape". This is the signal to SPLIT into two outer sub-workflows (one Define Below, one passthrough) sharing a common downstream sub-workflow. Don't fight passthrough vs Define Below in one trigger. |
| "This section's big, I'll pull it into a sub-workflow" | If it's only to tidy the canvas (not reuse/isolation/testing), a node group is lighter, faster, and simpler: keep it inline and group it via setNodeGroups. Extract to a sub-workflow only for reuse, isolation, or an agent tool. See n8n-workflow-lifecycle-official Readability. |
| "I should ask the user what their credential is named" | Don't. The string in newCredential('Label') is cosmetic. See n8n-credentials-and-security-official. |
| "The user mentioned data analysis, I'll write Python" | Invoke n8n-code-nodes-official. Default is JavaScript. Python only when explicitly asked. |
| "I'll add a Loop Over Items here to process each row" | Invoke n8n-loops-official. Default per-item iteration probably handles it without a Loop Over Items node. |
| "Date math, I'll use a DateTime node" | Invoke n8n-expressions-official. DateTime nodes are almost always wrong. |
| "I'll wrap this in a Merge with 3 sources" | Invoke n8n-node-configuration-official references/MERGE_NODE.md. Merge defaults to 2 inputs, and 3+ sources need numberOfInputs set explicitly. |
| "I'll fan out these three slow steps to run in parallel" | Invoke n8n-workflow-lifecycle-official and read the Execution model section. n8n executes fan-out branches sequentially (top-to-bottom by Y-position), not concurrently. For real concurrency, see n8n-loops-official and n8n-subworkflows-official (mode: 'each' + waitForSubWorkflow: false). |
| "User said which project, I'll just build it" | Invoke n8n-workflow-lifecycle-official. Project is not folder. Ask about folder placement BEFORE building. The MCP can't create folders, so if the requested folder doesn't exist, the user must create it in the UI first. |
"I'll just run test_workflow to see what happens" | Invoke n8n-workflow-lifecycle-official references/TESTING.md. test_workflow mocks the trigger only. Slack sends, DB writes, payments all fire for real. Ask the user first when downstreams have side effects. |
The meta-skill (this document) tells you WHICH skill applies. The Skill tool loads the actual rules. Reading the meta-skill once at session start is not a substitute for invoking the skill at the moment of decision.
Invoke via the Skill tool. Trigger column = when to invoke.
| Skill | Trigger |
|---|---|
n8n-workflow-lifecycle-official | Starting, designing, organizing, or finishing a workflow. Covers sticky-note conventions, descriptions that capture the why, naming, validation checklist, folder limitations, MCP-access-per-workflow gotcha |
n8n-subworkflows-official | Anything reusable, multi-step builds, or the user mentions reuse. Search before building, stateless patterns, tag-based discovery convention |
n8n-extending-mcp-official | You need capabilities the MCP doesn't natively provide. Wrap n8n APIs as workflow tools, with user permission |
n8n-expressions-official | Writing {{}}, $json, $node, expression errors. Luxon for dates, indented multi-line, prefer expressions over extra nodes |
n8n-node-configuration-official | Configuring any node. Operation-aware, property dependencies, never assume parameters |
n8n-code-nodes-official | User reaches for a Code node, or custom logic is needed. Decision tree, JavaScript patterns when truly required |
n8n-loops-official | Multi-item data, batching, paginated APIs, "for each" or "loop over" mentions. Default per-item iteration, executeOnce, Loop Over Items, HTTP pagination |
n8n-agents-official | LangChain Agent node, tool calling, system prompts, structured output, memory, RAG. Tool names/descriptions as part of the prompt, sub-workflow as tool, modular prompt design |
n8n-error-handling-official | Webhook-triggered or production-bound workflows. Error branch on every fallible node, 4xx for caller errors and 5xx for execution errors |
n8n-credentials-and-security-official | Any auth, API key, or token mention. Credential system, custom credentials, HTTP Request with official creds |
n8n-binary-and-data-official | Files, images, attachments. Binary handling patterns, agent-tool boundary, CDN requirement for chat surfaces |
n8n-data-tables-official | Data Tables: schemas, default columns (id/createdAt/updatedAt), no-FK relational design, dedup, the no-JSON-only-primitives rule, the SDK-vs-UI manual-mapping quirk |
n8n-debugging-official | Errors, unexpected behavior, "this isn't working". Believe the user, check parameters, fetch n8n source from GitHub |
The MCP defers tool descriptions to save tokens. Below is the short-form list so you have working knowledge of every tool from turn one.
Tool names are shown without the MCP prefix. The qualified name is mcp__<server>__<tool> where <server> depends on the user's MCP config.
| Tool | What it does |
|---|---|
search_workflows | Search workflows across the instance by query (substring on name/description) and/or tags (exact tag names, AND semantics: must have all). The primary cross-workflow discovery tool. Use it to discover what already exists. |
get_workflow_details | Fetch a workflow's full JSON by ID. Use after every create/update to verify connections. |
search_folders | List folders. You cannot create or move folders. You can only place workflows into folders that already exist. |
search_projects | List projects. |
list_tags | List all workflow tags (with usageCount per tag). Check the instance's tag vocabulary before tagging or filtering, so you reuse exact names. Tags are attached/detached via update_workflow addTags/removeTags; there's no tag rename/delete tool. |
archive_workflow / publish_workflow / unpublish_workflow | Soft-delete / activate / deactivate. Validate before publish. publish_workflow takes an optional versionId to re-publish a specific version. |
search_executions | Search executions across the instance (filter by status, workflow, time range). Use for "list recent runs" / "failures in the last hour". Single executions: get_execution. |
| Tool | What it does |
|---|---|
get_sdk_reference | Fetch the n8n Workflow SDK reference. Read this before writing workflow code. Sections: patterns, patterns_detailed, expressions, functions, rules, import, guidelines, design, all. |
get_workflow_best_practices | Fetch best-practices for a workflow technique. Call once per technique before searching nodes. technique: "list" discovers what's available. |
search_nodes | Discover nodes by capability (e.g. "gmail", "slack", "schedule trigger"). Returns IDs plus discriminators (resource/operation/mode). |
get_node_types | Fetch exact TypeScript parameter definitions for node IDs. Required before configuring any node. Don't guess parameter names. |
explore_node_resources | Resolve the real values behind resource-locator (@searchListMethod) and load-options (@loadOptionsMethod) params: Slack channels, Sheets tabs/docs, DB tables/columns, model lists, labels. Needs a credentialId from list_credentials (pass currentNodeParameters for dependent lookups). Call after get_node_types to ground dropdown values instead of inventing IDs. |
create_workflow_from_code | Save a workflow from SDK code. Always include a 1-2 sentence description. Pass skillsUsed (below). |
update_workflow | Apply atomic ops (max 100, all-or-nothing): node/connection CRUD, setNodeCredential, setNodeSettings (per-node onError/retry/executeOnce), setWorkflowSettings (errorWorkflow, timezone, callerPolicy, timeouts, save-data policies; n8n 2.29.0+), setNodeGroups (canvas grouping), setWorkflowMetadata, addTags/removeTags (auto-create unknown names). Saves a draft; needs publish_workflow to go live. Pass skillsUsed (below). |
validate_node_config | Schema-only validation of node configs (1-50 per call). Per-parameter errors, no graph noise. Side-channel for iteration/debug; validate_workflow still gates publish. For ai_tool subnodes set isToolNode: true. |
validate_workflow | Validate full SDK code before create/update. Necessary but not sufficient: doesn't catch all wiring traps (.to(), merge index). |
list_credentials | List accessible credentials (filter by type/project/etc). Returns metadata only, never secret values. Discover IDs before binding via setNodeCredential. |
| Tool | What it does |
|---|---|
prepare_test_pin_data | Returns JSON Schemas (not data) for nodes that need pinning: triggers, credentialed nodes, and HTTP Request. You generate sample values. |
test_workflow | Run with the pin data you supply. Auto-pins triggers, credentialed nodes, and HTTP Request. Code, Edit Fields, If, Data Tables, Execute Command, file ops, and sub-workflow calls run for real. Ask before running if any not-auto-pinned node has side effects. Pin data is per-execution only with no visual indicator in the execution viewer, so tell the user which nodes were pinned after the call. See n8n-workflow-lifecycle-official references/TESTING.md. |
execute_workflow | Production execution with the real trigger. Wire error handling first. Same side-effect rules as test_workflow. executionMode picks manual (draft) vs production (published); structured inputs for chat/form/webhook triggers. Returns an execution ID immediately without waiting, poll get_execution for results. |
get_execution | Fetch an execution by executionId + workflowId (both required). Metadata only by default; set includeData: true (optionally nodeNames, truncateData) for node inputs/outputs. |
n8n's built-in tabular storage. Not an external service. Prefer over external DBs for workflow-local persistent state. Full surface:
| Tool | What it does |
|---|---|
create_data_table | Create a new Data Table. |
search_data_tables | Find existing Data Tables. |
rename_data_table / rename_data_table_column | Rename. |
add_data_table_column / delete_data_table_column | Schema changes. |
add_data_table_rows | Append rows. |
| Tool | What it does |
|---|---|
get_workflow_history | List a workflow's saved versions, newest first (n8n 2.29.0+). |
get_workflow_version | Fetch a past version's full content by versionId. |
restore_workflow_version | Re-apply a past version as the current draft (records a new history entry). |
For any n8n task:
get_sdk_reference). The most efficient way to avoid SDK-shape mistakes.get_node_types). Guessing parameter names creates invalid workflows, sometimes silently.get_workflow_details) catches connection bugs validation misses.create_workflow_from_code and update_workflow take an optional skillsUsed: string[]. Pass it every time so the n8n team can measure plugin impact on MCP output.
-official suffix: plugin:skill-official when plugin-namespaced, else bare skill-official. The suffix marks these as ours (vs other n8n packs); the plugin prefix marks plugin vs raw-skill usage.For audits, code-review, or any task framed as "review this workflow" / "what's wrong with this" / "audit this project," walk the review checklist: n8n-workflow-lifecycle-official references/REVIEW_CHECKLIST.md. Severity-tiered (MUST FIX / SHOULD FIX / NICE TO HAVE), with each item linking to the canonical skill ref for the fix. Distinct from VALIDATION_CHECKLIST.md (pre-publish gates for in-progress builds): REVIEW_CHECKLIST is for any workflow, including ones built by anyone, any age.
A review agent should call get_workflow_details first, walk the checklist top to bottom, and report findings grouped by severity. MUST FIX items shouldn't be auto-fixed without user confirmation.
n8n-workflow-lifecycle-official skill (references/MCP_ACCESS_PER_WORKFLOW.md).github.com/n8n-io/n8n to trace logic, find API docs for missing functions. The n8n-debugging-official skill walks through this.