How do you install n8n-skills, and why does an AI agent need skills for n8n?

What are n8n-skills, and what is inside a skill?

n8n-skills is a repository of 14 skills for AI agents, plus one skill router and a hooks layer, that together teach an agent to build n8n workflows through the n8n-mcp server. n8n-mcp on its own gives the agent tools: node documentation, validation, operations on an instance. Skills add the knowledge of how to use those tools well, so the workflow works on the first try instead of after three rounds of fixes.

A skill in this repository is a SKILL.md file with a short description at the top telling the agent when to activate it, and content that teaches one thing well. I opened n8n-error-handling as an example: it explains that by default, a failing node halts the whole workflow, and that for a webhook or a scheduled job that default is wrong. It teaches the two-step error output setup, onError: continueErrorOutput plus a wired connection to a handler, and shows what happens if you do only one of those two steps: the workflow either silently drops the error, or the handler never fires.

Fourteen skills each cover a separate area, and each one teaches a single discipline well:

Skill	What it teaches
n8n-expression-syntax	{{ }} syntax, $json/$node access, the $json.body gotcha on webhooks
n8n-mcp-tools-expert	picking the right n8n-mcp tool for the task, nodeType formats, validation profiles
n8n-workflow-patterns	five proven architecture patterns: webhook, API, database, AI, scheduled
n8n-validation-expert	reading validation errors, telling must-fix apart from quality advice
n8n-node-configuration	required fields per operation, dependencies between node properties
n8n-code-javascript	JavaScript in the Code node, data access, this.helpers, loop patterns
n8n-code-python	Python in the Code node and its limits, no external libraries
n8n-code-tool	the Code Tool an AI agent calls, a different contract from the plain Code node
n8n-error-handling	per-node error output, retries, Error Trigger workflows, 4xx/5xx codes
n8n-binary-and-data	working with files and images, $binary versus $json, the agent-tool boundary
n8n-subworkflows	extracting shared logic, typed inputs, all versus each execution mode
n8n-agents	building the AI Agent node, $fromAI, memory, structured output, RAG
n8n-multi-instance	working across several n8n instances at once, avoiding silent address mistakes
n8n-self-hosting	deploying n8n on a VPS with Docker Compose and Caddy, single versus queue mode

The skill router, using-n8n-mcp-skills, loads at session start and points the agent to the right skill instead of relying on a match against the query alone.

Why does an agent build worse without them?

Without skills, an agent only has the raw n8n-mcp tools and whatever it remembers from training. n8n changes faster than any model's training keeps up with: parameter names, node versions, default behaviors. An agent that relies on memory instead of live documentation writes a parameter that looks right, passes validation as a plain string, and does nothing at runtime.

The classic mistake the skills describe is putting the agent's working instructions into the user message instead of the system message of the AI Agent node. It looks like a small detail, but on later turns of a conversation it leads to hallucinations: the model treats the instruction as content to process rather than a standing rule. The n8n-agents skill teaches that tool names and descriptions in the AI Agent node are themselves the prompt, and that memory, structured output and chat topology each carry traps that don't show up in the node's documentation.

Other recurring mistakes without skills: the wrong nodeType shape in MCP calls, one shape for get_node, a different one for validate_workflow, a Set node that feeds only one downstream node instead of writing the expression directly at the consumer, or validation that passes while nobody checked the connections after an edit, so the break surfaces only in production. n8n-mcp telemetry shows the scale of the problem from another angle: 74% of AI-built workflows contain a Code node, 60% an HTTP Request node, and only 2% have an Error Trigger. The agent reaches for code and a raw API call instead of looking for a ready-made integration, because searching a catalog of 525+ nodes needs the search_nodes tool, and n8n-mcp-tools-expert is the skill that teaches when and how to use it. The agent builds the happy path because that is what it was asked for, and skills only partly fix that; the rest is a job for a person or a rule in the agent's memory.

How do you install them in Claude Code?

The simplest route is the plugin install, because it also gives you the hooks layer, not just the skill files.

/plugin install czlonkowski/n8n-skills

An alternative is adding the repository as a marketplace, then picking the plugin from a list:

/plugin marketplace add czlonkowski/n8n-skills
/plugin install
# select "n8n-mcp-skills" from the list

A third route, manual, works anywhere but skips the hooks:

git clone https://github.com/czlonkowski/n8n-skills.git
cp -r n8n-skills/skills/* ~/.claude/skills/

Once the files are copied, Claude Code loads the skills automatically, no restart needed. The plugin install has one advantage worth knowing about: the skill router using-n8n-mcp-skills loads through a SessionStart hook at the start of every session, and comes back after a resume, a clear, or a compact, so it survives context compression. With a manual install, the router activates only when a query matches its description, like any other skill.

How do you install them in Claude.ai and ChatGPT?

In Claude.ai you don't install from the repository with one command; you upload files. Download a skill's folder from the skills/ directory in the repository, zip it, or grab prebuilt zips from the latest GitHub release, and upload it under Settings → Capabilities → Skills. Repeat for each skill separately, since Claude.ai has no concept of a "skill pack", only individually uploaded files.

In ChatGPT the install works similarly, but in a different part of the interface than most people check first. You upload skills under Settings → Skills, not Plugins, and pick every entry prefixed n8n-. That distinction matters, because Plugins in ChatGPT is a different thing, integrations with external services, while Skills is its own mechanism for uploaded instructions.

In both places, Claude.ai and ChatGPT, skills activate purely by the description in the file's header: the model decides on its own when a given skill matters for the query. That is the same activation rule as in Claude Code, but without the hooks layer, so nothing reminds the agent mid-task to check a given skill before it calls a specific n8n-mcp tool.

How do you install the Codex plugin, and what do the hooks do?

The n8n-skills repository also works as an agent plugin for Codex, alongside Claude Code. The install follows the same path as any client that supports agent plugins: add the repository as a plugin source and install n8n-mcp-skills from the list. In Codex, as in Claude Code, you get the hooks layer along with the skills, because both environments support the mechanism.

Hooks are three kinds of automatic reminders that don't wait for the agent to stumble onto the right skill description on its own:

• SessionStart loads the skill router using-n8n-mcp-skills at the start of every session, so the agent knows the full skill index and the rules from the first turn, instead of waiting to match a description.
• PreToolUse fires before high-impact tool calls: calling get_node on a Set, Code, Merge, Loop Over Items, DateTime, Data Table, or AI Agent node triggers a reminder specific to that node, and fires again on every repeated lookup of the same node, because a re-lookup usually means the agent is reconsidering the same decision. Calls to n8n_instances and n8n_manage_credentials trigger a one-shot reminder about multi-instance discipline and credential handling.
• PostToolUse fires after validate_workflow: it inspects the workflow's node types and routes the agent to the skills that own the remaining risks, with the reminder that passing validation is necessary, not sufficient.

Every hook runs fail-open: if something goes wrong in the reminder mechanism itself, it never blocks the tool call. In Claude.ai and ChatGPT none of these three mechanisms exist, because hooks need a client environment that supports them.

How do you keep them updated in a team?

Skills don't update themselves. This is worth knowing before installing them for more than one person: the repository on GitHub moves forward, but your copy, whether in ~/.claude/skills/ or as zips uploaded to Claude.ai and ChatGPT, stays on whatever version you uploaded.

In the teams I work with, one person maintains the skills centrally for the whole workspace. For a manual install, that means a git pull in the repository and re-copying the files into the skills directory. For Claude.ai and ChatGPT, it means downloading the new zips from a release and replacing the old versions in the skills panel, for each user separately, since skills uploaded by one person do not propagate automatically to the rest of the team. The update works by replace, not merge: the new version of a file simply overwrites the old one.

This model works well when the person maintaining skills centrally understands what changed in a given version and tells the rest of the team before swapping the files for everyone at once. Without that, the team ends up running different skill versions in parallel, which is not disastrous on its own, but makes it harder to debug why two people get different results on the same query.

What changes in practice after installing them?

At an e-commerce client I work with, the team builds workflows through ChatGPT and Claude with the n8n skills installed, in weekly sessions. One team member described the difference after installing them as "night and day": the same query to the agent, the same n8n-mcp server underneath, but with the skills layer the agent sticks to proven patterns instead of guessing from scratch each time how to configure a node or which architecture pattern fits the task.

Several things change at once, in concrete terms. The agent falls into fewer validation-error loops, because n8n-validation-expert teaches which errors must be fixed and which are just quality recommendations. It builds fewer Set nodes that feed a single consumer, because n8n-expression-syntax teaches writing the expression directly instead. It puts the agent's instructions in the wrong field of the AI Agent node less often. None of these changes is dramatic on its own, but together they mean fewer rounds of fixes between the brief and a working workflow.

Where to start?

If you work in Claude Code, start with the plugin install: one command, /plugin install czlonkowski/n8n-skills, gives you all fourteen skills together with the hooks layer. If you build workflows in Claude.ai or ChatGPT, upload the router first, plus two or three skills closest to what you do most often, for example n8n-mcp-tools-expert and n8n-workflow-patterns, rather than all fourteen at once, and add more as you need them.

The repository is free and open source, at github.com/czlonkowski/n8n-skills, and also listed on skills.sh. It needs the n8n-mcp server installed and configured: without it, the skills teach patterns, but the agent has no tools to act on them against a live instance.