Design n8n Agent Workflow
Use this skill when n8n is the right visual, trigger-driven integration surface.
It does not turn n8n into a default durable multi-agent runtime.
Source Check
Workflow
- Map trigger, deterministic transforms/routes, external reads, model step,
draft output, and any external write separately.
- Select credentials/storage ownership, webhook exposure, retry policy,
idempotency key, concurrency limit, and failed-run record before connecting
an AI node.
- Treat the local model as one bounded node: prove structured output, tool
compatibility when used, no-call behavior, malformed result recovery, and
the exact model/server combination before exposing a write node.
- Keep source data minimised and trace/log retention explicit. Do not put
secrets in workflow exports or prompts.
- Leave write nodes disabled or draft-only until
auto-with-escalation names
the recipient, action, payload evidence, and approval event.
- Export a testable workflow fixture and verify happy path, duplicate event,
failed model response, credential failure, rejected approval, and recovery.
Output Shape
Return trigger, node graph, credentials, idempotency/retry behavior, model
contract, draft artifact, approval event, and recovery evidence.
Guardrails
- Prefer deterministic n8n nodes over agent planning whenever the route is
known.
- Do not expose a self-hosted instance publicly without an explicit deployment
and authentication decision.
- Do not claim a local model is reliable for tool use without a capability gate.
1---2name: design-n8n-agent-workflow3description: Design safe n8n workflows with deterministic routing, credentials, idempotency, recovery, local-model checks, drafts, and exact approval gates.4---56# Design n8n Agent Workflow78Use this skill when n8n is the right visual, trigger-driven integration surface.9It does not turn n8n into a default durable multi-agent runtime.1011## Source Check1213- n8n hosting: <https://docs.n8n.io/hosting/>14- n8n and Ollama: <https://n8n.io/integrations/ollama/>15- Ollama: <https://ollama.com/>1617## Workflow18191. Map trigger, deterministic transforms/routes, external reads, model step,20 draft output, and any external write separately.212. Select credentials/storage ownership, webhook exposure, retry policy,22 idempotency key, concurrency limit, and failed-run record before connecting23 an AI node.243. Treat the local model as one bounded node: prove structured output, tool25 compatibility when used, no-call behavior, malformed result recovery, and26 the exact model/server combination before exposing a write node.274. Keep source data minimised and trace/log retention explicit. Do not put28 secrets in workflow exports or prompts.295. Leave write nodes disabled or draft-only until `auto-with-escalation` names30 the recipient, action, payload evidence, and approval event.316. Export a testable workflow fixture and verify happy path, duplicate event,32 failed model response, credential failure, rejected approval, and recovery.3334## Output Shape3536Return trigger, node graph, credentials, idempotency/retry behavior, model37contract, draft artifact, approval event, and recovery evidence.3839## Guardrails4041- Prefer deterministic n8n nodes over agent planning whenever the route is42 known.43- Do not expose a self-hosted instance publicly without an explicit deployment44 and authentication decision.45- Do not claim a local model is reliable for tool use without a capability gate.
gaelic-ghost