Workflow Automation
Some projects want a layer of automation that does not belong in the application: a nightly export, a webhook that fans out to three services, a Slack message when an order fails, a scheduled reconciliation. A workflow engine such as n8n does that, and this skill builds those workflows cleanly. It starts from two refusals: it does not assume a project wants automation, and it does not pretend to have a connection it has not been given.
The concrete target here is n8n, reached over its MCP server. The method is
written to be general, because the next engine is reached the same way; the
n8n specifics are in resources/n8n-connector.md.
1. Per project, ask first
Automation is opt-in, per project, and the answer is recorded, not assumed.
before building anything
ask the project owner, once: does this project want workflow automation,
and for what. A project that does not want an external engine in its
stack should not get one because a workflow was easy to write.
record the answer
yes, with the workflows wanted; or no, and nothing is built. A silent
default of yes adds an operational dependency nobody chose.
A workflow engine is an operational dependency: another system to run, secure, monitor and pay for. Adding it to a project that did not ask is the same defect as adding a library nobody needed.
2. The connector is detected, explained, never invented
n8n is reached over an MCP server that must be configured and authorized outside this skill, interactively, by the person running the agent. This skill does not fabricate that connection.
detect determine whether the n8n MCP connector is configured and
reachable in this environment
present if it is not, state exactly what is missing: the MCP server is
not configured or not authorized
guide point to the configuration step (the MCP configuration, the
authorization flow), which is the user's to run interactively
stop until the connector is real, no workflow is built against it, and
no access is assumed or simulated
This is the general rule for any external connector, stated in the source material: detect the requirement, explain it, guide the user to configure it, and do not invent access. A skill that pretends a connection exists produces a workflow that cannot run and a report that lies.
3. What a clean workflow has
Once the connector is real and the project has opted in, every workflow built carries these, and none is optional.
| Property | What it means |
|---|---|
| Idempotency | a workflow that runs twice on the same input produces the same effect once, because a scheduler or a retry will run it twice |
| Error handling | every node that can fail has a failure path: a caught error, a notification, a dead-letter, not a silent stop |
| Bounded retries | a failed step retries with backoff a stated number of times, then stops and reports, rather than retrying forever |
| Timeouts | every outbound call has a timeout and a decided behaviour when it is hit |
| No embedded secret | credentials come from the engine's own credential store or the environment, never hardcoded in the workflow definition |
| Signed inbound | a webhook the workflow receives verifies its signature before acting |
| Observability | a run leaves a trace an operator can read: what ran, what it touched, what failed |
| Versioned definition | the workflow definition is exported and version controlled, so it is reviewable and restorable, not only living in the engine |
4. Secrets never live in the workflow
The workflow definition is exported, committed and shared. A credential in it is a leaked credential.
credentials held in the engine's credential store, referenced by name,
or read from the environment; never written into a node
the exported JSON reviewed before commit for any token, key or connection
string, exactly as any diff is
a leaked one rotated, not just deleted, because export and commit may
already have carried it
This is secrets-management applied to a workflow definition: the definition
is source, and source carries no secret.
5. Verifying a workflow
A workflow is not done because it was created in the engine. It is done when it was run and observed.
run it on a real input, or a representative test input
observe the trace: every node reached, the effect produced once, the failure
paths exercised by a deliberately bad input
confirm idempotency by running it twice and checking the effect happened once
confirm the retry and timeout behaviour by forcing a failure
A workflow reported as working without a run observed is the same fabrication
implementation-integrity forbids in application code.
6. Prohibitions
- No workflow engine added to a project that did not opt in.
- No workflow built against a connector that is not actually configured and authorized; the connection is never assumed or simulated.
- No secret written into a workflow definition.
- No unbounded retry, no failure path that silently stops.
- No workflow reported as working without a run observed.
- No inbound webhook acted on without verifying its signature.
7. Protocol
- Ask the project owner whether this project wants workflow automation, and for what. Record the answer. If no, stop.
- Detect whether the n8n MCP connector is configured and reachable. If not, state what is missing and hand the interactive configuration step to the user; do not proceed against a connection that does not exist.
- For each wanted workflow, design it with the section 3 properties: idempotency, error handling, bounded retries, timeouts, observability.
- Keep every credential in the engine's credential store or the environment, and scan the exported definition for secrets before committing it.
- Run each workflow on a real or representative input, observe the trace, force a failure to exercise the error and retry paths, and run it twice to confirm idempotency.
- Export the workflow definition, review it, and commit it under version control so it is reviewable and restorable.
8. Auto-critique
Score from 0 to 5: the project actually opted in rather than being assumed into automation, the connector was detected and its absence handed over rather than simulated, every workflow has idempotency, error handling, bounded retries and timeouts, no secret is in the definition, and each workflow was run and observed rather than declared.
Threshold: no axis below 3, average at least 4. A workflow built against a connector that was assumed rather than confirmed scores 0 overall, because it cannot run and the report that says it can is false.
9. Interfaces
- Upstream:
engineering-core,devops-core. - Lateral:
secrets-managementfor the credentials the workflow uses,background-jobswhen the automation belongs in the application's own job system rather than an external engine,observabilityfor the run traces,input-validationfor the inbound webhook payloads. - Reference data:
resources/n8n-connector.md.