# Working with Claude

> What to ask Claude in Claude Code to build, edit, test, diagnose, connect, extend and deploy APIANT automations, and how Claude behaves while it does.

In Claude Code with the APIANT skills installed, you describe what you want and Claude picks the procedure for it: building an automation, editing one, testing, diagnosing a failure, building a connector. Each procedure is a skill that tells Claude which APIANT tools to call, in what order, and when to stop and ask you. You do not name skills or tools; plain requests are enough.

## How Claude works on your account

- **It looks things up before it builds.** Apps, operations and field names come from the APIANT catalog, not from memory.
- **It builds without stopping for approval,** and asks only what it cannot decide: which folder to use, a webhook payload it needs to see, an app you have not connected.
- **It pauses before saving when a mapping is at risk.** Dates, phone numbers, currency and lists that differ between apps are listed with options: check your existing automations, research the apps' API documentation, add a conversion step, or keep the mapping.
- **New automations are saved turned off.** Claude tests them first and turns them on when you ask.
- **It tests every build.** After a clean run, it works through the condition branches the run did not take, changing the recorded data to force each one, until every reachable branch is covered or it tells you which cannot be.
- **Edits happen in a draft.** Claude's changes stay in a draft owned by that Claude session until it saves them as a new version. It asks whether to overwrite the automation or save a copy.
- **It links you to the result.** After saving, Claude gives you the automation's diagram link, and the webhook URL when there is one.

## Build an automation

```text
When a Typeform response comes in, create a HubSpot contact and add a note with the answers.
```

Give the apps, the event that starts it, and what should happen. For a vague goal, Claude asks which apps, which event and what data should move before it builds. Claude recognizes common patterns and follows the matching procedure, for example hosted forms, chat widgets, web services, approval steps, delays, collecting items for a daily digest, fan-out and fan-in, two-way sync, and automations exposed as MCP tools.

## Build an integration of several automations

```text
Build a two-way sync of contacts between HubSpot and Salesforce, with a nightly reconciliation.
```

Claude first asks about entities, sync direction, triggers, business rules and volume, then designs the set of automations, builds the shared pieces, wires the automations to each other, tests the data flow end to end, and organizes them in a folder.

## Edit an automation

```text
In the Typeform to HubSpot automation, skip responses without an email address.
```

```text
Swap the Slack step for a Microsoft Teams message to the Sales channel.
```

To stage several changes by pointing at steps on the diagram, see [Review a diagram and send comments to Claude](/docs/mcp/diagram-comments).

## Test an automation

```text
Test it with a response from ada@example.com.
```

```text
Replay the webhook from yesterday's failed run.
```

```text
Test the whole HubSpot and Salesforce sync end to end.
```

Claude can run a test with sample trigger data, replay a recorded webhook, restart a run from a step, and change a run's recorded data to force a branch. For an instant trigger with no sample, it may build a temporary automation that creates a record in the source app to fire the real event, or ask you to cause the event. See [Testing an automation](/docs/automations/testing).

## Diagnose a failure

```text
Why has the Shopify to Slack automation been failing since this morning?
```

Claude finds the recent failed runs, reads the failing step's input and output, compares them with the automation's configuration, and explains the cause in plain language with a link to the run. It suggests a fix and applies it when you agree, then retests. It can also retry failed runs once the cause is fixed.

## Tune error alerts and retries

```text
My automation keeps emailing me "Connection reset by peer". Stop alerting on that one.
```

```text
Why was this automation turned off after a 401?
```

Claude can explain why an alert fired, why a run did or did not retry, and why a 401 turned an automation off. It can rewrite or suppress alert messages for a step or an automation. Alert rules that apply to every automation on your account need an administrator. Changing which errors are retried, or which services' `401` responses stop turning automations off, goes through APIANT support; Claude gathers the evidence to send. See [Error alerts](/docs/runs/alerts).

## Connect an app with your own OAuth app

```text
Register my own Zoom app and connect it.
```

For apps that need an application registered on the vendor's developer portal, Claude drives a browser through the portal: it finds the page, fills the form, selects the scopes and pastes the exact redirect URI, then saves the credentials and connects the app. You sign in to the vendor and answer any two-factor prompt. This needs a browser-control tool available in your Claude Code session. See [Use your own OAuth app or API keys](/docs/apps/own-oauth-app).

## Add a missing operation or connector

```text
The Acme CRM connector has no way to list invoices. What operations should it have?
```

```text
Build a List invoices action for Acme CRM.
```

Claude reads the app's own API documentation and proposes a ranked list of operations without changing anything. On request it builds each operation, calls the real API to get a sample response, and tests the result. Operations you build are private to you and are not listed in the shared catalog. This requires a plan that allows app authoring. When a build needs an app the catalog does not have, Claude offers to build a connector for it with API-key authentication; a connector that signs in with OAuth needs an additional entitlement on your account.

See [Build a connector, action or trigger](/docs/apps/custom-connectors).

## Publish and deploy

> **Plan:** Enterprise

```text
Publish the onboarding folder to production.
```

```text
Deploy the new-client onboarding automation to all child accounts as a linked deploy.
```

Claude publishes template automations from your development environment to production, then deploys them to customer accounts, asking each time whether the deploy is linked (later deploys update the copies in place) or unlinked (fresh copies). Subroutines the automation calls travel with it. Credentials do not; each account connects its own apps. See [Publish and deploy](/docs/account/publish-and-deploy).

## Report a problem to APIANT

```text
Report to APIANT that the execution history page is slow to load.
```

Claude drafts the report in your words, shows it to you, and sends it only when you say yes. Separately, when Claude sees the APIANT server itself return something broken, it files a defect report on its own and gives you the report's reference in one line.

## Tips

- Name the apps, fields and folder you want. Claude asks fewer questions.
- Say what should happen to the data, not how to wire it. "Only for paying customers" is enough for Claude to add the condition.

## Next steps

- [Review a diagram and send comments to Claude](/docs/mcp/diagram-comments)
- [Testing an automation](/docs/automations/testing)
- [Claude Code skills](/docs/reference/skills)
- [Troubleshooting MCP](/docs/mcp/troubleshooting)