Create a durable Zapier workflow from natural language using @zapier/zapier-durable and the Zapier SDK CLI. Use when the user wants to build a Zapier workflow, create an automation, write a durable workflow, build me a Zap that, create a durable that, or automate a multi-step process involving Zapier-connected apps.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Instruções da origem · Visualização somente leitura
name
workflows-create
description
Create a durable Zapier workflow from natural language using @zapier/zapier-durable and the Zapier SDK CLI. Use when the user wants to build a Zapier workflow, create an automation, write a durable workflow, build me a Zap that, create a durable that, or automate a multi-step process involving Zapier-connected apps.
Create a complete durable workflow from natural language, test it when appropriate, and deploy it through the Zapier SDK experimental Code Workflows commands.
Use the public SDK CLI path. Do not use zapier-sdk-code-substrate.
Compatibility Gate
Before using this skill, run the workflows-doctor bundle compatibility check. If workflows-doctor is not installed or cannot be loaded, run workflows-install or install workflows-doctor from zapier/agent-skills before continuing. If workflows-doctor reports SDK/skill drift, follow its refresh instructions, stop this skill invocation, reload the agent workspace if needed, and ask the user to rerun the original request.
Pass latest for versions — the platform resolves it. The sandbox installs with pnpm install --config.minimumReleaseAge=1440, so a dependency published less than 24h ago is rejected, and @zapier/zapier-sdk publishes several times a day. You do not have to work out which release is old enough: latest is resolved for you to the newest release that clears that 24h gate (COSUB-956).
That only applies to @zapier/zapier-durable and @zapier/zapier-sdk. They are resolved server-side, at publish and at run submit, and the exact version is stored on the workflow version or the run — so a workflow stays on the version it resolved to, however long it runs.
A spec that isn't resolved server-side is stored as written, and every tick regenerates package.json from it with no lockfile. So "zod":"latest" would install whatever is newest at each tick, and a run that spans a zod major would replay against a different library than it recorded. The install itself would succeed — pnpm falls back from the latest tag to the newest release outside the 24h gate — which is what makes this worth stating: the failure shows up later, as a replay that doesn't match.
When publishing a workflow version you can omit zod entirely and the service injects a pinned version for you. run-durable injects nothing, so declare there whatever the source imports.
Ranges (^1.2.3, ~1.2) are rejected with a 400 for the durable runtime version — whether you pass it as --zapier-durable-version or as the @zapier/zapier-durable entry in --dependencies. So latest is the only non-exact value it takes. Every other dependency does accept a range and stores it as written; pin those exactly anyway, for the reason above. An exact version is always fine; it must be at least 24h old.
Every package the generated workflow.ts imports must still appear in --dependencies — the sandbox installs from --dependencies, not your local package.json, so a missing import (such as zod) fails the run with Cannot find package.
Read the user's natural language request and extract:
Steps and ordering.
Apps involved.
Data passed between steps.
Manual input fields or trigger input fields.
Conditional logic.
Waits, callbacks, or human approval gates.
Start mode (required — see below).
Summarize the proposed workflow back to the user before discovery. Ask focused clarifying questions for missing details like target channels, folders, recipients, or whether to stop when a search returns no results.
Do not generate code until the user agrees on the workflow shape.
Classify The Start Mode (Required)
Every workflow has exactly one start mode, and you must decide it here — it is a required output of this phase, not something inferred later from whether a trigger happened to get configured. There are two:
trigger — the workflow starts on its own, either on a schedule (scheduled) or in response to an external event (event-driven). It is published with --trigger and runs without anyone invoking it.
manual — the workflow has no trigger and runs only on-demand via trigger-workflow. This is a deliberate choice, not the absence of a trigger.
Default to looking for a trigger in the build instruction. Infer from intent cues:
Recurring or time words ("every morning", "daily", "each hour", "on a schedule") → trigger, scheduled.
"when X happens in <app>" / "whenever a new <record> is created" → trigger, event-driven.
Explicit "manually", "on demand", "when I run it", or a workflow clearly meant to be invoked by hand with input → manual.
If the start mode is ambiguous — no clear trigger cue and no explicit manual cue — ask the user: run it manually on-demand, or attach a trigger (and which)? Never silently assume manual just because no trigger was named. A triggerless workflow published on an unconfirmed assumption is the exact failure this classification exists to prevent.
This decision is required even though the platform treats its start-mode input as optional — the write-only manual flag sent at publish is not a required field, and enforcement is per-account and flag-gated, so the optionality is a staged-rollout artifact, not permission to skip the decision. Never publish without a start mode the user, or an unambiguous build instruction, chose.
Carry the chosen start mode forward: it is confirmed in the Phase 3 build plan, drives the publish in Phase 6, and is the gate for the Phase 7 verification.
Phase 2: Discover Apps, Connections, Actions, Triggers, And Fields
Use the standard Zapier SDK CLI for app/action discovery:
If several apps, connections, actions, triggers, or field choices are plausible, show the candidates and ask the user to choose.
Use "AI by Zapier" For AI Steps
For any AI / "call an LLM" step — summarize, extract, classify, generate, or analyze text — always use "AI by Zapier" (app key AICLIAPI) as the step and select the model inside it: if the user names a provider or model, set that as the model_id (see below); otherwise use its default model. It runs on Zapier's built-in AI credentials (no third-party account required) and bills as normal Zapier tasks, so an agent-built workflow does not silently route to a separate raw-provider app the user must connect and pay for. Discover it with list-apps --search "AI by Zapier"; its generic completion action is get_completion ("Analyze and Return Data"), alongside extract_content (from a URL) and search_content (confirm the current set with list-actions AICLIAPI --action-type write --json).
Configuring the get_completion step. Inspect its fields with list-action-input-fields AICLIAPI write get_completion --json. The ones that matter for a generated step:
instructions (required) — the prompt describing what the AI should do.
provider_id (optional) — the AI provider, needed only when the user names one. Choices are openai, anthropic, google, azure-openai, amazon-bedrock (list-action-input-field-choices AICLIAPI write get_completion provider_id --json). Setting it is what makes model_id's choices resolve.
model_id (required, default "advanced/auto") — the model. For a generic step, pass the default "advanced/auto" — auto-pick a model in the Advanced tier (tiers: standard/advanced/premium) on built-in credentials. When the user names a provider or model, set provider_id first, then resolve the valid model for it with list-action-input-field-choices AICLIAPI write get_completion model_id --inputs '{"provider_id":"<provider>"}' --json (the list is empty until provider_id is set) and pass the matching <provider>/<model> value (for example anthropic/claude-sonnet-5, openai/gpt-4o). Do not hardcode a model list — resolve it at build time.
authentication_id (required, default "0") — "0" is Zapier's built-in AI credentials (the models shown with a Zap icon). Keep "0" for the default and any built-in model. A model the user names may not be available on built-in credentials — those require the user's own AI provider account (a custom authentication_id); if so, tell the user and use their authentication. model_id depends on this field.
inputFields (optional, OBJECT) — extra context fields mapped from earlier steps, merged into the prompt.
So a default AI step needs only a prompt. model_id and authentication_id are required but have working defaults; pass them explicitly with those defaults ("advanced/auto" and "0") so the runAction inputs are complete, and no connection alias is needed for the built-in path:
Naming a provider or model is not a reason to leave "AI by Zapier" — set it as the model_id above. Reach for a raw-provider AI app (Anthropic, OpenAI, Google AI, and so on) only when the user explicitly asks for that standalone app, or needs a capability "AI by Zapier" does not offer. When you do, tell the user the step uses their own provider connection and billing, not "AI by Zapier."
Assign a short snake_case connection alias for each chosen connection, such as slack_work or gmail_primary. Track alias to connection ID. The alias goes in workflow code; the connection ID is passed to test/deploy commands through the --connections JSON.
For output mapping between steps, run a safe action test only after user confirmation. Use the current SDK command shape:
Use the version-pinned app/API identifier for selected_api, the trigger action key for action, the trigger source connection ID for authentication_id when the trigger requires auth, and trigger input values for params. Omit optional fields only when the trigger does not need them.
For selected_api, use the version-pinned implementation identifier — the implementation_id returned by SDK discovery (list-apps/get-app), such as GoogleSheetsAPI@2.3.0. Do not use the bare app key (GoogleSheetsAPI) and do not substitute a display name. A bare, unversioned selected_api makes the trigger claim fail silently at publish: the publish call returns success with no errors, but the workflow stays disabled and nothing surfaces the cause. If discovery only exposes a bare app slug and not a versioned implementation_id, treat that as a blocker and record it in the build plan before publishing — do not publish a trigger with an unversioned identifier.
For params, match each field's value_type from list-trigger-input-fields <app> <action>. ARRAY fields must be JSON arrays (for example "dow": ["1"]); STRING fields must be plain strings (for example "hod": "9:00 AM"). Passing a scalar where an array is expected (or vice versa) fails the trigger claim the same silent way.
Capture app implementation/version information from SDK discovery output when available, such as list-apps, get-app, list-actions, or trigger/action result metadata. Do not invent app versions. If no implementation/version binding is exposed, omit --app-versions rather than guessing.
"Webhooks by Zapier" and other apps with a catch-hook trigger (PayPal, Salesforce, Twilio, WordPress, Wufoo, Zillow, and others) are discovered and configured exactly like any other trigger app — nothing about them is special-cased. Search list-apps --search "webhook" (or the specific app name) for its implementation_id (for example WebHookCLIAPI@1.1.0 — an illustrative example, not a version to hardcode; confirm the current version via discovery), then list-triggers <appKey> for its catch-hook trigger action. "Webhooks by Zapier" itself is no-auth (authentication_id: null) with empty params, but confirm its action key via list-triggers WebHookCLIAPI rather than hardcoding one — as of this writing it exposes both hook_v2 (parsed payload; the common default) and hook_raw (unparsed body and headers, max 2MB), and that pair of action keys is specific to WebHookCLIAPI, not a pattern the other apps share. Other catch-hook apps (PayPal, Salesforce, Twilio, ...) commonly require a connection, because claiming their trigger means calling the provider's API to register a subscription. Do not assume no-auth or empty params for those — confirm each app's actual action key, auth, and param requirements via list-triggers/list-trigger-input-fields (see above) rather than generalizing from "Webhooks by Zapier." Configure them through --trigger at publish time (Phase 6) like any other trigger — do not treat them as "no trigger" / manual-only workflows.
Phase 3: Confirm The Build Plan
Before writing code, present:
Workflow: <kebab-case-name>
Input: { field1, field2 }
Connections:
alias = connectionId (connection title)
Start mode: trigger (<selected_api.action with params, including "Webhooks by Zapier" or other catch-hook apps>)
— or —
Start mode: manual — on-demand only via `trigger-workflow`
Steps:
1. <step-name> - <AppName>.<actionType>.<actionKey>
2. <step-name> - <AppName>.<actionType>.<actionKey>
Return: <summary of output>
The Start mode line is required and must state exactly one of the two modes classified in Phase 1. manual is a deliberate, user-confirmed selection — never render it as "no trigger" or leave it implied by an absent trigger. If the mode is still ambiguous at this point, resolve it with the user before proceeding (Phase 1).
Ask the user to confirm before generating files, including explicit confirmation of the start mode.
Generate durable source that can be turned into a meaningful step graph. Avoid overly dynamic construction.
defineDurable call shape — every call must resolve run to a function. Use either the bare form defineDurable("workflow-name", async (ctx, input) => { ... }) or the object form defineDurable({ name: "workflow-name", inputSchema, outputSchema, description, run: async (ctx, input) => { ... } }). ctx is always the first parameter of run; input is the optional second parameter, so async (ctx) => { ... } is also valid. These shapes are invalid and make the workflow fail on its first run with durable.run is not a function:
defineDurable(async (ctx, input) => { ... }) — a bare function with no name. The function is treated as an options object, so run is never set. This is the most common mistake.
defineDurable({ name: "workflow-name", run: someNonFunction }) — run is not a function.
durable.run is not a function is a code-shape defect in your defineDurable call, not a version mismatch. Do not change the pinned @zapier/zapier-durable or @zapier/zapier-sdk versions to fix it — correct the call so it passes a name and a run function.
Default to this parser-friendly shape — module-level sdk, hoisted app-key/connection constants, and a bare runAction body for each app action:
The editor renders a ctx.step as an app-action step (with the app icon) when its body is a single sdk.runAction({...}) call with appKey, actionType, and actionKey (object literal, or a const that resolves to one; the app / action spellings also work). A string-literal step id (ctx.step("create-todoist-task", ...)) and an inline async () => ... callback are the recognized form; object form ctx.step({ name, run }) works too.
Other steps render as plain code steps — for example a step with no runAction, or with more than one, or one created in a loop with a dynamic id (`process-item-${index}`). That is expected, not a regression; loops and fan-out legitimately need dynamic ids.
Build the connections JSON from the selected aliases. It's a nested object — each alias maps to an object holding a connectionId (never a bare string). The same shape is used for publish-workflow-version in Phase 6:
run-durable returns a run immediately, often before the workflow is complete. Capture the returned run ID, then poll until terminal status. Do not assume the first response contains final output.
Terminal success means the run has status: "finished", an expected output, error: null, and top-level errors: []. Terminal failure means status: "failed" or a non-null error. Continue polling while the run is initialized or started.
Fix code and retest until the behavior matches the confirmed plan.
Phase 6: Deploy The Workflow
Decide whether the workflow should be private before creating it. For EA users, default to private unless the user explicitly wants an account-visible workflow.
Omit --private only if the user explicitly wants the workflow visible to the broader account.
Capture the returned workflow ID. Then decide how to ship the code:
Direct publish (the default below): publish the first version straight away with publish-workflow-version. This is the legitimate no-open-draft case — the container was just created, so no draft exists to publish past.
Stage as a draft for review: if the user wants to look the workflow over in the Zapier editor before it goes live, put the generated code in a server draft instead of publishing:
zapier-sdk --experimental create-workflow-draft <workflow-id> --json
zapier-sdk --experimental update-workflow-draft <workflow-id> <draft-id> "$SOURCE_FILES" \
--draft-revision <draft_revision from the create response> \
--json
Pass the same --dependencies, --zapier-durable-version, --connections, --app-versions, and — for a Start mode: trigger workflow — --trigger values Phase 6 would have passed to the publish. Then hand the user the draft's editor link — https://zapier.com/durables-editor/<workflow-id>/draft/<draft-slug>/workflow.ts, using the slug from the draft response; the final segment is one of the draft's source_files keys (workflow.ts in this skill's flow) — to review and publish, or publish on their go-ahead. Carry the start-mode decision to the draft publish exactly as a direct publish would: a Start mode: manual workflow publishes with --manual (publish-workflow-draft <workflow-id> <draft-id> --manual --enabled --json); a Start mode: trigger workflow's draft already holds its --trigger, so publish without --manual (publish-workflow-draft <workflow-id> <draft-id> --enabled --json). Never pass --trigger and --manual together. Publishing consumes the draft. Skip Phase 7's version read-backs if nothing was published.
For a direct publish, the current SDK CLI expects source_files as a JSON object, not a path to workflow.ts.
For publish, use the same nested connections shape as run-durable — each alias maps to an object holding a connectionId:
Omit the entire --app-versions flag when no app implementation/version binding is needed. Likewise, omit --connections when the workflow has no connection bindings. Do not pass placeholder text like "if needed" to the CLI.
For trigger-backed workflows, build the trigger JSON from Phase 2. Keep selected_api version-pinned to the implementation_id (for example GoogleSheetsAPI@2.3.0) and keep each params field shaped to its value_type (see Phase 2) — a bare app key or a wrong param shape makes the trigger claim fail silently at publish:
A "Webhooks by Zapier" or other catch-hook trigger is a real trigger — publish it with --trigger using the config captured in Phase 2, the same as any other app trigger.
How you publish follows directly from the start mode confirmed in Phase 3 — the two are not co-equal defaults; you commit to the one the user chose.
Before publishing, confirm the payload matches the declared start mode:Start mode: trigger → the publish passes --trigger and not--manual; Start mode: manual → it passes --manual and not--trigger. Pass exactly one — the platform contract is a discriminated union (a version is either triggered or manual: true, never both) and rejects passing both together as a contradiction; the CLI also guards the both-case client-side. The workflow you publish must carry the trigger you decided on, or be explicitly marked manual. Catch any disagreement here, before the publish call, so a dropped or missing trigger is not discovered only in Phase 7.
Start mode: trigger — publish with --trigger, using the config built above. The trigger is the signal; do not also pass --manual (that is the contradiction the gate rejects):
Start mode: manual — and only when Phase 3 confirmed manual — omit --trigger and pass --manual to declare the on-demand start mode explicitly. Marking manual is the deliberate branch, not a fallback for when a trigger was hard to configure: if the user asked for a trigger, a failure to build its config is a blocker to resolve, never a reason to drop to manual.
--manual and --trigger are mutually exclusive: pass --manual here because there is no trigger. Never pass both in one publish.
Do not use the old --trigger-app, --trigger-action, --trigger-auth, or --trigger-params flags. The current trigger publish path is the single JSON --trigger object.
If the publish is rejected with a conflict about open drafts, someone (likely the user, in the Zapier editor) forked a draft on this workflow mid-flow. An open draft always holds unpublished work, so never publish past it silently. Tell the user and offer the same choices as workflows-modify: fold your changes into that draft and publish it (update-workflow-draft + publish-workflow-draft), or — with their explicit confirmation, since it drops the draft's unpublished work — discard the draft (discard-workflow-draft) and retry the direct publish.
Verification must confirm the deployed workflow matches the start mode confirmed in Phase 3 — not merely that it is enabled. A triggerless workflow reads back as enabled: true, so an enabled check alone silently passes a workflow that was supposed to have a trigger but doesn't. Re-read the workflow (the trigger claim is asynchronous and can fail silently, so allow a few seconds and poll if needed) and gate on the declared mode:
Start mode: trigger → require bothenabled: trueand a non-empty triggers[]. An empty triggers[] means the trigger was dropped or --trigger was omitted — the claim failed or was never attempted. Do not report the workflow as done. The most common cause is a selected_api that is not version-pinned to the implementation_id, or a params field with the wrong shape (see Phase 2); enabled: false after publishing with --enabled is the same failure. Re-publish with a corrected --trigger and re-check.
Start mode: manual → require triggers[] to be empty by design, and confirm this workflow was deliberately classified manual in Phase 3 (never triggered). A manual workflow with a non-empty triggers[] is also a mismatch — stop and reconcile with the user. It is invoked on-demand via trigger-workflow; there is no trigger claim to verify.
triggers[] is the authority for this gate. The platform's start-mode input (the write-only manual flag) is never surfaced on any read-back, so do not look for it on get-workflow or a version — read the trigger's actual presence instead.
Regardless of trigger type, check the matching entry in triggers[] from the get-workflow --json read-back above for details.webhook_url (re-run the same command if enough time has passed since that read that the claim state could have changed). If present, it is the catch URL external services call — show it to the user plainly; unlike the workflow-level trigger_url, it is meant to be shared. Most triggers have no webhook_url, and that is normal — do not flag its absence.
If you configured a catch-hook trigger in Phase 2 (a "Webhooks by Zapier" or similar catch-hook app/action) and details.webhook_url is still absent once the trigger is active, the installed @zapier/zapier-sdk may predate this field — run workflows-doctor to check for an update, and in the meantime tell the user to copy the URL from the trigger step in the Zapier editor (https://zapier.com/durables-editor/<workflow-id>).
If manual triggering is supported for the workflow, test it only after confirming side effects with the user:
Whether the workflow is private or account-visible.
The confirmed start mode: trigger (a Zapier app trigger or a catch-hook trigger — report its webhook_url if available) or manual (on-demand only via trigger-workflow), and that the Phase 7 gate confirmed the deployed workflow matches it.
The Zapier editor link: https://zapier.com/durables-editor/<workflow-id>.
Durable Patterns
Waits
await ctx.wait("wait-before-followup", 3600);
Place waits at top-level workflow scope, not inside ctx.step.
Loop/fan-out steps use a dynamic id (`process-item-${index}`), so the editor renders them as code steps — expected for this pattern (see App-Action Step Shape (Editor Recognition)).
