Workflows

Tidying up text

Doing sums

The arithmetic formatters are how a workflow calculates without a formula column in the source. number first if the cell carries a currency symbol or grouping — ₹1,499 is text until it is stripped.

Dates, times, and values that change every row

date: 'DD MMM YYYY' formats any date — a column, or {{now.iso}}. The pattern is the spreadsheet one rather than strftime: YYYY, MMMM/MMM/MM, DD, dddd/ddd, HH/hh, mm, ss, A. A second argument names a timezone. Dates are read the way a sheet writes them, so 15/08/2026 and a Unix epoch both work, and addDays: 7, addHours, addMinutes and addMonths shift them.

Test a step to see what it returns

Sources and http.request carry a Test button. It runs that one step as currently configured — without saving — and lists the fields that came back with the values they held. Those fields then populate the dropdowns of every step below.

Testing is free and cannot write anything. It reads sources and calls read-shaped endpoints; it never renders, never spends credits, and never touches your spreadsheet. Steps that would cost money are not testable at all — they declare their output instead, which is why {{item.render.url}} is offered the moment a render step is added.

Templates ask for everything they need

Choosing a template loads every value it takes: the fields it declares, and any {{PLACEHOLDER}} sitting in the design that was never turned into one. Both are filled the same way at render time, so both are asked for. Required fields are marked, and the header counts how many are still empty.

Changing the template afterwards clears values that belonged to the old one — they would otherwise keep being sent on every render, invisibly, with no row in the form to show them.

Capture presets

The capture step takes a saved screenshot preset. Choosing one lists what it will actually do — viewport, format, selectors to hide — because a preset is a name over twenty settings, and “Product pages” tells you nothing about the width it captures at. Anything set on the step refines the preset rather than being overridden by it.

Endpoints that page

source.http reads more than the first page. Three schemes are supported because APIs never agreed on one: page numbers, offsets, or a cursor the previous reply handed back. All three stop on an empty page and all three are bounded, so a misconfigured cursor cannot loop against somebody’s API.

Without this a source would quietly automate the first hundred products and ignore the other nine thousand — the kind of wrong that looks like it worked.

On a schedule

A five-field cron expression, read in an IANA timezone you choose. The timezone is not decoration: the schedule is evaluated in that zone through daylight saving, so 6am means 6am where you are, all year. A schedule computed in UTC would walk an hour twice a year, which for a “post at 9am” automation is the whole thing failing quietly.

On an event

The same events webhooks deliver — render.completed, render.failed and the rest. The event payload is available throughout as {{trigger.payload.…}}.

A workflow never reacts to a render it produced itself, so “when a render finishes, render something” cannot loop. Each event also fans out to a bounded number of workflows, which caps the blast radius of a configuration that tries.

From a URL

A private hook URL any other system can POST to — a Shopify webhook, a Zapier action, a step in your deployment pipeline. The whole body arrives as {{trigger.payload.…}}.

curl -X POST https://pixbix.app/api/hooks/workflows/whk_9tK2… \
  -H "Content-Type: application/json" \
  -d '{ "campaign": "diwali-2026", "region": "west" }'

# 202
# { "success": true, "data": { "runId": "wfr_2b8e41", "status": "running" } }

Every method works— GET, POST, PUT and PATCH. A write method’s JSON body becomes the payload; a GET’s query string does, so ?orderId=1234 arrives as {{trigger.payload.orderId}} just as a posted body would. That means a system that can only fire a GET, and a person testing the URL in a browser, both work.

Send one call before you build the steps. The builder records whatever arrives — whether or not the workflow is switched on — and reads the field names out of it, so every value dropdown below offers the real paths with the real values beside them instead of asking you to type {{trigger.payload.customer.email}} from memory. A call to a switched-off workflow answers 202 and says so; nothing runs.

Hardening, all optional.The URL alone is a bearer credential, and a URL travels where a header does not — browser history, a third party’s audit log, a screenshot in a ticket. Three controls layer on top of it, each off by default so a hook never starts demanding something nobody configured:

When new rows appear

A repeating check over a spreadsheet, for the shape everybody wants and nobody wants to build: “when a contact is added, make their video”. The workflow remembers the last row it collected, so each check reads only past it — a sheet of fifty thousand contacts costs one small read every fifteen minutes, not fifty thousand renders.

By hand

Always available, whatever the trigger. Run it once from the dashboard or with POST /v1/workflows/:id/run before you trust it to a schedule.

Writing back to the row it came from

deliver.sheet has two modes, and the default is the one personalisation needs. In update mode it fills named columns on the row each result came from — the contact already has a row, and what you want is their video URL beside their name. Nothing else is configured: the spreadsheet, tab and row number travel on the row from the source step.

{
  "type": "deliver.sheet",
  "config": {
    "mode": "update",
    "updates": [
      { "column": "Video URL", "value": "{{item.render.url}}" },
      { "column": "Rendered on", "value": "{{now.date}}" }
    ]
  }
}

A column is named by its header — "Video URL"— or by its letter. Headers are resolved against the live header row at write time, so a column inserted in the sheet last week cannot send this morning’s URLs into the wrong cells. A name that matches nothing is reported in the run log rather than written somewhere arbitrary.

append mode is the exception: it adds new rows at the bottom, for when the results are a fresh log rather than an update. It inserts rather than overwrites, so a totals row or a second table underneath survives.

Personalised voice-over

audio.voiceover speaks a templated script through your own ElevenLabs connection, once per row — so every contact hears their own name. The MP3 lands on the row as {{item.voice.url}}, ready to be the audio of a video template.

Personalised videos for a contact list

Render a video for every contact, write each URL back beside that contact, then tell your own system about it — one call per contact, carrying both the sheet data and the render.

[
  {
    "type": "source.sheet",
    "name": "Read the contacts",
    "config": { "spreadsheetId": "1BxiMVs0…", "tab": "Contacts" }
  },
  {
    "type": "render.template",
    "name": "Render their video",
    "config": {
      "templateId": "welcome-video",
      "waitForCompletion": true,
      "values": {
        "name":    "{{item.First name}}",
        "company": "{{item.Company}}",
        "city":    "{{item.City}}"
      }
    }
  },
  {
    "type": "deliver.sheet",
    "name": "Write the URL back",
    "config": {
      "mode": "update",
      "updates": [
        { "column": "Video URL",   "value": "{{item.render.url}}" },
        { "column": "Rendered on", "value": "{{now.date}}" }
      ]
    }
  },
  {
    "type": "deliver.webhook",
    "name": "Tell your system",
    "config": {
      "mode": "perItem",
      "url": "https://api.acme.com/hooks/video-ready",
      "payloadTemplate": "{\"email\":\"{{item.Email}}\",\"name\":\"{{item.First name}}\",\"videoUrl\":\"{{item.render.url}}\",\"row\":\"{{item._row}}\"}"
    }
  }
]

Capture, voice-over, then render

The same list, with two more steps in front of the render: capture the page each contact’s row points at, and speak a script personalised to them. Both land on the row, and the render step consumes them as ordinary field values.

[
  {
    "type": "source.sheet",
    "name": "Read the contacts",
    "config": { "spreadsheetId": "1BxiMVs0…", "tab": "Contacts" }
  },
  {
    "type": "screenshot.url",
    "name": "Capture their page",
    "config": { "url": "{{item.Dashboard URL}}", "options": { "viewport": { "width": 1280, "height": 720 } } }
  },
  {
    "type": "audio.voiceover",
    "name": "Speak their script",
    "config": {
      "voiceId": "21m00Tcm4TlvDq8ikWAM",
      "text": "Hi {{item.First name}}, your {{item.Plan}} report for {{now.date}} is ready."
    }
  },
  {
    "type": "render.template",
    "name": "Render their video",
    "config": {
      "templateId": "report-video",
      "waitForCompletion": true,
      "values": {
        "screenshot": "{{item.screenshot.url}}",
        "voiceover":  "{{item.voice.url}}",
        "caption":    "{{item.voice.script}}",
        "name":       "{{item.First name}}"
      }
    }
  },
  {
    "type": "deliver.sheet",
    "config": { "mode": "update", "updates": [{ "column": "Video URL", "value": "{{item.render.url}}" }] }
  },
  {
    "type": "deliver.webhook",
    "config": { "mode": "perItem", "url": "https://api.acme.com/hooks/video-ready" }
  }
]

Order is the whole design here. The capture and the voice-over run before the render because the render reads what they produced; move either below it and the template receives an empty field rather than an error. The builder shows what each step leaves on the row underneath it, so the dropdown for a field only ever offers things that already exist at that point.

A folder as a source

source.drive gives every file in a folder its own row — filtered to images, videos or audio if you like, and paged, so a folder of ten thousand assets streams in windows exactly as a spreadsheet does. Paired with the when new files appeartrigger it lists only what changed since the last check, which is what makes “drop a product photo in this folder and get a video” cost one small query rather than a re-render of everything.

Making a file usable

A Drive link needs your credentials, so handing one to the render engine produces a 401 rather than a picture. drive.import downloads the file once and hosts it for the run, leaving {{item.file.url}} — drop that into any image or video field.

Google Docs, Sheets and Slides have no bytes to download and are refused with a message saying so. Export them from Google first.

Captioning the voice-over

captions.generate sends the voice-over to Google Speech-to-Text and gets every word back with its own timing. Put it under an audio.voiceover step and it needs no configuration at all — it defaults to {{item.voice.url}}, the clip that step just produced.

It leaves two things on the row, because they are wanted in different places. {{item.captions.cues}} carries per-word timings and is what a template embeds for burned-in captions that highlight each word as it is spoken. {{item.captions.vttUrl}} and {{item.captions.srtUrl}} are subtitle files — write one into a spreadsheet column, or hand it to a platform that takes a subtitle track.

Delivering on WhatsApp

deliver.whatsappsends one approved template per row from your MotherBot number, with that row’s render in the template’s media header. It is the step that makes a personalised-video run finish rather than stop at a spreadsheet column — the video reaches the person it was made for, on the channel they actually open.

Two WhatsApp rules shape the step, and both are Meta’s rather than ours. A business can only open a conversation with a template Meta has approved, so the template mode is the default and a plain message reaches only people who wrote to you in the last 24 hours. And a template can only carry a file of the kind its header was approved for — the picker greys out templates with no media header at all, and a video sent to an image template is refused before it costs a conversation.

The step asks two questions before anything else: which channel, and which connected number or address to send from. Both narrow what follows — templates are approved per WhatsApp number, so the template list is that number’s, and which message types are even possible is read off the chosen account rather than assumed from the channel’s name. The “send from” picker stays hidden until a workspace has more than one identity on a channel, because until then it has only one possible answer.

Attach takes any URL, from anywhere above the step: {{item.render.url}} for a video or image render, {{item.screenshot.url}} for a capture, {{item.file.url}} for a file imported from Drive, or a spreadsheet column that already holds a finished URL — in which case the workflow renders nothing at all and simply delivers. Leave it empty to send a template that has no media. There is no language setting: the template is looked up by name and sent in its own approved language.

Delivering back to Drive

deliver.drive uploads each finished output. The sub-folder is a template, created if it does not exist and reused if it does — Renders/{{now.date}} files a run by day without anyone maintaining folders, and naming a column instead files by region, campaign or customer.

What happens when something stops

Every row carries its own state and a step cursor. A row that was captured and voiced but not yet rendered resumes at the render — it does not pay ElevenLabs or the capture fleet a second time. A claim is a lease: if the process holding it goes away, the row returns to the queue and another picks it up. A run whose worker vanished mid-queue is resumed automatically.

None of that is best-effort recovery bolted on afterwards. It is the reason the queue is in the database rather than in a closure — a deploy in the middle of a six-hour run should cost seconds, not the run.

Retries

A row that fails is retried up to three times, unless retrying cannot help. Being out of credits, or pointed at a template that does not exist, fails identically every time — those stop immediately rather than burning three attempts and three log lines each. Running out of credits stops the whole run, because the alternative is one identical failure per remaining row.