All articles

Add timed polls and CTAs to a webinar by asking an AI assistant

Place polls and call-to-action buttons on a pre-recorded webinar's timeline by asking an AI assistant. The prompt to paste, the duplicate check, and the limits.

Josh Miller

WebinarStack puts polls and CTAs on the video timeline at a second-level offset, separately for live sessions and replays, and an AI assistant can place them for you through our MCP server. It has a tool that checks for an existing poll or CTA before anything is written, and the assistant is told to use it first, so asking twice should not leave you with two copies.

"Should" is doing real work in that sentence, and the section on the duplicate check explains exactly how much. First, what a timeline interaction is.

What is a timeline interaction?

A timeline interaction is a poll or a call to action pinned to a moment in the video. It has a type (poll or cta), a headline, a start offset in seconds (appearsAtSeconds), an end offset (endsAtSeconds) and a timeline, which is either live or replay.

The timeline setting is the detail most tools gloss over. An interaction on the live timeline appears during a scheduled session, and one on the replay timeline appears when someone watches the replay. They are configured separately. An offer that lands well at 24:00 of a session may deserve different timing, or a different offer, on the replay, and you can set that without touching the live version.

A poll has between two and eight options and can be single-choice or multi-choice. A CTA can carry a description, a button label, a button URL and a countdown.

What timeline are we going to build?

Three interactions, covering both types and both timelines:

  1. A qualifying poll at 05:00 on the live timeline: "What is your biggest struggle right now?" with four answers.
  2. A CTA at 24:00 on the live timeline, shown for 90 seconds, linking to your offer.
  3. A second CTA at 20:00 on the replay timeline only, shown for two minutes, with a bonus for people who act on the replay.

What do you need first?

An MCP key with the interactions write scope, an assistant connected to it, and an existing webinar with its video set. The WebinarStack MCP server guide covers the key and the connection. The sibling tutorial, building an email sequence by prompting an AI assistant, is shaped the same way and covers the automation side.

What is the prompt?

Paste this, replacing the bracketed parts. Timings are MM:SS because that is how the product reports them back to you:

Use WebinarStack to add three interactions to my webinar "[webinar title]".
Before creating each one, check whether a matching interaction already exists
and tell me what you found.

1. A poll on the live timeline at 05:00, visible for 30 seconds.
   Headline: "What is your biggest struggle right now?"
   Options: [option 1], [option 2], [option 3], [option 4]. Single choice.

2. A CTA on the live timeline at 24:00, visible for 90 seconds.
   Headline: "[offer headline]". Button label: "[button label]".
   Button link: [offer URL].

3. A CTA on the replay timeline at 20:00, visible for 2 minutes.
   Headline: "[replay offer headline]". Button label: "[button label]".
   Button link: [replay offer URL].

When you are done, list all interactions on the webinar.

What does the assistant do with it?

For each interaction it calls interactions_find_matching, and if nothing matches it calls interactions_create. Put the first of those first, because it is the detail that matters most if you are nervous about letting an assistant write to your webinar.

interactions_find_matching is a read-only check, and what counts as a match depends on the type:

  • For a CTA, a match is another CTA on the same timeline within ten seconds of the same start offset.
  • For a poll, a match is a poll on the same timeline whose headline is the same after trimming whitespace and lowercasing.

If it returns an empty list, the tool's own description calls that safe to create. If it returns matches, the assistant sees the existing interactions and can leave them alone, update one or ask you.

In practice, ask for the same poll twice and a well-behaved assistant finds the first poll the second time and does not create another. The check is also scoped to the timeline, so a CTA at 24:00 on the live timeline does not count as a duplicate of one at 24:00 on the replay.

How firm is that? The create tool does not run the check itself. interactions_create and interactions_find_matching are separate tools, and the create tool's description tells the assistant to call the check first. That is guidance the assistant follows, not a lock the server enforces. If an assistant skips the check, the create call goes through and you get a second poll. The practical advice is to put the check in your prompt, as the prompt above does, and to read the list at the end.

How do you read it back and change it?

Ask the assistant to list the interactions. It uses interactions_list, which can be filtered by type or by timeline. interactions_get returns one interaction in full.

When a write tool confirms an interaction back to you, it uses one format: type, headline in quotes, timeline and time. interactions_update and interactions_delete return it, so a change reads as a line you can check at a glance. For the timeline above:

poll 'What is your biggest struggle right now?' @ live 5:00
cta 'Get the course now' @ live 24:00
cta 'Replay bonus' @ replay 20:00

To move a timing, ask for it in plain language. The assistant uses interactions_update, and there is one thing to know: an update replaces the interaction rather than merging into it. Fields you leave out reset to their defaults, so the safe sequence is that the assistant fetches the interaction, changes the one field and sends the whole thing back. The tool's description says so, but say it in your prompt too, and check the result with interactions_get.

interactions_delete removes one, flagged as destructive and returning a snapshot of what it deleted.

How does a poll answer drive an automation?

A poll answer is one of the events the automation engine can act on. poll_response is an automation trigger, so a rule can fire when an attendee answers a poll. branch_taken is another trigger type, for when an attendee is routed down a branch.

Rules can also use the answer as a condition, so different answers can lead to different emails. The exact way to express an answer in a condition is in the product's interaction authoring guide, shipped on the MCP server as the resource webinarstack.docs.interaction_authoring. Attach it to the conversation and the assistant will write rules against it.

The email sequence tutorial shows how to build the rules themselves, and it lists all 15 triggers an assistant can use. This page covers the timeline those rules react to.

Where does a human still have to look?

The timings. An assistant can place a CTA at 24:00. It cannot watch your video and tell you that 24:00 is thirty seconds after the point where the pitch actually lands, or that the poll at 05:00 pops up in the middle of your best story.

Watch the video with the timeline open, once, before you publish. The assistant handles the placement and the consistency, and you handle the judgement. Publishing is a step in the dashboard's webinar editor, not a tool the assistant has, and the editor lists anything still missing before it lets you.

For how interactions fit with the rest of the product, see the features page. To see why a session can start minutes after someone registers, read what is a just-in-time webinar. For a comparison with the closest product to ours, see WebinarStack vs eWebinar.

Run this on your own webinar

WebinarStack records your webinar once and runs it every day after that, with follow-up driven by what each attendee actually did.