Taalk Help Center

Tutorial: Fix a Plugin Without Losing the Call

A 15-minute walkthrough of Sample Data. You turn real calls into test cases, see what's covered, test "what if" without placing a call, and check a change against every call before it goes live.

Think of it the way a QA lead works with a human team: keep a file of real calls, write down what should have happened on each one, coach the fix, then replay the calls to confirm it. Sample Data is that file, shared with your whole workspace.

New to Sample Data? Start with Test Plugins and Agents with Sample Data.

What you need

  • A plugin with at least one saved action, used by an agent that has made calls.
  • Permission to edit plugins.

Step 1: Turn real calls into samples (3 min)

  1. Open the plugin: Outreach › AI Agents › Plugins, then scroll to Sample Data.
  2. Click From calls, then Use recent calls. Or paste links to calls you care about, one per line.
  3. Each plugin run is listed with its action, what the live call decided, and Match or Differs. Leave the Differs rows unchecked.
  4. Click Save.

Each sample keeps the data the action received, the AI's answers, the transcript up to that point and what the live call sent. It expects the live result, so there's nothing to type.

Step 2: Read the panel (2 min)

After saving, every sample is checked right away.

  • The status line says All N checked samples pass, or how many fail.
  • Each action groups its samples by the outcome they expect, such as Spoke to Contact (1) and Unable to Contact (6).
  • Dashed outcomes have no sample yet. That is your test gap. Click From calls on that line and paste a call that ended that way.
  • Only 1 example next to an outcome means one call is all you have. Click Try variations there to cover the rules around it.

Step 3: Test "what if" with Quick edit (5 min)

Pick a sample and click Duplicate, so the original stays as it is. The dialog opens on Quick edit.

  1. Under AI answers, change one answer. For example, set voicemail_left to true.
  2. Click Try it. You see It gives Left a Message and whether that matches the expected result.
  3. If the new result is the right one for this data, click Use this result as expected.
  4. Rename it to say what it tests, for example Voicemail left, and click Save.

You now have a test for a case that no real call has shown yet.

Call facts the JS uses works the same way: only the facts this action's JS reads are listed, such as callStatus, durationVM or durationAfterTransfer. Durations are in milliseconds.

Step 4: Let Variations do it for you (2 min)

Click Variations on a sample. Taalk changes each AI answer and each call fact the JS reads, one at a time, and shows which changes move the result. Keep the suggested keys, and click Save as samples for the ones that matter.

Run it on one sample of each outcome. A variation changes one thing, so it only tests the rules near that call's outcome.

Step 5: When a sample fails, find out why (3 min)

A failing sample shows a red mark, a one-line reason and a spot under Needs attention, for example:

UTCReason: expected Answered - Dropped, got Other

  1. Open the sample in Quick edit.
  2. Find what feeds the field that failed. Often it's an AI answer with the same name (here utc_reason).
  3. Change it to what the call really said and click Try it.
What you see What it means What to do
It matches now The sample's saved data was wrong Save the sample
The reason moves to another field Progress: more than one value was wrong Fix that one too and Try it again
It still fails on the same field The plugin is making the decision Check the call facts, then the JS or prompt
The live call itself was wrong The plugin has a real bug Keep the right expected result and fix the plugin (Step 6)

Step 6: Fix the plugin and check every call (3 min)

  1. Click an action to open it, and change the JS, the prompt, the request body or the messages. Don't save yet.
  2. In the Run with Sample Data bar at the bottom, click Run all. It runs your unsaved edit against every sample for that action.
  3. Click any failing row to see its steps, checks and a side-by-side with the live call.
  4. When everything passes, click Save. The change goes live on the next call. The panel's Check all confirms it against the saved plugin.

Watch the JS step's warnings as well as pass or fail. Live calls don't support ?., ??, async, or default values in destructuring, and the preview flags them before they break a real call.

Fixing the extraction prompt

If the wrong value came from the AI step, edit the Extraction Prompt, turn on Run with AI and run again. The prompt runs on the sample's saved transcript. It approximates live calls closely, but not exactly, and each sample is one AI call.

Step 7: Confirm on the real runtime (admins and developers)

In a result panel, click Rerun on source call. This runs the saved plugin on the real runtime with the sample's data. It may send real requests to your webhook, so use it only as the last check.

Agents: reuse contact data in test calls

  1. Open an agent and go to Knowledge and Skills. Add samples under Sample Data with the contact fields you test with (first_name, state, loan_amount...).
  2. Click Test, open the advanced options and pick one in Load sample next to Contact data. Samples from the agent's plugins appear too.
  3. Typed new contact data you want to keep? Click Save as sample.

The habit

Before saving any change to a plugin, click Run all. When a call is classified wrong, import it with From calls, set its expected result to the right answer, then fix the plugin until it passes.

Tips

  • One sample per behavior: qualified lead, voicemail, transfer declined, dropped after hello. Name them that way.
  • Aim for at least one sample per outcome, so no outcome line is dashed.
  • Samples are shared: what you add, your teammates see and run.