Update prompting-guide.mdx (#2758)

Co-authored-by: Shuchang Zheng <wintonzheng0325@gmail.com>
This commit is contained in:
kati3-design
2025-06-20 15:42:11 -04:00
committed by GitHub
parent 27add941ff
commit ae9f288ebf
2 changed files with 72 additions and 6 deletions

View File

@@ -70,7 +70,7 @@ navigation:
path: getting-started/quickstart.mdx
- page: Skyvern In Action
path: getting-started/skyvern-in-action.mdx
- page: Task Prompting Guide
- page: Prompting and Troubleshooting Guide
path: getting-started/prompting-guide.mdx
- section: Tasks
contents:

View File

@@ -1,5 +1,5 @@
---
title: Task Prompting Guide
title: Prompting and Troubleshooting Guide
slug: getting-started/prompting-guide
---
@@ -8,11 +8,11 @@ slug: getting-started/prompting-guide
## Choose the Best Block Type
- Most deterministic: Action block
- single action like click, input text, etc.
- Happy medium: Navigation block
- **Most deterministic:** Action block
- single action like click, input text, etc
- **Happy medium:** Navigation block
- single goal, like your goal is to fill out the form
- Most flexible: Navigation V2 block
- **Most flexible:** Navigation V2 block
- multi-goal, like your goal is to login, then fill the form, etc.
@@ -42,6 +42,8 @@ slug: getting-started/prompting-guide
- **Make an action list unless absolutely necessary.** If needed as extra details, make sure its accompanied by a main goal
- **Name the HTML element for items.** Some HTML element items are accessible, but not all. Wouldnt recommend this as a first course of action!
<br>
### Examples
**Example Navigation/Task Block prompt:**
@@ -64,3 +66,67 @@ Now that you have the location, your goal is to go to google maps and extract th
**Example Action Block prompt:**
Click the continue button
<br>
# Workflow Troubleshooting Guide
## Process
Sometimes we test things and they run perfectly. Huzzah! Other times, issues crop up. The process of overcoming these issues by identifying errors is **troubleshooting**
There are a two main reasons why a workflow might fail
- prompt issues/configuration issues
- bugs
<br>
## When you get a Failure or Termination
### When you need to make a prompt change
Common signs that indicate a prompt change
- **Completed too soon:** the task wasn't finished, but the agent thought it was finished and moved on to the next block too early
- **Didn't complete soon enough:** the task finished but the agent didn't realize it was finished and kept taking actions on that block
- **Payload issue:** The information provided may be incorrectly structured or incomplete. This could mean:
- The format needs to change: JSON parameter, string parameter, breaking a parameter into smaller parameters
- Content is missing or invalid: check the payload to see if the parameter is missing or mislabeled
- Parameters might not be configured properly: does each parameter have a double bracket {{sample_parameter}}? Does the label in the bracket match the label in the parameters tab?
### When youre unsure
What do you do when youre unsure:
1. Identify where the problem is happening
- **Check the screenshots with every step:** is everything happening as you expect it to?
- **Check the recording:** is there something missing in the screenshots that can explain why it failed?
2. Identify the problem
- **Check the parameters of the problem task:** Are the inputs what you expected them to be? Do they match whats in the workflow?
- **Check diagnostics for the problem task:** Is the navigation goal correct? Were elements mapped properly? Do they “exist” according to the bot? What actions in the action list did it say it took?
### When you need to file a bug
- Usually, youve tried making multiple prompt changes and the error persists
- Sometimes its clear from the gate that there is a bug
- Issues with common bugs like dropdowns, like checkboxes, etc.
- Action lists referencing the wrong element or bounding boxes not mapped on properly
<br>
## When you get a Complete
### If it aint broke dont fix it…
- First, celebrate
- Check **reliability**: run again and again, are any errors popping up on subsequent runs? Add guardrails, change payloads, refine completion criteria
- Check **accuracy**: is the content correct? Improve prompting and add validation blocks to check the content
### …or maybe fix it
- If you have a **content-sensitive workflow,** plan for edge cases ahead of time
- What happens if I end up on a different screen? Set termination criteria preemptively
- Add validation blocks to check a review page and make sure the content matches. Make sure to pass in the payload to check!
- **In ForLoops,** it is especially important to add guardrails since each block builds on the last
- For instance, if you are clicking into each item in a table, consider passing in termination criteria that stops the task in the case youre not looking at the correct page. Pro Tip: set continue on failure and a first block so it continues even if it messes up one!