AI Prompting for Validating Assets

Using AI
1

Hey ya’ll!

I’ve been working on few ideas to help get the most out of having AI parse designs for quick, iterative feedback for early prototypes, and had a couple of methods that I was experimenting with:

  • Chain prompting - splitting up tasks into smaller requirements
  • Database of embeddings - taking needed contexts and having them pre-embedded
  • Context overload - providing as many attributes, assets, and data as possible

Here are a couple of findings that I’ve found through personal experience:

1. Chain Prompting Can Reduce Context Overload

This seems to help a lot if you want specific outputs from AI and in a more accurate way.

2. Embeddings

Turning data into compressed, searchable vectors, seems to help also decrease immediate context and give the models more potential for accuracy.

ALSO, getting better prompts seems to correlate with reducing, or at least compressing context.

I’m curious if we should also be somehow training the models, especially if we want them to be design, or product oriented, on something like Glare.

Wondering if anyone has any experience in these areas- I think more insight here would be super impactful!

1 Like
2

Love this and have spent more time this week thinking about how design shapes this prompt work. The overlap is not always clear.

These advanced techniques require engineering overlap, so it’s super fun to learn these new working styles.

Here are two interesting articles on the design side of the equation:

This one is from Alex Klein:

He uses use six categories to structure the prompt design process:

  • Flow
  • Role
  • Mission
  • Guidelines
  • Examples
  • Output parameters

And this one is from Emily Campbell

Here are her heuristic titles:

  1. Purposeful and Needful
  2. Input Clarity and Ease
  3. Result Quality and Context
  4. Customization and Tunability
  5. Branching and Recall
  6. User Autonomy and Control
  7. Logical Transparency
  8. Continuous Learning
  9. Ethical Integrity and Trustworthiness
  10. Identification and Honesty
2 Likes
3

Nice, these articles are very informative.

Funnily enough, they all lead down a similar path- managing context (almost obviously so).

I’ve probably been “too technical” in the thinking, but it seems like being able to open up how people receive the results (with something like a thumbs up button) and give them more control of the input might be useful.