A hyper-realistic image of a hipster man with a beard and colourful dreadlocks next to a robot with exposed mechanics and a teal mohawk. Both are eating ice cream, standing next to an open mailbox with an open package enjoying their success. Background is a relaxed modern, sunlit office space with potted plants, furniture and floor to ceiling windows, evoking a sense of casual communial space.
Collections: Can its AI agents make ‘ice cream in the mail’ a reality? Created with Microsoft Designer.

Perplexity AI: Collections as Mini AI Agents

Brian W Reaves
8 min readSep 5, 2024

--

UPDATE: Perplexity AI made 2 changes: 1) Renamed Collections to “Library” 2) Created a new section called “Spaces”, which now is used for the example below.

From Search Engine to Answer Engine with AI Agents.

DISCLOSURE: Article written with AI-assistance for grammar, spelling, readability, and/or fact verification — <10%. The only AI generated image is the header.

If Google is a search engine, then Perplexity AI is an answer engine! I’ve been using it for nearly a year now, and it’s a powerful AI once you know how to use it. Just tell it what you need and your genius intern gets to work. It performs the needed searches, reads through the links a search engine returns, summarises what you’re looking for, and provides the answer with reference links. Now you have the ability to verify it or dig deeper on your own.

For recurring or complex task, throw in one of its most powerful feature: Collections. Now you’re upgrading your intern to your very own AI agent to perform those tasks on command. You outline a task, including how you want it completed, what you want provided in its response, and save it.

My AI Agent Team

I have ≈20 AI agents (collections). I may use 5–8 of them each week. Here are a few of my top agents:

  • Comparison Table Agent — Researching and testing 20 tools myself ‘ain’t my bag, baby…’
    PROMPT: Create a table comparing the top 5 AI video automation tools with feature01, feature02, etc. List them by highest user rating, include columns for: Product (as a clickable link), ValProp (its value proposition, how it differentiates itself), Price (in AUD), Rating
  • Promo Code Agent — Because who wants to pay full price… launch an agent to find a discount
    PROMPT: Search for 5 promo codes less than a month old for Widget AI Video. Verify they work and list them in a table with the headings: Code (as a clickable link where you found it), Discount, Age
  • Legal Council Agent — Reading T&C, Privacy Policy or similar, I have ‘people’ for that…
    PROMPT: Review the attached subscription T&Cs to Widget AI Video and create a table of significant red flags I need to be aware of before agreeing.

Let’s Build Your Agent

Ready to upgrade from an intern to an assistant? Let’s build one of my highly skilled assistants, “AI Prompt Agent”. This agent will improve prompts, which in-turn returns better responses.

In theory, AI should be better at writing its own prompts than you, but in reality, it depends on how well it understands what it is meant to do. Essentially, if I need to talk to a robot, maybe I should ask the robot how to use robot talk.

Step 1

  1. You’ll need an account, fortunately Perplexity AI offers FREE accounts that include Collections.
  2. Once logged in, click on “Library” in the left column. Yes, they have taxonomy issues, and much more, as with all AI tools.
  3. In the right-hand column, next to Collections, click the “+” icon.
  4. The “Create Collection” modal appears with a Title and Emoji fields, which are using good taxonomy, so fill them in.
  5. Then there are 2 other fields, Description and AI Prompt followed by the Privacy setting. This is where the magic happens when creating your AI agent. We’ll complete them in Step 2.

Here’s a screenshot of mine with a title and emoji for its kick-ass dance moves:

A screenshot of the “Create Collection” modal. The collection is titled “AI Prompt Agent,” and an emoji of a person dancing is selected. The description field and the AI prompt field are blank.  The privacy setting is set to “Shareable,” and there is a blue “Create” button at the bottom.
A screenshot of the “Create Collection” modal with the titled and an emoji completed.

Step 2

Before we begin, a PSA to the Millennial/Zoomer readers. You’ll notice lots of little “things” and spaces near words & numbers. They’re called punctuation and formatting. Turns out they’re actually important in communication, not just for humans but robots too. So for best results, mimic the details.

First is the Description field. Just copy/paste the following:

Please use the following variables: 
VARIABLES
• role = professional AI prompt creator ||
• nn = 3 (number of versions) ||
• uses = are optimised for an improved AI response ||
• intent = used to reduce errors by the AI as we work on tasks together ||
• prompt = <insert_the_prompt_here> ||
• other = Otherwise, you may begin. OR (any other details/instructions
you wish to include) ||

NOTE: The pipe “||” characters are for visual references. We’ll removed them late, which I’ll explain more a bit further down.

For the AI Prompt field, copy/paste this:

Use the variables provided to complete the following instructions. 
Acting as a [role], review the draft prompt provided and create [nn]
alternative version that [uses].
The improved prompt will be [intent].
Here's the draft prompt:
"[prompt]"
Do not action the draft prompt I just provided, only provide improved
versions of it.
If these instructions aren't clear, are nonsensical or lack enough
details to successfully complete this task, please ask questions before
beginning. [other]

Last step, set the Privacy preference and click “Create”. The modal closes and your AI agent is ready to use.

Putting Your AI Agent to Work

Like me, you may not be a developer but you’re probably familiar with algebra, back when you were hitting puberty, and they both use variables. In our collection we defined the variables in the Description field and our AI agent uses them to complete the AI Prompt.

Essentially, doing this creates a flexible and reusable tool without having to recreate the prompt each time. Now on to Step 3 to put your agent to work!

Step 3

On the Library page click on “AI Prompt Agent” on the right and its page will open. It will look something like this:

A screenshot of a newly created collection titled “AI Prompt Agent” with a purple dancing person emoji as the icon. The left panel contains options like search, threads, and more. The main section shows a field to start a new thread with a “Focus” toggle and an attachment option. The right panel has a description field with detailed instructions for the AI using variables like role, nn (number of versions), uses, and intent.
A screenshot of a newly created collection titled “AI Prompt Agent”. The right panel has a description field with detailed instructions for the AI.

Copy everything under Description and paste it into the “New Thread” field. Here’s where things kinda suck… as you see, the original formatting we used isn’t honoured, so you have to manually move each variable to a new line. This is why I include pipes, to help you see where new line should be. Once reformatted, delete the pipes. With your list of beautifully formatted variables ready, on to Step 4.

NOTE: By now you’ve probably used enough AI chats to know pressing Enter sends the message (still can’t believe that isn’t an option). Use shift+Enter for new lines.

Step 4

We are finally ready to customise our agent’s task by setting variables. For this example I only added a draft prompt and remove unneeded copy:

Please use the following variables: 
VARIABLES
• role = professional AI prompt creator
• nn = 3
• uses = are optimised for an improved AI response
• intent = used to reduce errors by the AI as we work on tasks together
• prompt = "Be a business consultant and make my business idea of,
'getting ice cream in the mail', so investors love it!"
• other = Otherwise, you may begin.

Before we continue to Step 5, here’s a screenshot of the variables ready to go:

A screenshot of the “AI Prompt Agent” collection. The thread field includes detailed AI instructions using variables, one per line. A blue arrow indicates the message is ready to be sent.
A screenshot of the “AI Prompt Agent” collection of detailed AI instructions, correctly formatted one variables per line.

Step 5

The AI Prompt Agent will replace the variable names with the variable values, resulting in the actual prompt Perplexity AI used to complete the task:

Use the variables provided to complete the following instructions. 
Acting as a [professional AI prompt creator], review the draft prompt
provided and create [3] alternative version that [are optimised for an
improved AI response].
The improved prompt will be [used to reduce errors by the AI as we work
on tasks together].
Here's the draft prompt:
"[Be a business consultant and make my business idea of, 'getting ice
cream in the mail', so investors love it!]"
Do not action the draft prompt I just provided, only provide improved
versions of it.
If these instructions aren't clear, are nonsensical or lack enough
details to successfully complete this task, please ask questions before
beginning. [Otherwise, you may begin.]

With that understood, Step 6 to finish up.

BONUS: AI prompts with correct grammar and spelling will improve responses from AI. When you’re adding variable values, make sure they form logical sentences. Meaning, they should read as if they weren’t composed from variables.

Step 6

Now just put it to work! Press Enter or click on the right arrow icon and your agent is doing its thing, which is essentially Perplexity AI creating 3 enhanced prompts. No guarantees one of them will catch the eye of an investor for your ‘ice cream in the mail’ business idea, but you should have one heck of an initial concept analysis.

Here are my 3 awesome prompts in my Ice Cream in the Mail AI Prompts.

So, That’s It?

Well, yeah pretty much. I mean yes you did all that just to receive 3 enhanced prompts, but what you’ve really done is create a reusable tool to help with future prompt optimisation. Creating more advanced AI agents is where the real value lies. Collections like my Promo Code Agent or Comparison Tables Agent or whatever you dream-up.

Technically, our AI Prompt Agent isn’t a full-on AI agent, but it does offer insight into a simple to implement tool that mimics them. This is important as AI agents are the future, in terms of our personal and professional lives. All of us will need to learn how to use AI agents to remain relevant, regardless of our profession.

You’ll be hearing lots about:

  • Single AI agents
  • AI agent swarms and AI agent teams
  • MoE (Mixture of Experts) and MoA (Mixture of Agents) — want to be one of the cool kids, mention that at your next party

I suspect it won’t be long before we transition into roles like, human AI agent curator or AI agent manager. Eventually, AI won’t need oversight, but for now learn what you can about them and continue to provide value to your employer.

But Wait, There’s More…

Perplexity AI has heaps more.

Standard plan (FREE) in addition to Collections:

  • Browser plug-in for Chromium browsers to ask questions about a webpage or an entire domain
  • Summarise a webpage instead of having to… you know, read it
  • Profile customisation to “Introduce yourself for personalised answers. Share any information or instructions that the AI should know… to be more useful to you.” i.e. Talk to its audience.

Professional plan (≈US$20/m) when you’re ready to upgrade your assistant to a consultant:

  • “Pro Search” is your copilot with Q&A (your genius BFF)
  • Unlimited document uploads for RAG (Retrieval Augmented Generation) — another hit at parties
  • Multiple LLM options (GPT-4o, Claude-3, Sonar Large (LLama 3.1), and more)
  • Multiple image generators (DALL-E, SDXL, and more)
  • And $5/month in API credits

P.S.

If you’re keen on the Professional plan leave a comment and I’ll share a referral link good for $10 off your first month. In the interest of not SPAMing everyone I’m not going to add it as part of the article.

--

--

Brian W Reaves
Brian W Reaves

Written by Brian W Reaves

UI ≠ UX - Brian W Reaves is a Senior UX Researcher | Leveraging AI to Enhance UX Research

Responses (1)