Manus High Effort
Banked up a chunk of credits, so I’m running a Manus High Effort mode prompt for the first time. Watching it mirrored on tv. So satisfying. I gave it some time so I could think of a good use case first.
Choose Complexity
I chose a complex task to figure out, that involves mission critical information, and complex regulations, and complex data structures. So, it’s a high value, high complexity ask.
Think Meta Level Synthesis vs Boxing It In
I'm not going to talk about the content (although it would be amazing), but I can talk about the techniques I use to get the most out of a tool like this - the techniques I use to try and make sure I'm not limiting my thinking when interacting with an agentic LLM that has access to knowledge unforeseen.
In this case, it involves asking meta level questions, and meta level prompt thinking (vs trying to create a specific piece of code for example).
I want to leveredge that level of synthesis and build on it, rather than assume I understand what it will synthesize. Because, in this case, I don't want to box it into some output that is much less useful than what it is capable of just because my prompt chaining was not strong.
Use the Output to Get the Improved Prompt for Next Time
After it gave the initial output (which is very comprehensive and useful), I then asked it for the improved initial prompt. I want to account for the portability of this ask and the shareability. I want to make sure I can just reuse it in many other contexts, and potentially in collaborative ways in the future. It's just a habit, really.
I usually ask for the improved initial prompt at some point in the conversation under the assumption that it knows best how to communicate with itself. I want to use it's language. Also, these can vary from model to model. I also like to templatize things for reuse in general if they are complex.
So, asking for the improved prompt instruction insures that I prompt for this particular output in the best (which is context specific) possible way for Manus to execute this task in the future.
The idea is to leveredge what was learned, as it executed the first prompt, and then use that understanding to create the prompt with the best instuctions for use next time I want to ask for this output.
Ask What's the Most High Value Thing I Can Ask for Next?
Now that it has completed the inital prompt, I can ping questions against it and build on it.
(Knowledge quantity limit reached.)
Dispite the knowledge quantity limit reached, I am going to proceed and choose what knowledge to delete from previous runs as I go, so that I can keep this knowledge pertaining to the current run.
A Side Note on Limiting Crawl Access to Mission Critial Information Online
Something I have noticed in this run is that the thing I was working on can help humans. But, some sites try to block tools like this. It's an interesting case to think about. Also, sites that are monitized by ad impressions, are they counting the crawing of these cognitive agentic tools? Bot doesn't have the same meaning it used to in terms of site traffic. That needs a sophisticated id handshake. Probably exists. I saw a person promoting a talk about crawlers that are not detectable. There's a lot to unpack under that topic. Maybe I'll do it in a future post.
High Value
Contextual inquiry, surfing the high value, next most likley, next most effective action... The next "best thing to do" is really context specific, but the idea is that you are given this cognitive model tool that you may be familiar with or may not be familair with. Yet, to get value out of it, you need to unlock it's potential. The "next best action" technique is helpful to leveredge the model understanding.
For contrast, it's kind of the opposite of "explain it to me like I'm a beginner" sort of thing where you just want the AI to output simple to understand information.
Summary using Perplexity Deep Research
I ran my notes through Perplexity to see what gems of knowledge I could pull out of it.
Then I published it as a Perplexity Page which you can interact with here:
Advanced Techniques for Maximizing AI Agent Performance in Complex Workflows https://www.perplexity.ai/page/advanced-techniques-for-maximi-Ue3agAg9QmaIILVBJbo39g