RossAI Blog

Frameworks

Here’s my personal framework ladder for applied AI, based on complexity:

---

1 – Fundamentals – Prompting

2 – Complex Applications – Chaining Prompts, Generative Apps

3 – Agents / Automation – No-code / low-code

4 – Advanced Automation – CLIs, Custom Automations & Agents (pro-code)

---

You might not intend to build an agent or learn to code.

But understanding how AI automations work opens up a whole new way to see problems.

That perspective creates new possibilities.

We’ll get to that soon. But the truth is, you can accomplish a hell of a lot by just getting good at the fundamentals. And you’ll never do the advanced stuff well without mastering the fundamentals.

In boxing, its the jab. In LLMs, its prompt structure.

So here are some of my thoughts on getting the most out of your prompts.

---

LLM Structure

Caveat here – I think ‘Prompt engineering’ is going to matter less as the tech gets better.

Big AI is working tirelessly on making every new model better at reading our most garbage inputs.

But there are some fundamental concepts that have staying power.

Here’s my TLDR for LLMs in November 2025:

• Role Prompting – almost always worth including; eg “You are a McKinsey consultant”.
• Return Format – where you want to spend your energy for complex prompts.
• Context Management – LLMs have memory limitations. If you pass the limit, they blow up.
• Hallucination and BS – LLMs make stuff up. You have to anticipate and mitigate.

---

Role Prompting

Role Prompting is so effective because it gives the most useful context with the fewest words.

“Role: creative director”.

“Role: tax accountant”.

Two very different roles. Two totally different perspectives.

Different skills, ways of speaking — and totally different knowledgebases for the LLM to access.

Role Prompting is a shortcut for the LLM that immediately orients so much context for what its doing, how it should process, and what language it should use in its response.

Its the best bang for the buck.

So it’s almost always worth assigning a role at the beginning of your prompt.

When you tell the LLM “Role: Financial Analyst” in 3 words you’ve already conveyed a whole world of expertise it should invoke in whatever response follows.

I’m far from the only one to talk about this, so here are some good resources straight from the horse’s mouth.

• Anthropic
• OpenAI
• Google Gemini

---

Return Format

LLMs are incredibly good at guessing. But they can’t read your mind.

They work best when they understand exactly what output you’re trying to fit.

The more you can detail the exact return format, the happier you are going to be with the results. Specificity of return format is like superfood to LLMs.

You don’t need a fancy elaborate prompt structure, you just need to be specific.

Prompting “give me a 5 page report” is far superior to “give me a detailed report”.

You’re helping cut out the guesswork, and it will focus more energy on getting the right content.

Asking for results in a table format with defined fields is even better.

Here are some examples.

A)

“Read the change history in this document and give me a 5 page report on what has happened over the last 7 days. 

B)

“Return the results in table format with the following setup:

Property Name | Year 1 | Year 2 | Year 3 | Year 4 | Year 5
Net Operating Income by year | 
Net Cash Flow by Year  |

C)

“Score the following responses based on the Categories I have defined above & Produce strict valid JSON matching exactly the following structure and keys: 
{ "question_content": { "q1_why_acquisitions": "", "q2_invest_100m": [], "q3_new_deal_criteria": [], "q4_strengths_weaknesses": [] }, "question_approach": { "q1_why_acquisitions": { "communication_clarity": "", "conciseness_efficiency": "", "specificity": "" }, } }” 

---

Context Window

They say brevity is the soul of wit.

And so it is with AI. Pruning makes better prompts. If you don’t need the word – cut it out.

Just like any good copy editor would tell you, every word costs money.

LLMs are still fundamentally limited by their “context window”. This is another way of saying there’s a limit to how much information they can store in their memory before they start to break down.

Have you ever done a really long LLM chat where the responses started to get worse and worse? That’s because you’re pushing the context limit.

The bigger your LLM conversations get, the more more the LLM will hallucinate.

Every word in a prompt or an artifact you’ve uploaded to an LLM takes up context.

That means there is a sweet spot between giving too much information and not enough information.

Good reference data is essential to get the answers you need, but if you provide too much data, you’ll overload the context window and you’ll start to get poorer results.

So you want to keep LLM conversations lean.

Don’t go off on a million tangents in one megathread. Start new a conversation for each new topic.

Often I’ll run a two step process: first I ask an LLM to process a large data set and come up with a consolidated analysis / summary.

Second, I take that summary to a fresh conversation, where the next LLM can leverage the data summary without having to burn through all its context on the raw data.

This two-step approach can dramatically improve results if you’re dealing with big documents or large spreadsheet like datasets.

One of the easiest hacks to manage this is the “handoff prompt”.

If your conversation is getting long-winded, here’s a prompt to transition to a fresh conversation:

“prepare a detailed 1-3 page handoff prompt explaining what we’re doing and what we still have left to do, along with any important context so I can continue this conversation in another chat.”

Hallucination & BS

AI lies with high confidence and zero remorse. Tell it to cite its sources.

Ask it to double check every important data point against multiple sources,

Ask it to provide a confidence rating on the accuracy of the information. This seems to trigger some statistical analysis neurons in the machine that help it call out its own BS before it gets further.

Cross-check your work by putting content from one LLM into another.

For example, take content from ChatGPT into Gemini.

Tell Gemini “fact check this response from ChatGPT against live sources, verify all information via multiple sources, and cite references in your response to me”.

It will go HAM looking for mistakes and they seem to love telling on each other.

---

LLM Do’s & Don’ts

Don’t	Do
Assume the first response is the best response	Expect to correct and prod the LLM a little bit
Assume the LLM is 100% accurate	Cross examine. Ask the LLM to fact check itself.
Assume the AI can read your mind	Provide a detailed explanation of how you want the AI to answer your question (”Return Format”).
Try to ‘one-shot’ solutions to complex multi-step problems in a single prompt	Break up complex tasks into multiple discrete prompts that build on each other & chain them together
Continue iterating forever in an endless conversation that blows through your optimal context window	Ask the LLM for a “handoff summary” or “transfer prompt” to start the conversation with a fresh context window
Give the LLM too little context or too much context	Find the goldilocks window where you are giving all the information it needs, but not added fluff or redundant data.

---

Imperfect Genius

One of the biggest things about prompting success is mentality.

Working with LLMs means working with an imperfect but incredibly powerful toolset that requires some persistence & effort to get that value back.

AI can’t read your mind…yet. So until we all get neuro-linked, you have to articulate.

LLMs are also Lazy by Design. They’re programmed to save energy and calibrated for low effort inputs. Be prepared to say “that’s good, but keep going”.

Think of LLM tools like a PhD level intern: incredibly bright, fast learners, but sometimes lacking the common sense earned through experience.

Working with a person like this, you would be aware of the need to provide a little more guidance, and that they might occasionally sound persuasive while being totally off-base.

But given a few extra nudges, they can create some impressive results. Basically AI tends to be book smart, but not street smart. Proceed accordingly.

So you should expect the first pass will often come back incomplete, ill-conceived, or simply half assed.

This doesn’t mean the LLM is incapable, it means it wants you to confirm its on the right track before expending more energy. Which coincidentally is the same thing interns tend to do.

---

Thanks for reading and hope you got something out of this!

Next week, I’ll share some of the generative AI apps I follow and use every day.

Have an AI application you want me to cover?

Let me know in the comments or shoot me a DM.

---

References