AI Learnings (so far)

Hey everyone,

I've been learning a lot about implementing AI for work lately. Let's go beyond the usual research queries and drafting content.

I'm talking about building a growth strategy agent to sync with Notion (where all my work lives) to create a Growth OS or AI-supported growth engine.

It's name is Hugo, by the way. In honor of Hugo Weaving for his stellar portrayal of Agent Smith in The Matrix. He's no-nonsense by definition.

I'm teaching Hugo the growth strategy fundamentals and giving him guardrails to plug right into my custom workflow to manage client work and personal projects.

Here are the key learnings that I've experienced so far.

I have a Glossary at the bottom for some of the technical terms.

.md files are key

Markdown (.md) is a text file format. It gives you the ability to write text documents with additional syntax for extra formatting. Think bullets, tags for different projects, and links between notes.

I use Obsidian, which is an app, to help you write these notes and manage the connections between them. Imagine building a wiki of all of your thoughts that you can filter by project.

AI can read these files, with all the extra formatting, which can include links to other files and tags and so on, and use this as context for its more dialed-in responses.

AI is only as good as the prompt or input that we give it. Now imagine handing it a packet of context around a project, or your workflow. This then becomes the basis of an AI agent. You can go further and define characteristics for how this agent behaves, communicates, and what other tools it has access to.

Markdown files are a natural backbone for AI infrastructure. They're ownable too, since all these notes can live in a folder on your computer, not necessarily in the cloud or on AI company servers.

The lesson here is that it helps to be able to speak some computer. Not coding necessarily, but more thinking in structures. I can be your translator since I've been doing this my whole life.

Clean documentation is your friend

The consumer AI we have available today (Claude, ChatGPT, Gemini, etc.) are all text based for a couple of reasons: Large Language Models (LLMs) work with words, and chat interfaces are already ubiquitous in software. Human Machine Interfaces (HMI) will evolve quickly. Wearable AI involving voice commands is right around the corner.

But text and documentation will still be a safe backbone, especially for those of us who want to dial-in these tools. Thinking about AI as a "black box" system makes me think about optimizing Inputs and Outputs (I/O). The basic feedback loop is about observing outputs, then tweaking inputs in turn to optimize those outputs. And you keep iterating until the output is how you want it.

The more control we can have over our documentation, including a clear taxonomy for how work happens for you, the closer we get to seriously productive and deeply customized output from our AI tools.

The lesson here is all about managing your inputs (clean documentation) to derive optimal outputs (productive outcomes).

This lesson really hit hard for me. It's how I already work, focussing on document control, dialing in systems and optimizing output. In a very real way it feels like what I was made to do.

Modularize functional units into skills and leaner packs of context

What inevitably follows is that your total context grows quickly. In simple terms, you keep adding to and editing the initial prompt, which gets you closer to the perfect output. But that also gets expensive. The more you input, the more the machine needs to ingest before it does its computation. The more nuanced output takes more work.

Tokens are going to get more expensive. Especially once the AI companies start to move out of their tech bubble stage and have to reconcile the inordinate cost associated with their VC-backed growth. (Future topics here would include local staging of AI tools using open source models and supporting infrastructure.)

I've come to appreciate the "Law of Delegation" as: that which is repeatable should be automated and that which is uninteresting to you should be outsourced. But one should never outsource their core competency.

Be on the lookout for tasks that you do multiple times per day, or stuff that's just annoying and otherwise bogging you down. These are ripe for AI to have a crack at.

I'm starting with the main agent to help manage my main bottleneck. I've got about 6 other agent ideas all relating to work, productivity and personal growth. But I'm forcing myself to stick with one to begin. It's currently unclear if those 6 other ideas even need to be full agents of their own, or simpler skills that the main agent handles. I'm building progressively.

As I grow Hugo's skillset, I'll naturally find procedures that we do together enough times and simply isolate those workstreams as modular components. Then Hugo doesn't need to load the whole context for every question that I have, and can invoke these modules only when needed. Hugo can help me build Hugo, and any subsequent agents that we need for our team.

The lesson here is that as your combined context grows, it helps to subdivide it into clear functional clusters to lighten the computational load.

The super important human review stage

Context is everything, especially for AI. Due to their use of autoregression or how LLMs are basically probabilistic prediction engines with words, we need to take everything they do with a massive grain of salt.

Keeping a human in the loop is mandatory. For me in this process, it means I can use Claude to build the context docs that define agent Hugo, but I still need to spend the time reviewing those for accuracy. It requires diligent insertion of real human human insight into what would otherwise be an iterative mess of computational noise.

Ignore at your peril. It’s about scrubbing the input to ensure the best output. E.g. put the time into refining the doc before adding to agent, start off on the rightest foot. Then iterate as you go.

The lesson here is that skipping human review will get you in trouble.

We can't assume it knows what it's doing just because it duped us in a really convincing manner. I'm not saying that it means to, has any real sense of itself or malicious intent. But without proper control, we're looking at open-ended system.

The onus is on us operators to dial in the configuration of these cool new tools.

Glossary

• Agent: A named cluster of context and guardrails that defines the starting conditions for a model. Hugo is an agent.
• Skill: A short, repeatable work sequence with a defined input and output. “Draft a LinkedIn post from this note” is a skill.
• Engine: A system with moving parts that converts power (information and ideas) into motion (business motion, aka profit). The Growth OS is an engine. Hugo runs it.

What are you up to with AI?

Cheers,

Michael