Your AI assistant doesn't remember yesterday.
IWE fixes that.
There's a specific frustration that anyone who works regularly with AI assistants knows well:
You've had a great session — you explained your project, your goals, your constraints, the language you use, the things that matter. The AI was genuinely helpful. And then you close the window, open a new one, and start from scratch. The context is gone. You're strangers again.
My knowledge base isn't an archive — it's an ASSET!
I use it all the time. And keep it updated all the time. And then I share it with artificial intelligence.
It wasn't always like that.
Before most notes ended up dying quietly deep in a folder structure.
That's not a problem with the notes. It's a structural problem. And that's the problem I've spent the last month solving.
You probably know the situation:
In the world of functional programming, there are two titans of iteration: Unfold (generating a stream from a seed) and Fold (collapsing a stream into a single value).
In Nushell, these are represented by generate and reduce. Having built agents with both, it's clear that while they can often achieve the same goals, their "philosophy" is fundamentally different.
In agent.nu, we explored using generate to build an LLM agent. That pattern is an unfold: starting from a seed, you grow a stream of values. It's perfect for UI-centric agents where you want to stream "thoughts" to the user as they happen.
You can read more about it here: A very simple and slightly better way to create an AI agent using nushell generate instead of a normal agent loop | docujAI
Functional Agent Loops: Why Nushell's generate Beats the Imperative Loop
Building AI agents often involves a predictable but messy loop:
When working with LLMs from the command line, constructing well-structured prompts can be tedious and error-prone. I recently built a Nushell module that leverages custom completions to make prompt generation interactive and consistent.
Here's how it works and how you can build similar tools.
Instead of memorizing prompt structures or copying from a notes file, this module lets you build prompts step-by-step with tab completion. The system supports both English and Danish, with three configurable components:
For a company to truly adopt AI and be transformed by it at the deepest level, it must be process-driven rather than person-driven.
A process-driven business delivers value through documented, up-to-date, and repeatable protocols rather than the "personal expertise" of the founder or specific staff members.
To be ready for AI at a deep levet, you must first perform "pre-AI" work.
This "pre-AI" work include, but is not limited to:
Document existing processes, systematize them, and improve them before AI can be applied effectively.
For a simple editorial workflow, I would almost certainly choose reduce, but the "cool" choice (and the one that feels more like an autonomous agent) is generate.
Here is how to decide based on what your "workflow" actually looks like:
reduce (The "Linear Polisher")Use reduce if you have a fixed set of operations you want to run on a piece of text (e.g., "Fix Grammar" "Shorten" "Change Tone").
In Nushell, reduce is all about taking a list and folding it down into a single value. When you plug an LLM into that "folding" process, you’re essentially giving the model a "memory" of everything it has processed in the pipeline so far.
Here are three high-level patterns for using reduce with an LLM in your terminal.
reduce Works (The LLM Mental Model)Before we dive in, remember the syntax: list | reduce { |it, acc| ... }.
Imagine being able to create graphics in all kinds of popular formats.
Directly from the command line!
I know, that is not the dream of everyone. But then again, you really should try the command line with Nushell.
So many things can be done faster and more consistently that way.
Anyway, here is a small Nushell module I made to work with Google's Gemini 3 pro image preview model. (Google, you gotta come up with a shorter name.)