One of my most rewarding efforts working with AI has been teaching it how to teach me. `` When I write, I try to thread the reader’s attention through the text. I’m mindful of how much I’m demanding of the reader at any moment, how hard a concept is to grasp, and what prior knowledge the reader may or may not have. I try to vary my sentence lengths but keep them overall simple and self-contained, and I like to build up from simple to more complex ideas as I progress.
AI, on the other hand tends to wordvomit. It will unload paragraphs on you. It doesn’t feel like it has a conception of you as a reader, and it will throw many concepts at you at once. Of course, this is subjective, but I often get lost, overwhelmed, or mentally fatigued reading AI writing.
I tried to distill the antidote to tedious AI explanations. What I wanted was for AI to hold my hand and walk me step by step through whatever I want to learn. I thought about math textbooks, which introduce small ideas and slowly build them up. Each step builds directly on previous steps. This format traces all the way back to Euclid’s elements:
Book I.
Definitions.
I. A point is that which has no parts.
II. A line is length without breadth.
III. The extremities of a line are points.
IV. A straight or right line is that which lies evenly between its extremities.
V. A surface is that which has length and breadth only.
VI. The extremities of a surface are lines.
VII. A plane surface is that which lies evenly between its extremities.
Each step only introduces one piece of knowledge, and each step is numbered, so it’s easy to refer to it later. I wrote a skill that takes any context and explains it with the same principle.
To help write the skill, I used Anthropic’s skill-creator and Matt Pocock’s writing-great-skills. It was helpful to get started using the knowledge of people who know what they’re doing and then to tweak it myself afterward. skill-creator suggested the language of a ladder:
Teach the subject by building a ladder: a numbered sequence of single-sentence statements where each rung adds exactly one new idea that rests on the rungs below it. The reader climbs one step at a time, never having to hold too much in their head at once, and never having to jump ahead to understand the step they are on.
I won’t go through the skill line by line, but I’ve published it here. I do want to talk about some of the “features” of the skill.
- The line numbering makes it easy to ask AI about a specific point or ask for more information. For example, you could say “rung 8: explain the standard”
- When multiple rungs elaborate on one previous rung, they appear as indented children (“sub-rungs”). I added this after a few weeks but it really helped provide structure.
- The skill instructs the agent to stop around every ten rungs to summarize the points that came before. This isn’t instructed in the skill, but Claude will often look ahead in these resting points (“Which raises the obvious question”)
- Headings introduce each section. This was also a more recent addition to help with readability.
It’s interesting to see the difference between how Claude and GPT write these explainers. Fable 5 is more more verbose than GPT 5.6—in this example, Claude’s sentences averaged 25.6 words to GPT’s 16.6. Claude seems to be cramming more information into the skill’s constrained rungs.
I could more tightly control this - mandate a maximum sentence length, or create a second pass that divides a long rung into two. I originally had each rung be 1-2 sentences before constraining it to 1. But I like the way GPT 5.6 and Fable have each interpreted this, and I appreciate having both versions. GPT’s version is arguably closer to the spirit to the skill, but I find myself preferring the more elaborative but still focused sentences of the Claude version.In addition to learning general concepts, I’ve used this to understand PRs in progress. I’ve been working on an internal wiki and this skill has been instrumental in filling the gaps in my understanding of our codebase and infrastructure. This skill is primarily built for lightweight single-use explainers, but I’ve also used it to compose multi-page explainers, telling the AI to create related explainers and link between them using a nav bar.
After exploring code tours in March, this is my new favorite way to mitigate cognitive debt. I’ve made AI talk in a way that’s much more enjoyable for me to read. I’ll keep tweaking this skill–I don’t think it’s in its final form yet. I hope I’m not the only one that finds it useful!