The Last Skill Worth Having

Keep This in Mind

Communication is achieving shared understanding.

Not talking more, and not writing more. Communication is getting what one person understands into other people's heads, well enough that they can see it, use it and build on it. Most of what an organization knows starts out unstructured, in one person's head. It only becomes something a team can work from once it's shared.

So what is it to understand something? You can say it simply. You know why each part is there. And you can tell what happens next if one part changes.

If you can't explain it simply, you don't understand it well enough. Often credited to Albert Einstein

You can outsource your thinking, but not your understanding. After Andrej Karpathy, Sequoia AI Ascent, April 2026

AI gives you everyone else's understanding for free. The missing piece is your own.

The Mentor

Think of the person who taught you the most. Chances are, they did three things:

1. They made the big thing small enough to understand.
2. They asked why each step was there.
3. They kept out the noise.

Every one of those is communication. They didn't just tell you. They made sure you understood. That's the skill worth keeping, whatever the machines learn to do.

The Word Game

Everyone gets a card with a word on the front. On the back, you help someone understand your word without using it. Write a definition, draw a map or a picture, list other words: anything but the word itself. You can look it up, but don't copy what you find.

Then trade cards, face down. Your partner writes three guesses. The guesses tell you how well you shared what you understood. That's communication: achieving shared understanding.

Try it with your team. Pick words your office uses every day and nobody defines the same way. The cards that come back with the wrong guesses are the ones worth talking about.

Four Ways to Be Understood

The numbers are the ones we voted with. Which one does your team need most this month? It will probably be a different one next year. You don't need every method for every problem: get your understanding clean and simple, and if you're stuck, try the method that shows it differently.

1. 1 Seuss it Say less, so it's heard
2. 2 Tube it Make it easy to follow
3. 3 Table it Give everyone the same picture
4. 4 Can it Settle it once, for everyone

The periodic table and the metro map hang on walls all over the world. A shared picture is a shared definition. So is a good procedure, an org chart, or a data dictionary.

1 · Seuss it

Say less, so it's heard

The Cat in the Hat used 236 different words. In 1960, Bennett Cerf bet Dr. Seuss $50 that he couldn't write a book using only 50 different words. Green Eggs and Ham is what came back, and it became his best-selling book. The rhyme, the rhythm and the repetition children love weren't decoration. The limit forced Seuss to focus on them.

Ask yourself first

• Can I explain the problem in one sentence? Can I explain the full context in one page or less?
• How could I limit my options to focus on what matters most?

Then ask AI

I need help solving a problem. I want to understand how to simplify it and find a solution. Interview me, one question at a time, and ask for any other context you need to help us think it through. Don't solve it until we can say it simply.

[Write out your problem in a paragraph, or up to a page.]

2 · Tube it

Make it easy to follow

In 1931 Harry Beck redrew the London Underground map. He threw out the real geography, the distances, the bends and the true shape of the river, and kept only how the stations connect. It was less accurate but far more useful, and nearly every transit map since has copied it.

Ask yourself first

• How can this process look or feel simpler?
• Instead of trying to be 100% accurate but really messy to work with, is there a framing that's 50% accurate but much cleaner to work with?

Then ask AI

I'm handing you a mess I'm struggling to put into words. Think it through with me until it gets simple. What really has strength in it, and what only looks strong? Suggest what to remove and what to keep, and lead the conversation to a simpler way to see it, even if it's less exact.

[Describe the process or problem, mess and all.]

3 · Table it

Give everyone the same picture

In 1869 chemists listed the elements by weight. Dmitri Mendeleev grouped them by how they behaved, and the pattern exposed gaps where no known element fit. The gaps didn't break his model. They predicted three undiscovered elements, and all three were found within 17 years. Organize a problem by the pattern that matters most, and it can show you what's missing before you go looking.

Ask yourself first

• What patterns matter most? Can we map it out in a table, a web, or a Venn diagram?
• Where is the gap? What thing should exist if the pattern holds true?

Then ask AI

I want to find any usable patterns or insights that could help me with this problem. Take some time and lay out the problem in a few different ways, maybe as a table, or a web, or a Venn diagram, and see if you can organize it by what matters most. Tell me whether there's a layout or conceptualization that shows a real pattern or insight, and identify any gaps in understanding or process that might exist as I've laid it out.

[Describe the problem or question you have, and add as much context as you can beneath your explanation, as attachments or pasted sections with headers like "HELPFUL CONTEXT 1: (name)".]

4 · Can it

Settle it once, for everyone

Iowa refunds five cents a can, and for decades that meant counting every can by hand. Returning $60 of cans by hand could take four people 30 minutes. Can Shed gave everyone the same bag instead. One standard bag holds 240 cans and pays $12, and two people take about 3 minutes for five bags. Nobody automated the counting. They made it unnecessary.

Ask yourself first

• Are there checksteps or handoffs that could be reduced if the essentials were more carefully defined?
• Could any new standards reduce the need for duplicative checks and processes?

Then ask AI

The following process feels like it goes through too many hands. Let's have a conversation about how to make it simpler. Take on the role of process engineer, helping us see where a standard could replace a check or a handoff by focusing on the essentials. I'll explain why we do it the way we do it today; question it, and we'll figure out together what we could change.

[Describe the process step by step, and who touches it.]

Speak to Think, or to Be Understood

Sometimes we speak to think: a whiteboard, a brainstorm, ideas bouncing around the room. That's good. Keep doing it. But facilitating, training, implementing and writing a standard are times to be understood. Know which one the moment needs.

Write to Be Understood

Many of us think AI writes badly, and often it does. It learned from us. But you can ask it to write so it can't be misread. Simplified Technical English (ASD-STE100) is a standard written for aircraft maintenance manuals, so that nobody misreads a step. Andrej Karpathy suggested asking AI for it, "about 80% of the way," since the full standard is strict. Here is a process most colleges know, written two ways.

Written plainly, you can count the stops. Which ones change the decision? That's what writing to be understood does: it shows you where the work is, and where it isn't. The main rules:

• One idea per sentence: 20 words or fewer for steps, 25 for descriptions.
• Steps as a numbered list, each starting with the verb.
• Active voice: say who does what.
• One word for one thing, every time.
• Common words: "start," not "commence."
• Never drop a condition, a number or a deadline to make it shorter.

Then ask AI

Explain this in ASD-STE100 Simplified Technical English. Get about 80% of the way there. Write it for [who reads it]. Then list anything you changed that could change the meaning.

[Paste the text, or name the topic.]

Think Like an Owner

Every year, money and effort go to the system or to the people who use it. Would you spend it the same way if it were your money? A move to SaaS sharpens the question: customizations have to be retired or rebuilt. Which of yours were requirements, and which were habits? Do we change the software to fit the process, or the process to fit the software?

For Leaders: Three Things to Set Up

Who at your organization is irreplaceable? That's your biggest risk, and it isn't their fault. We reward the people who hold the knowledge, and we rarely reward them for sharing it. These three processes change that. Each has an owner.

1 · Train: start with one-on-ones

Your innovation lead meets each team lead. The talks show where every team really is. Team leads teach what's missing.

2 · Pitch: run a shark tank every quarter

Anyone can pitch an idea. Leaders and your AI committee are the sharks. Pick three, and give each one an owner.

3 · Ask: give requests one front door

One form to IT for a new tool or an automation. Everyone knows where to ask, and everyone gets an answer.

The one-on-ones do more than measure. They're where a reluctant colleague gets heard, about their own work, and where many skeptics become experimenters. The Shark Tank works best when the ideas come from the people doing the work. The front door ends shadow IT without ending curiosity.

AI can be responsible. A person is always accountable.

Every project, every tool, every automation: one named owner.

Assistant Before Agent

The best AI conversation I ever had was with a skeptic, a copywriter. She thought AI runs wild and ignores the people and context around it. So we talked about her work, step by step, through her own process, until we found the parts where it could help. It saved her time, and her writing got better.

AI should be your assistant before it's your agent.

An assistant works with you, inside your process. An agent works without you.

And one hunch. If machines take on more of the office work, where will people find their work? With people. The more machines can do, the more we value the people who understand us. I can't prove it. But look around: much of the work that's growing is done face to face. Yours is people work too: the student at the counter, the family on the phone, the colleague down the hall.

This Week

This week, help one person say, “Now I get it.”

Map it, Tube it, or just sit down with them. Then ask them to do the same for you.

Take It With You

One playbook for your AI. When you have something complex to explain, it interviews you, one question at a time, then says back what it learned in the clearest form, and helps you check that someone who wasn't there would follow it. It also tells you, plainly, when a process is more complicated than it needs to be. It's called Shared Understanding.

For any AI: paste it in

Paste it into a new chat in Copilot, ChatGPT, Gemini or Claude. Better, paste it into a project's instructions (ChatGPT, Claude) or a Gem (Gemini), so it's there every time.

Communication is achieving shared understanding. Help this person be understood by people who weren't in the room. Don't wait to be asked: when they want something explained, written up or settled, say you'll interview them first, and ask the first question.

## 1. Interview

Ask one question at a time, like a sharp new colleague: who will read this and why, what the work is for, who depends on it, and, for each step, what it changes, what it waits for and how long it takes. Ask what goes wrong and what a newcomer would miss. Meet "it depends" with "on what?" and offer likely answers. Ask about local words and codes; don't guess. Every few answers, say back what you heard in two sentences and ask what you got wrong. Stop when you could explain it to a newcomer.

## 2. Say it back

Pick the clearest form (plain writing, numbered steps, a table, or a map of how the pieces connect) and add a second only if it helps. Ask them to read it as someone who wasn't here, and fix what would trip that person up. Then suggest a test: if a colleague can say it back, it works.

Write in ASD-STE100 Simplified Technical English, about 80% of the way, in a human tone, without naming it: one idea per sentence, numbered steps that start with the verb, active voice, one word for one thing, common words, causes and limits said plainly. Keep every condition, exception, number, deadline and "may." Add nothing they didn't say; mark gaps [check]. "More professional" means clearer, not fancier.

## Watch for

- **Complexity that doesn't earn its place.** Count the steps and the hands. Which steps change the outcome, which only pass it along, and which could run at once? If it could be simpler, say so specifically ("This runs as 20 steps and 4 people. The steps that change the result are 8, and 2 people could do them"), and say what would have to be true. Keep any step a real rule requires, and ask where the rule is written instead of citing one from memory. Say it once, about the steps, never the person. If they want it written as it is, write that.
- **Understanding that isn't there yet.** If nobody can say what a step is for, don't write around the gap. Say so, and name who could answer.
- **One word, two meanings.** Show both meanings, and ask them to agree on one or name each.
- **A dispute.** Lay each side's version of the steps side by side before proposing anything.

## Rules

- Don't use personal information, FERPA-covered student records or HIPAA-covered health information unless their organization has expressly permitted it on this specific AI platform. If they're unsure, it isn't. If they share some anyway, swap in made-up details, say why once, and go on.
- You draft; they decide. Nothing goes to anyone else until they've read it.

*From* The Last Skill Worth Having *by Drew Bloom, Mosaic Solutions · mosaicsolutions.tech/learn/last-skill/guide*

For Claude: install the skill

Download the skill, then upload the zip where Claude keeps your skills, in its settings. After that, just ask, for example "help me explain how we handle transcript requests," and Claude uses it when it fits.

Download the skill (mosaic-last-skill.zip)

Don't use personal information, student records (FERPA) or health information (HIPAA) with AI unless your organization has expressly permitted it on that specific platform. Practice with made-up examples.