Attention Allocation

Our product team is heavily using AI and agentic coding. We use GitHub Copilot for background tasks; Graphite for AI reviews on Pull Requests (PRs); Claude Code for coding, data analysis, and building custom tools.

We’re synthesizing more data, exploring more possibilities, opening more PRs, shipping more features, and delivering more value.

More, more, more, more, more.

Increasing output develops new bottlenecks in the process. We increased output without increasing headcount. The downstream consequences put pressure on the work both before and after execution.

AI can summarize data and explore possibilities, but it won’t make the best decision on what to build.

That’s direction.

And AI can review outputs, but it won’t bring taste, judgement, and nuance to the work. An LLM response is predicated on the data it’s trained on. There’s a reason asking Claude Code to build you an app without style direction ends as a dark-themed app with purple gradients. It’s echoing an amalgamation of what it knows. And what it knows has already been created.

To diverge from that, you need human judgement to review and refine AI outputs.

Humans also develop taste from data we’re exposed to. Using software, reading books, admiring art builds our understanding of what’s possible. The difference? The human experience is messy and unpredictable. Ideas emerge from seemingly random events because our brains crave patterns and connections. AI doesn’t generate novel ideas because it’s not engaged in the real, messy, meaningful process of the lived experience. Humans don’t gain understanding through exposure; we gain it through the process of discovering meaning. It’s cognitive.

What matters is what we pay attention to.

I was asked to help train the sales team to leverage AI tools like Claude. Seeing the product wins piqued their interest. They want the same gains in sales.

I started thinking about a mental model for considering when to use AI; and, more importantly, when to ignore AI and do the work as a human. Where’s the line?

If AI can do so much work now, where should humans actually focus their attention?

I address this question in both Intelligence Orchestration and Human Calibration.

But there’s a bigger idea at play. One that sheds light on how humans can best allocate attention to the right tasks—the tasks combining engagement and expertise.

AI is shrinking execution time. It’s in a constant state of decline. While this is an opportunity to create the future of work, it’s also a moment of existential uncertainty.

Attention is the new bottleneck. AI expands what’s possible, but human advantage comes from deciding where attention goes.

This idea is called Attention Allocation, which lives in the Clarity Codex of the Claritorium and Quality Refinement of Equilio.

The three pillars are:

1. Allocation Matrix shows where human attention should actually go.
2. Intelligence Loop shows how work moves between humans and AI.
3. Human Leverage shows where humans create the most leverage in AI systems.

Allocation Matrix

Where should human attention actually go?

Engagement + Expertise

When considering AI use-cases, I think of work across two dimensions:

1. Engagement: How much you enjoy it.
2. Expertise: How good you are at it.

These aren’t static dimensions. What you like doing and what I like doing may not be the same. And what you’re good at and what I’m good at is different. Your team will also have different aptitudes as a collective unit.

That’s the point!

This is a diagnostic tool to help calibrate in real-time so you can keep adapting. “Survival of the fittest” was never about strength. It was about adaptability. The species most adaptable to change is the strongest.

The Four Types of Work

The matrix created by engagement and expertise creates four quadrants to represent the four types of work:

1. Creative Work
2. Exploratory Work
3. Judgment Work
4. System Work

This is a lens for doing work in an AI-accelerated world that requires Intelligence Orchestration and Human Calibration.

AI is part of every type of work, but it’s expressed differently within each type.

Creative Work

High Engagement + High Expertise

Creative Work is the highest and best use of your time as a human. This is the craft. You should prioritize your attention, energy, and focus for this work. If AI shrinks execution time, then expand your Creative Work.

If you’re a PM, this means shaping the problem space with the team—finding the user problem worth solving.

If you’re a designer, this means exploring conceptual ideas to transform requirements into a cohesive experience.

If you’re an engineer, this means designing the architecture and technical approach.

If you’re a salesperson, this means running a discovery conversation, identifying pain points, and guiding a prospect to a solution.

Creative Work is for human ingenuity, but AI can be part of it. I explore ideas and insights by having ChatGPT reflect back questions. I discuss ideas when they’re still raw, distilling them through rigorous debate. If you’re diligent in expressing your critical thinking, you can escape the sycophantic trap of AI.

The creative process is beautiful. You get to decide what work brings you energy and allows you to focus your skills. Extend that out to your team and you create magic. Most of my work with teams is finding ways to put everyone in this area. You can’t always do this work, but it’s the most important you can do.

Prioritize Creative Work.

Exploratory Work

High Engagement + Low Expertise

The speed of AI is profound. You not only can create something in seconds, you can create multiple variations in seconds.

10 example analogies for a concept.

5 possible solutions for a problem.

3 specific outreach messages.

Divergent thinking is difficult alone. And the cost of bringing enough people in a meeting to create divergent thinking is not cheap. It’s worth it for the hardest problems, so don’t underestimate the value of the collective intelligence of diverse groups. It matters.

But AI is an accelerant for exploration, expanding the space of possibilities.

If you’re a PM, this means generating alternative feature concepts or drafting early versions of specifications.

If you’re a designer, this means exploring multiple UI directions or rapidly prototyping interaction ideas.

If you’re an engineer, this means experimenting with implementation approaches or generating prototype code.

If you’re a salesperson, this means exploring outreach angles or different approaches to land the next deal.

Exploration matters. You diverge before you converge, going wide before going deep. But time and attention are limited resources, so we often have to pick a path too early.

There’s no time to explore.

AI excels at Exploratory Work. If you guide the intelligence using your own judgement, you can simulate several paths before you commit and move forward. Time stays on your side while you explore.

Accelerate Exploratory Work.

Judgement Work

Low Engagement + High Expertise

As a human with unique skills and experiences, you cannot sit by and let your critical thinking disappear. Human cognition is intricate and special. Human Calibration is knowing when and where to flex your judgement muscles.

Imagine if you spent months defining and executing a project only to ignore the output.

Working with AI is the same.

You can’t generate outputs without review, refinement, and a consistent quality bar.

Judgement Work enables you to review AI outputs, engage your critical thinking, and hold the line on quality. Evaluating outputs and making intentional decisions requires deep expertise in what you’re reviewing. If you don’t know about the work, you’re just offering an uninformed opinion. It’s helpful, but it won’t create quality work.

If you’re a PM, this means making sure a feature actually solves a user’s problem.

If you’re a designer, this means reviewing an implemented design and identifying where the user experience breaks down.

If you’re an engineer, this means reviewing a Pull Request and spotting architectural or performance risks in the code.

If you’re a salesperson, this means reviewing a deal to assess whether it’s worth pursuing.

As AI increases output, reviewing work and maintaining quality is critical. And it’s not just the quality; your team needs to share understanding in the work. If you don’t, coherence is lost and quality drifts further.

Protect Judgment Work.

System Work

Low Engagement + Low Expertise

Every Monday morning, I review a set of key product metrics from the previous week and manually add them to a spreadsheet. It’s a lot of opening various source websites, changing date inputs, and logging numbers.

It’s data entry.

I enjoy reviewing the numbers to use them as data to inform priorities. But spending 30+ minutes a week with data entry doesn’t sit high on my list of engaging activities.

System Work is low engagement and requires low expertise. Even before the prevalence of AI, this work was at the top of the list for automation. It’s work for machines.

So I automated it. There’s now a daily digest email with daily metrics, and a script I run to pull the weekly numbers. Now I can spend more time digging into the details. That’s now fuel for Creative Work and Judgement Work.

If you’re a PM, this means writing meeting summaries or pulling metrics.

If you’re a designer, this means organizing files or documenting design decisions.

If you’re an engineer, this means writing boilerplate code or generating documentation.

If you’re a salesperson, this means updating the CRM or scheduling follow-ups.

If you don’t enjoy it—and your skills aren’t necessary for it—then automate it. Building a custom script or automation has never been easier. I have a command I run in Raycast that pulls the latest data from a range of sources, generates HTML files, and opens them in browser tabs automatically. I get beautiful data visualizations in seconds, freshly built from the latest data sets.

The time you don’t spend doing manual work you don’t need to do is time you can spend on Creative Work. That’s where you’re needed.

The Allocation Matrix

Intelligence Loop

How does work move between humans and AI?

The Allocation Matrix is a model. It’s a representation of specific types of work.

But work isn’t static.

It’s a fluid movement of information passing through contexts. How and where the information flows is an intricate dance of human and artificial intelligence.

If the Allocation Matrix shows the types, then the Intelligence Loop shows how attention and intention move work to completion.

• Direction sets the intention.
• Exploration sees what’s possible.
• Execution moves intention to outcomes.
• Review evaluates the results.

Direction

We get a lot of feedback—from customers, stakeholders, and our product team. I use the Linear MCP to funnel context back into the Work Registry. I look for patterns to drive a higher signal frequency. Figuring out what we should work on is still one of the highest and best uses of my time. AI supports, but the team’s collective intuition strengthens the bet.

Direction is intention. Without a clear direction, humans and AI miss the mark.

This is Creative Work. I prioritize it.

One such signal surfaced recently. Missing functionality in a core part of our product came from multiple sources of feedback—both B2C and B2B customers.

I started by defining the intention. This work was tracked on the roadmap, but wasn’t prioritized within our current capacity. So I decided to define the shape of what we wanted to do, and use that to guide next steps.

I wrote a detailed prompt, which is slowing down to think about what you’re trying to do before you do it. The “prompt” language is what we use for LLMs, but the same structure has always worked well for humans, too.

This is the Direction. It’s a clear intention expressed within practical constraints.

Exploration

Without AI, work is tightly constrained by time. Most work environments can’t tolerate endless exploration without execution. It’s good to explore a few different options, but not dozens (or more). Beyond the time constraint, it’s hard to come up with several divergent paths without bringing together a diverse group of thinkers.

AI changes things considerably. You can take your original direction and explore multiple possible paths in minutes. For an engineer to explore multiple implementation paths, it would take weeks or months. With the right prompt, Claude Code can review the code and provide a plan with multiple implementation paths for you to choose from. Technically, you could use concurrent worktrees or Claude Teams to implement them simultaneously.

Exploration precedes execution. You slow down to speed up. You explore wide before you converge and execute with precision.

Once I set the direction, I had Claude Code review the codebase and determine possible solutions. Software is a calculus of trade-offs, so there’s almost always at least two options:

1. Solve the problem faster, but worse.
2. Solve the problem slower, but better.

Temporary relief vs. long-term stability.

In this case, the feature was actually a really simple change. The proposed implementation plan was slightly over 100 lines of code; about half of which were tests.

Execution

Execution is narrow focus on a clear solution.

If you encounter unknowns or new trade-offs to consider during execution, then you return to the intention and explore paths.

AI is shrinking execution. Relatively speaking, it takes a small percentage of time to deliver a working solution of anything. That’s why I emphasize what happens before (direction + exploration) and what happens after (review). If the machine does the work, then you need to shift your attention to inputs and outputs.

Claude Code implemented the plan in a few minutes, I tested the change, and let it open a Pull Request for the next step.

Review

I track our “cycle time”, which measures how long it takes for an opened Pull Request to be reviewed, approved, and merged. The less time it takes, the better. But it’s also a balance because you don’t want to rush work.

Review is the new bottleneck.

I have the coding agent review its own code, following a specific prompt to adhere to codebase standards. I review the code, have the agent review its own code, and then another engineer reviews it. There are automated tests, linters, and accessibility checks all code must pass before merge. And then most work still goes through a human QA process to break it like a user would.

AI shrinks execution and expands output, so expanding the quality of review matters. Like all software, it’s a delicate balance of speed and quality. Hold them in constant tension.

The Intelligence Loop

• Direction → humans set intent
• Exploration → AI expands possibilities
• Execution → systems produce outcomes
• Review → humans evaluate results

Human Leverage

Where do humans create the most leverage?

I work with great engineers. One of them is brilliant. He not only has a masterful understanding of the entire software development stack, he also understands the work decomposition process—breaking down work into meaningful value iterations. This is a skill that separates good engineers from great ones. Knowing how to isolate work into independent slices is an art. It’s what makes or breaks delivery. You can’t understand trade-offs or map unknowns when the work is invisible, scattered, or misunderstood.

This engineer single-handedly built an entire new surface area of our product. He orchestrated the work with Claude Code (using multiple Claude Teams), combining his own skills, experience, and knowledge into compounded intelligence. He used Claude Code and the Linear MCP to help break out distinct Linear projects. He documented where all the work is, what’s left, and how we can execute with rapid precision. He’s going on a much-deserved vacation, so knowledge-transfer was critical for continued momentum.

Historically, humans were responsible for the execution of work. As a software engineer, that means writing every line of code.

But now? Agentic coding can write code faster and with higher precision than humans. It’s like everyone is moving from an individual contributor to a manager, setting direction and reviewing the outputs. The difference? You’re still in the details orchestrating the outputs.

The human advantage is direction and review.

Creative Work and Judgement Work.

Use AI in Exploratory Work to improve the direction. And then use AI in System Work to execute with precision. It’s like letting an industrial machine create high-fidelity physical products based on an artist’s sketches. The human toiled in the Creative Work, while the machine facilitated the automated System Work. The craft remains, directed intentionally across intelligences and leveraged in new ways.

Leverage should change when it’s no longer working. At that point, there’s no real leverage. You must change what you’re doing to adjust the outcomes you aim to achieve.

If you’re a PM writing PRDs and requirements by hand, that’s not your leverage.

If you’re a designer building high-fidelity Figma mockups only, that’s not your leverage.

If you’re an engineer writing the same React component by hand, that’s not your leverage.

If you’re a salesperson updating details in the CRM, that’s not your leverage.

Humans guide the system; AI expands the work. This is where Attention Allocation becomes leverage.

The Throughline

AI compresses execution and expands exploration. That gives you more time to engage with work you enjoy, but also requires discipline in reviewing work critically.

Writing code and designing are isolating activities for me. I devote immense energy to focus and block out all distractions. But it’s never been my only job. I also have to track work progress, triage feedback, and empower the entire team to deliver with quality—along with dozens of other things.

Creating an Emergent Environment requires focused energy. The output compounds as it’s scaled across a team, so it’s worth my time. I can’t afford to solely allocate my attention to writing code or designing. But, with the right system, I can generate the same outputs without the attentional overhead.

That’s Attention Allocation.

Prioritize Creative Work. Focus on setting clear intentions through intentional thinking and high-leverage strategic output.

Accelerate Exploratory Work. Explore a wide range of possibilities to improve the quality of your decisions and solutions.

Automate System Work. Let AI execute on your intentional inputs with speed and precision to validate assumptions faster.

Protect Judgement Work. Be an active participant in reviewing every output to improve your intuition and the quality of work.

Attention is now the scarcest resource.

So where will you allocate your attention?