A Model for Goal Setting

March 14, 2026

The question I keep returning to is not “what should I do” in the advice-seeking sense, but something closer to: what is actually happening when I decide what to work on? What is the mechanism?

This is an attempt to describe that mechanism from first principles. The description is useful not because it tells me what to do, but because it gives me language for diagnosing why things go wrong.

1. Definition of a Goal

A goal has three components: a want (a desire for some state to be different), steps (a series of actions that move toward it), and an outcome (a well-defined state where it is achieved).

Scrolling Instagram is a goal. The want is the feeling of dopamine when I look at reels. The steps are picking up my phone, opening the app, and swiping. The outcome is that I have scrolled. Qualifying for Team Canada is also a goal. The want is the credential and what it represents. The steps are years of structured training. The outcome is making the team.

These two things have identical structure. There is no separate category for “distractions” or “wasted time.” Every action I take is a goal being selected. The difference between these goals is not structural. It is in the properties described below.

2. Properties of a Goal

Clarity

A clear goal has tangible steps, can be easily visualized, and has a definite start and end. Clarity falls on a spectrum.

High clarity: I’d like to go to the gym today. (Get clothes, go to gym, work out.)

Low clarity: I want to be athletically extraordinary at the national level. (Which event? How to train? How much time to dedicate?)

Motivation

Motivation is how much I want to pursue a goal. A very boring task has near-zero motivation. Motivation is a psychological state, subject to rapid changes — music, exercise, mood, a conversation.

Probability of Success

Probability of success is the estimated chance that pursuing a goal actually results in the desired outcome. Going to the gym today: 99%. Qualifying for Team Canada: 40%. Getting hired at a frontier AI lab: 5%.

This is a property of the world, not a psychological state, which is why it is separated from motivation. It estimates how likely a goal is reached given that you have done everything you can.

Alignment

Alignment captures how good achieving a goal actually is for me.

This requires an assumption: there exist many possible timelines, and my well-being across them is a function of wealth, happiness, health, and peace. A high alignment goal pushes me closer to outcomes near the top of that distribution. A low alignment goal pushes me toward the bottom. Negative alignment actively makes things worse.

Alignment is the hardest property to measure. It is also the only one that connects to actual long-term outcomes. This asymmetry is a problem for the model: the variable that does the most explanatory work is the one I can access least reliably.

One thing I keep returning to is that alignment is not always smooth. Some goals have alignment that increases linearly with effort. Others jump discontinuously at a threshold. Any goal with a binary qualification — a credential, a hire, a ship date — has a step function in its alignment curve.¹

Comparing Goals

Given these properties, there is an obvious temptation to reach for something precise: multiply motivation by alignment by probability and call it expected value. I don’t think that’s honest. These inputs aren’t on commensurable scales. I can’t put a number on alignment. Motivation shifts hourly.

What I can do is use these properties as a rough ordering heuristic. Given two goals, I can ask: which one has better alignment? Which one has a more realistic probability? Which one am I actually motivated to pursue? The answers won’t be exact, but they produce an ordering, and a crude ordering beats having no model at all.

The ordering implies some basic structures. Goals with high motivation and low alignment feel rewarding but produce little. Goals with high alignment and near-zero probability are technically worth pursuing only when the attempt cost is low. Goals with negative alignment and high motivation are probably the most dangerous class, because they concentrate my energy on outcomes that actively make things worse, and the motivation makes them hard to resist.

I suspect a large fraction of how most people spend their time — myself included — falls into the high motivation, low alignment category. This is hard to verify because alignment is exactly the thing that’s difficult to measure.²

3. The Priority Queue

A sufficiently rational actor would model their goals as a priority queue sorted by value: one goal selected at a time from the top, executed, then the queue re-evaluated.³

The useful thing about this framing is that it collapses failure modes into a single category: input quality. If the queue is correct, every bad decision must be traceable to a bad input. You don’t look for the failure at the selection step. You look upstream.

A blocked task, in this model, is not abandoned. It is removed from the queue temporarily and added back when new information changes its properties.

The queue is also not static. Selecting from it changes it. Completing a goal tends to increase motivation for similar tasks as confidence builds — a positive feedback loop that moves related goals higher. The inverse is also true: repeated failure or abandonment depresses motivation for similar tasks, which moves them lower, which makes further avoidance more likely. The ordering is self-reinforcing in both directions.

The compound structure of alignment complicates this further. As goals are actualized, future goals’ alignment shifts. Finishing a degree changes the value of jobs that required it. Making a national team changes the value of training volume that was necessary to get there. The queue is not just re-sorted by selection — the values themselves evolve as circumstances change.

4. Distortions

The priority queue represents an ordering by actual value. This is a theoretical object I don’t have direct access to. What I actually select from in practice is a perceived queue — one shaped by mood, systems, and a set of distortions that warp my perception of the scores, or cause me to ignore the ordering entirely.⁴

Several categories of distortion are worth distinguishing, not because the categories themselves are particularly clean, but because they point toward different kinds of interventions.

False inputs fabricate goals or data that don’t correspond to reality. The clearest example I have is OCD. Untreated, my OCD generated goals like “protect yourself from someone who is trying to harm you.” That goal scores extremely high on motivation and alignment, because self-preservation is about as aligned as it gets. The queue correctly prioritized it. The problem was that the threat wasn’t real. OCD was injecting fabricated data into the system. No amount of better deliberation at the selection step would have helped — the selection step was working correctly. What was broken was the input.

Sertraline is an intervention that targets this. It doesn’t fix the queue logic. It corrects the data feed. That’s a substantively different kind of fix than willpower or strategy, and it only became clear to me once I had language for distinguishing the selection mechanism from its inputs.

Noisy inputs introduce real but unpredictable perturbations to motivation — fluctuations that randomly reshuffle the queue in ways that can’t be anticipated or controlled. Music is a case I’ve thought about. At 5-6 hours a day, music is not a discrete goal being selected. It is a persistent system-wide modifier on motivation. A song that hypes me up for training increases motivation on a high alignment goal. A song that makes me nostalgic kills my drive to work. These shifts are real — they actually change the ordering, not just my perception of it — but they are unpredictable and untargeted. The volume at which I listen means it functions as an uncontrolled substance affecting psychological inputs that cascade through every decision. Unlike sertraline, it isn’t calibrated to correct a specific problem. It just adds variance.

Biased inputs systematically push selection in a consistent suboptimal direction. Unlike noise, the bias is predictable. Unlike false inputs, the underlying signal is real — it is just weighted wrong.

Ego is an example. Ego skews the weights toward staying on the current path. It doesn’t directly change motivation or alignment. It causes me to treat the cost of switching as higher than it is, because switching feels like admitting the previous choice was wrong. This is a systematic bias toward inertia, and it operates silently precisely because it works through the plausible-feeling mechanism of “I’ve already invested in this direction.”

Instant gratification is another. I systematically penalize tasks with the same alignment and probability of success if one requires more upfront effort to start. This is not irrational in the sense of being incorrect about the difficulty — the easier task genuinely is easier. But it means I’m adding a penalty to high alignment tasks that have high startup costs, which suppresses them relative to their actual value.

Feedback distortions are not a single bad input but a compounding process. Each time a low-value goal is selected, motivation for high alignment tasks drops slightly, which lowers their ranking, which makes them less likely to be selected again, which leads to another low-value selection. The queue progressively degrades through repeated bad selections.

Procrastination is the canonical example. It is not laziness — or at least, that description doesn’t seem mechanically accurate. What I observe is a negative feedback loop: avoiding a high-alignment task makes future avoidance more likely, because motivation for that task is genuinely eroded by each avoidance cycle. The queue is corrupted from within.

The loop typically breaks when a deadline approaches. Deadlines change alignment directly: as the deadline gets closer, the cost of not completing the task increases, so alignment rises. Urgency also raises motivation. At some point the ranking overwhelms every active distortion and the task gets selected. In retrospect this looks like success. What’s less visible is the entire window before it, where the feedback loop kept a high-alignment task suppressed while lower-value goals accumulated time they did not deserve.

5. Interventions

An intervention is anything that modifies the inputs to the priority queue. The OCD treatment was an intervention targeting false inputs. Reducing music exposure would be an intervention targeting noise. Both work not by improving deliberation at the selection step, but by changing what the selection step receives.

This framing has a specific implication: interventions aimed at willpower or motivation in isolation are working at the wrong level. They treat the selection mechanism as if it were broken when the mechanism is fine. What’s broken is upstream.

The more interesting and harder question is how to identify which distortion is operating in a given case. The symptoms are similar — goals that feel important not getting done, time going toward things that don’t compound. But the cause matters for what would actually help. Noise calls for reduction of variance sources. Bias calls for a counterweight. False inputs call for something more like therapy or medication. Conflating them produces interventions that address the wrong thing.

6. Open Questions

The model gives me language for reasoning about goals. It does not resolve several things I find genuinely unclear.

Alignment is doing most of the work, and I have no reliable method for measuring it. The entire framework depends on a variable I cannot quantify. I can make rough qualitative judgments, but the model’s precision is entirely illusory at this variable. This might not be a fixable problem — alignment may be intrinsically hard to measure because it requires predicting your own future preferences, which are themselves shaped by what you do.

The non-linearity of alignment (the step-function case) suggests that the structure of value across outcomes is itself worth examining. Threshold effects mean that the ordering between goals is not just about which has higher value on average, but about which has the right kind of value distribution. I explore this in the next piece.

Motivation as an input risks circularity. If I select by ranking and motivation is part of the ranking, the model may be restating “I do what I feel like doing” with additional formalism. I don’t think it fully collapses into this, because the other inputs — alignment, probability — can override motivation when they are strong enough. But the circularity is real and the model doesn’t fully escape it.

The queue being “never wrong” is unfalsifiable by construction. Any bad outcome is attributed to bad inputs. This is useful as a modelling assumption — it directs diagnostic attention toward inputs rather than the selection mechanism — but it should not be mistaken for a claim about reality. It is possible that the selection mechanism itself is flawed in ways the model can’t represent. I’m not sure what that would look like, which is part of why I’ve left it as an assumption rather than trying to test it.