Marginal Gains

Marginal gains is a method for choosing what to improve next: the smallest change you are near-certain to pull off, chosen so it builds on the one before it. A small improvement in the wrong place is not a small gain; it compounds with nothing.

The question is which part is capping the rest. What could I improve is the wrong question.

Core takeaways

• A gain counts when it builds on the last one and sits on the part of the process that caps everything else. A gain that opens an unrelated thread compounds with nothing.
• The capping part, the rate limiter, is found by a check every one to two weeks. Procrastination, time management, prioritisation, focus, experimentation and critical reflection are the first places to look, and the limiter changes over time.
• Two or three processes are trained at once. Each skill is a problem that has to be decoded, and decoding costs working memory. Spreading that capacity across five skills produces competence in none.
• Fewer at once also keeps a later layer of a different kind in the system, and that layer is what catches what an earlier pass missed. Collecting isolated details out of fear of missing them produces weakly encoded notes.
• The increment is picked with five steps (goal, dissection, target rating, current rating, increment) and tracked in one notebook with a written win criterion, because a 1% change cannot be felt.
• At a one-to-two week cadence a compounding 1% is about 1.7× over a year. The position of a gain matters more than its size. Large strides fail more, teach later, and can succeed at the wrong thing.

What makes a gain a gain

A gain stacks when it adds value on top of existing progress rather than scattering effort. The test is one question: does this next gain add value to the progress I just made, or open an unrelated thread?

Three requirements have to hold together. The improvement targets an important skill that actually moves the goal; chasing a peripheral habit while the bottleneck sits still wastes the compounding, which is why Upgrading Your Dimensions separates short-term foundations from later cognitive growth. It has to persist long enough to find the mistakes, fix them, and let the new way become a habit. And it has to fit the larger operating system, not just score well on a local metric. Miss any one and progress caps. When all three line up, stacking buys two things: growth, and habits. Without the habit, every repetition keeps costing effort and concentration.

One worked case: a first gain in note-taking, compressing difficult concepts into fewer keywords, improved retention and left the processing slow. Of the two candidates for next, working out the big-picture ideas first stacked, because it strengthens the same deep processing the first gain moved. Cutting procrastination to get started sooner helped and built on nothing; it left the processing untouched, a barrier beside the skill rather than a gain on top of it.

Poor stacking looks like a switching spiral: non-linear notes, then flashcards, then a method for reducing procrastination, then time management, then back to note-taking, unimproved. Weeks or months pass. Divided effort slows acquisition because each skill needs enough experience and reflection to make mistakes, understand them, and build each increment to a meaningful level. Mistakes surface more slowly, experiments lose focus, and each skill takes longer. Weeks pass with surface familiarity across many areas and competence in none, the plateau The Technique Is Only as Good as the Thinking It Produces names, where better methods leave results unchanged. A clean acquisition creates headroom. Switching creates work without creating competence.

Effort without progress has two causes. Fluctuating gains come from inconsistency, real increments that do not build on each other, and are repaired by stacking. Negative gains come from unreliable information, calibration, or feedback, and are repaired by getting feedback, not by persisting harder. From inside they feel identical. Persisting harder on a negative gain makes it worse. Increments build on each other just as reliably in the wrong direction when the feedback is unreliable; a stack is only as good as what is telling you it worked.

Whether a gain will compound at all is a removal question, not a size question: if the input vanished, would you still be more capable? That test lives in Compounding vs Additive Gains.

Where the gain has to sit

A rate limiter is the part of a process that prevents every other part from improving. Fixing anything else while it remains does not raise the ceiling. The comparison is a bucket with a hole in its side: perfect everywhere else, it still will not fill above the hole. Work on the handle or the rim changes nothing. The water you pour in meanwhile is wasted.

The same principle one level up is Dimensions of Learning: five capabilities, and the weakest sets the ceiling for the rest.

Limiters are not always obvious. A wrong guess still beats not looking: attempting the wrong limiter still tends to produce smoother progress than charging ahead without looking for one. The limiter is the part most consistently constraining, and that is a different part from the one most recently frustrating; keeping that distinction is what stops the hunt from turning into a complaint.

The first place to look is the enablers. Self-management (procrastination, time management, prioritisation, focus) and the growth skills, experimentation and critical reflection, limit execution across the board. A strong learner can be capped by heavy procrastination alone. Old cue-response habits that fire faster than a new skill can stabilise are a candidate limiter of their own, treated in How to Unlearn Old or Bad Habits Efficiently.

Limiters rotate. Limited by a Fixed vs Growth Mindset early, a learner may be limited by time management a month later, because time management got worse, or because the mindset improved. When a constraint breaks, return to the hunt and do not let inertia keep you optimising the old one.

The check runs every one to two weeks: is there a part of my process that seems to be holding everything else back? If yes, that is the next target, ahead of other improvement work. You may already be working on it, in which case continue. A plateau triggers the same check without waiting for the calendar date. Which factor is actually holding performance back, and so which gain is worth stacking next, is the job of Measuring Learning.

A technique can be improved forever, so “good enough” cannot come from the skill. It comes from position: you work a skill while it caps the rest, and you move when it stops being the cap.

How many at once

Two fronts is the working number. Three is the ceiling. More than three overwhelms. The cap is two or three processes in the active set being trained, not the whole skill list. Each process in the set is a problem sitting in the way of the next hour. Running that few feels too slow, and it is how a skill actually lands.

The integer is this system’s working default, the same ceiling used when stacking small improvements on the current few. No trial has shown that two or three beats four or five as a measured optimum. The instruction is the cap, not the constant.

A student who needs critical thinking, essay writing, encoding, retrieval, and time management in two months, from a basic baseline, cannot grow all five in time. The move is still to grow each as fast as possible, by running fewer at once. The others wait in the pool. The price is that wait, and it is paid at the same time as the speed is gained.

Measuring Learning says which two or three: the rate limiters. The cap says how many. Inside the structure The 30 Day Plan sets, the gains run on the two or three skills that matter most for the goals of that plan.

A named change inside a fortnight is the checkable sign that a process is actually active. If none of the two or three is moving, the pick is the problem, not the cap.

Why fewer is faster

The picture in the way is skill-building as collecting tools. A tool can be picked up and used. A skill cannot. Each one is a problem sitting in the way of the next hour: something blocks performance, the block has a cause, a solution has to be tried, and the loop repeats. That decoding costs load. Exposure is not the cost. The expectation is a kit that grows by addition: another method, another technique, another item on the list. The work is solving one problem well enough that the next one can use what was learned.

Split the same limited capacity across many skills and none of them get enough of that work. Each skill has variables that have to be decoded before it will run. High element-interactivity already strains working memory inside one skill. Across five at once, the same hours produce confusion and frustration instead of a landing.

What the cap is protecting

Those two or three processes are not only a count. They are how a later layer stays available. What fails is a missing kind of method, not a skipped line in the notes. A system with no method that forces comparison, judgment, and relationships misses interrelational questions on every pass. The skipped fact is not the risk. The missing layer is.

A learning system works like stacked layers, each with holes. Encoding, then retrieval, then encoding again: each pass misses something different, and a later pass can catch what an earlier one missed, if a later layer of that kind exists. The image is a Swiss-cheese model: several imperfect layers covering different misses. It is a metaphor. It is not a proof that misses get caught. The model those layers are borrowed from is a model of residual failure when holes align. When every layer is the same kind, the miss goes through.

A detail skipped now often gets picked up later, by a technique better suited to it. That consolation holds only when the later layer exists. A fact that is too detailed for the current network is overload if forced now, and a cue once the surrounding network exists. Isolated elements first, then interacting ones: the same sentence is load for a novice and easy later. Not all information is equal.

Order Control is taking information in the order the current network can hold, not the order the chapter printed it. Prestudy is the layer that decides what is too detailed to take now.

Focusing on details because of a fear of missing them makes them more likely to be forgotten. Isolated notes encode weakly. Organisation is what retention uses, and isolation has none. Tidy notes then feel like fluency: the page looks complete, the network was never built. The harvest is the slowdown. Skipped details do come back later, when a later layer of the right kind exists. The warning is the other half: the frightened harvest is itself the reason they do not come back.

Two or three at a time still looks slow. It is how a later layer stays in the system, the layer that can actually catch. The frightened harvest of isolated details is slow in the same way, and it looks like thoroughness. Fix the missing method, not the individual detail. A deadline that truly needs five skills from a basic baseline will not grow them all. The cap does not invent time.

One selection rule, then you run it

Does it build on the last one, and is it on the thing capping the rest? Two questions, one selection rule. Selection is half the method. The other half is running it.

Direction is the job of the selection rule. The improving lives in Kolbs Experiential Cycle, the four-stage loop that turns one real attempt into a better next one. Selection without the cycle produces wishful planning. The cycle without selection produces unfocused reflection, which is rumination. You cannot run a reflection cycle on every problem every day, so something has to choose which problems get one.

The 30-Day Plan describes what to do for the next thirty days: a month-long plan that removes barriers and sets up a path. Marginal gains describes how to improve inside that plan. Combined, the order is: name the goal, then the performance goals, then the habits and environment in the way; then two or three skills tracked and stacked; then pace adjustments.

Aspect	Marginal Gains	30-Day Plan
Primary focus	Skill development (getting better at a skill)	Self-management (taking action productively)
Core insight	Skills improve faster and more sustainably through compounding tiny gains than through large jumps	Goals are more likely to be achieved when plans account for barriers in advance, especially around habits and tendencies
Method	Tracking and stacking 1% gains through experimentation and reflection	Clear goal-setting and habit review to create a high-probability plan
Usage	Continuously, wherever skill development is desired	When a concrete, structured plan is needed to make rapid progress toward a specific goal
Scope	Breaks bigger skills into smaller focused gains	Breaks bigger goals into 30-day mini-goals with a clear plan
End goal	High competence in the target skill	Achievement of the medium-term goal

The loop, run forward

REDO is the sequence that turns a marginal gain into a stacked improvement. It runs after each serious attempt, looking forward, and it ends in a plan.

R, Reflect. What have the previous attempts taught, looking for trends across multiple attempts, and what might the next increment be? One attempt cannot show a trend.

E, Evaluate. Which direction. The fork is: unblock yourself by moving to another skill, or unlock this one with more experimentation, theory, and reflection.

D, Define. Name the increment, specific enough to test, small enough to stay with, and estimate how long results will take. Early improvement usually looks like finding mistakes rather than performing better. One worked case allowed at least ten hours across a week to surface the major mistakes, then another week or two to test fixes.

O, Optimise. Make a specific plan, run it, reflect again from step one. The goal is data, not perfection.

You will not see improvement on every attempt. What you should see is a trend toward clarity and awareness, with performance later. That is the checkable expectation. With no attempt there is nothing to reflect on; that is a precondition for running the loop, and it is not what the loop is. Not knowing how to start is a finding: the next move is a best attempt, because working out how to start can run forever.

Run the Kolbs Experiential Cycle after each serious attempt. The form you actually open is Kolbs Template.

Picking the 1%, and being able to see it

The five steps, used once to set the field and then kept light:

1. Anchor a meaningful goal, nine months to three years out. Outcomes are controlled indirectly through the skills, attributes, habits, resources and actions that raise their probability, the mechanism in Reverse Goal Setting.
2. Dissect it: what knowledge, attributes, skills, processes, resources does it need, then ask how sure you are. When the honest answer is not sure, that uncertainty is the finding, and the next gain is to get the information. You do not need the full list, only the parts clearly necessary as the next step. At the very beginning, the step is sometimes just taking a step. The structured walk that produces this inventory is Skills Audit, run after rushing, after a break, or at a plateau.
3. Rate the target level out of ten and say in words what that number means. Seven out of ten on time management means time reliably goes to what moves the goal, with enough room for error that it need not be perfect.
4. Rate the current level out of ten and justify it as objectively as you can.
5. Name the increment that closes the gap.

The justification names the gain. No knowledge → go learn. Knowledge but no practice → run one real attempt. Knows how but has never done it → do it once. Later reviews repeat only steps 4 and 5, because the first three barely change. Early on the next step is unclear because everything looks like it needs work; as each part gets worked, which step comes next gets easier to see. The stacking test and the five-step ladder are the same question at two ranges: what does this build on.

When you cannot see a gain, information is the gain. You cannot take action before knowing what action to take. Know enough to attempt it and start making mistakes → experiment. Not enough even to make a mistake → go learn first. Gains progress knowledge → actions → behaviours and habits → position, meaning how much control you have over reaching the goal.

You cannot feel a 1% change. The brain registers small gradual movement badly, so early on there is no honest intuition about whether you are gaining, which is why invisible progress is one of the most common causes of early demotivation. Progress that is not written down is invisible. Stacking keeps the direction positive; tracking keeps it visible. The artefact is one notebook: the targets you are on, a written win criterion that tells a gain from a loss, reviewed weekly. That is not a second dashboard. The gain should make the next attempt easier, cleaner, or more motivating; once the record has become a dashboard, keeping it has stopped being a gain.

The same logic runs outside studying. In fast-moving work, small improvements to search, review, source hygiene, feedback speed and sense-making loops compound into a stronger operating system, the claim in Nothing Ever Happens Is Over. A skill gain pays out everywhere the skill is used.

Keep the increment small enough to start and meaningful enough to repeat. Large strides trade safety for an unlikely boost.

For Priority 0+1, one to three priorities for tomorrow, two the usual working number, a gain should usually improve one of five things: recurrence, emotional engagement, identity reinforcement, visible progress, action initiation. The operators named in the table below have their own pages: BHS, SIR, Shortcut detection, and the approach-layer work in Interleaving for Complex Problem Solving.

Priority 0 Area	Skill Or Process	Possible Marginal Gain
Agentic Engineering	Prompting agents	Save one reusable prompt pattern after it works.
Agentic Engineering	Agent workflow design	Add one clearer instruction to AGENTS.md or a project README.
Agentic Engineering	Code review with agents	Ask for one focused review category instead of a broad review.
Agentic Engineering	Verification	Add one repeatable build, test, or screenshot check to the workflow.
Learning Systems	BHS	Improve the Aim step by writing sharper why/how questions before reading.
Learning Systems	SIR	Add one interleaved prompt that forces comparison instead of recognition.
Learning Systems	Complex problem solving	Change one meaningful variable before executing again.
Learning Systems	Kolbs	Reflect on one real attempt instead of journaling about the whole system.
Learning Systems	Shortcut detection	Name one shortcut and add one constraint that makes it harder to repeat.
Vietnamese	Immersion recurrence	Pick one default YouTube channel or playlist for low-friction starts.
Vietnamese	Noticing	Capture one repeated phrase or grammar pattern from real input.
Vietnamese	Listening	Rewatch one short clip until it feels less noisy.
Vietnamese	Comprehension	Adjust subtitles, speed, topic, or lookup tools to keep attention engaged.
中文	Maintenance	Read or convert one useful sentence without reactivating the whole language track.
中文	Character contact	Review one character, phrase, or short clip.
Fitness	Strength	Add one small progression: weight, rep, set, tempo, or form cue.
Fitness	Mobility	Choose one mobility bottleneck and repeat a short routine.
Fitness	Cardio	Make the start easier: shoes ready, route chosen, timer preset.
Fitness	Recovery	Track one signal: sleep, soreness, energy, or readiness.
Relationships	Contact	Send one message that would otherwise stay vague.
Relationships	Repair	Clarify one small tension before it becomes stale.
Relationships	Appreciation	Express one specific appreciation.
Relationships	Planning	Put one next touchpoint on the calendar.

Five questions, and a discard rule. Does it reduce start friction? Does it make the skill feel more alive? Does it reinforce the identity? Does it leave visible progress? Does it improve the next attempt? A gain that scores on none of them is organisation work, not improvement work.

Position over size

Once the two selection questions have been answered, the position of a gain matters more than its size. At the one-to-two week review cadence used here, a compounding 1% is about 1.7× over a year. The famous 37× needs a daily cadence and every gain compounding, a cadence the method does not recommend.

Large strides fail more, teach later, and can succeed at the wrong thing. They cost three things: high failure risk, more variables and a smaller margin; low iteration frequency, feedback arriving in weeks or months instead of days; and misalignment risk. Rebuilding a whole workflow around a tool, then finding that the methods that matter are impossible in it, is the misalignment case: a large stride that can succeed and still leave nothing behind. Daily and weekly gains iterate far faster, the difference can be many-fold, and they compound both skill and the accuracy of judgement about progress. The ladder costs a sitting; stacking costs the appeal of variety.

It should feel like narrowing the improvement field: you leave with one small, believable upgrade that connects to a larger goal. Good signs: specific enough to test, on a real bottleneck, repeatable across sessions, makes progress concrete. It has become avoidance when you keep planning tiny improvements and never run the next attempt. Not knowing how to start is the same finding in other words: make a best attempt.

The gain also improves the picking. Learning your own tendencies both prevents the next mistake and makes the next reflection more accurate.

How to practice this

1. Every one to two weeks, ask which part of your process holds everything else back. Work on that part ahead of other improvement work. Notice whether the same limiter comes up again or a new one has taken its place.
2. Before taking a gain, ask whether it builds on the last one. Notice a gain that opens an unrelated thread; that is a switching spiral starting.
3. Keep two or three processes active and leave the rest in the pool. Notice a named change in each inside a fortnight; if none moves, change the pick.
4. Rate the target level and the current level out of ten, and write what each number means. Notice the justification; it names the next increment.
5. Keep one notebook with the targets you are on and a written win criterion, reviewed weekly. Notice that a 1% change cannot be felt, only read from the record.
6. After each serious attempt, run Reflect, Evaluate, Define, Optimise, and end in a plan. Notice mistakes becoming visible before performance moves; that is the early trend.

Related pages

• Mindset: parent dimension; the material the fixed-mindset limiter example depends on.
• Self-Regulation: second parent; the page is dual-parented.
• Measuring Learning: rate limiters tell which two or three to pick, and which limiter is live; how many is answered here.
• Order Control: take information in the order the current network can hold, not the chapter’s.
• Prestudy: the layer that decides what is too detailed to take now.

Open questions

• Early cycles that buy clarity rather than performance, and fluctuating gains that never compound, feel identical from inside, effort without visible progress, and they take opposite repairs. How early can the two be told apart?
• When a later layer of the missing kind cannot be added in time, what is the honest move: drop the fact, or drop the deadline’s implied scope?

Sources

• Theory of Constraints: the public name for the rate limiter, and the five focusing steps whose fifth supplies the inertia warning after a constraint breaks. https://en.wikipedia.org/wiki/Theory_of_constraints
• Compounding arithmetic, stated as computation rather than as a citation: 1.01^52 ≈ 1.68; 1.01^365 ≈ 37.8; 0.99^365 ≈ 0.026.
• Ericsson & Pool, deliberate practice: high-frequency feedback supports the small-and-frequent grain; this is not a trial of the named method.
• Koriat 1997, cue-utilisation; Bjork & Bjork, desirable difficulties: more frequent feedback sharpens judgement about one’s own progress.
• Locke & Latham: distal goals with proximal subgoals bound the nine-month-to-three-year window, which is this system’s planning grain rather than a finding.
• No controlled trial of marginal gains as a named learning method exists, and the size of the iteration advantage is not measured.
• van Merriënboer, J. J. G., & Sweller, J. (2005). Cognitive load theory and complex learning. Educational Psychology Review. Few interacting elements at once; the load half of the cap.
• van Merriënboer, J. J. G., Kester, L., & Paas, F. (2006). Teaching complex rather than simple tasks. High element-interactivity overwhelms working memory when too much is live at once.
• Pollock, E., Chandler, P., & Sweller, J. (2002). Assimilating complex information. Learning and Instruction. Isolated elements first, then interacting elements: why a fact too detailed now becomes easy later.
• Kalyuga, S., Ayres, P., Chandler, P., & Sweller, J. (2003). The expertise reversal effect. The same detail is load for a novice and a cue once the schema exists.
• Craik, F. I. M., & Lockhart, R. S. (1972). Levels of processing. Isolated, shallow encoding is what the frightened harvest produces.
• Bjork, R. A., Dunlosky, J., & Kornell, N. (2013). Self-regulated learning: Beliefs, techniques, and illusions of competence. Annual Review of Psychology. Tidy notes as fluency.
• Reason, J. (1990). Human Error; Reason, J. (2000). Human error: models and management. BMJ. Swiss cheese as residual failure when holes align: the model being scoped, not a proof that later layers catch.