A field guide for an uncertain world
How to Think
A practical field guide to models, systems, uncertainty, and changing your mind
Core premise: reality is richer than any model. Good thinking means choosing useful simplifications, knowing what they omit, testing them against reality, and switching models when they stop working.
Enter the atlasBefore you begin
There is no single best framework.
I use these thinkers as different ways into a problem. They often disagree, which is useful: each one helps me notice something the others miss.
Act 01
A label is not an explanation.
A label tells you what something resembles. An explanation tells you how it works and how it got there.
Gilles Deleuze
How did this become what it is?
Start with the thing you can see, then ask what made it. Calling a retailer “high quality” is not an explanation. Ask what concrete conditions—customer loyalty, buying power, store density, low costs, and capable operators—produced the result.
thing → category
difference → relation → process → emergence
Representation
Labels such as “emerging market,” “monopoly,” “luxury,” or “high ROCE” save time. But naming something can create a false sense that we understand it.
Same result, different causes
When two companies look similar, ask how each one got there.
One number hides many realities
A company or country is made up of people, incentives, contracts, technology, institutions, geography, and history. “Country X grew 7%” can be true even while agriculture is weak, youth unemployment is high, and some regions are falling behind.
Thresholds matter
A factory at 60% utilisation and the same factory at 90% can have very different economics. Look for bottlenecks and tipping points, not just gradual change.
Problems before solutions
“How do we increase app engagement?” invites notifications and gamification. “Why is there no recurring reason to open the app?” may reveal that the product solves an occasional problem and should not be optimised for daily use.
What produced this result—and what does the label hide?
Alfred Korzybski
The map is not the territory
Every useful map leaves things out. A road map ignores trees and building interiors; a DCF leaves out much of the day-to-day business. Simplifying is necessary. Forgetting what you removed is the danger.
- Territory
- What is actually happening.
- Map
- How you represented it.
- Omission
- What the representation cannot contain.
What did this representation have to omit in order to become useful?
Act 02
Look past the event. Find the system.
When a problem recurs, the structure is often a better explanation than the latest incident.
Donella Meadows
Structure creates behaviour
Look for what accumulates, what flows in and out, what feeds back, and where delays sit. A system can turn sensible individual choices into a bad collective result.
Reinforcing loops
A change produces more change in the same direction: more users → more content → a more useful platform → more users.
Balancing loops
A change activates forces that push back: high prices → lower demand → inventory builds → discounting → prices fall. Extrapolating the first leg misses the system’s response.
Delays
Cut marketing today; revenue stays strong for three months; management concludes marketing was wasteful; six months later the pipeline collapses. Delayed feedback makes the original decision look smarter than it was.
Stocks and flows
Customer base is a stock. Acquisition is an inflow; churn is an outflow. Celebrating record acquisition while churn quietly rises is like praising a stronger faucet while the bathtub drain widens.
Leverage points
Small parameter changes often do little. Information, incentives, rules, and goals usually matter more. If service agents rush customers because promotions depend on calls per hour, another training manual will not fix it. Change the metric.
What system would make this behaviour predictable rather than surprising?
Gregory Bateson
Look at the relationship, not just the thing
We tend to blame isolated things: the manager, the customer, the price. Bateson’s useful point is that behaviour often comes from the relationship between them.
A difference that makes a difference
A ₹100 cup of coffee is not “high” or “low” by itself. ₹100 relative to its ₹50 cost, a nearby café’s ₹120 price, a customer’s ₹140 willingness-to-pay, or yesterday’s ₹90 price are four different pieces of information.
Patterns over traits
A “bad manager” may improve dramatically under a different boss, incentive system, and team. The old environment helped produce the old behaviour.
Double binds
Leadership says “take initiative,” but punishes decisions not pre-approved. Employees learn that the safest form of initiative is to ask permission for everything.
What pattern of relationships is producing the behaviour I am attributing to an isolated object?
Act 03
Your starting assumption shapes what you notice.
Two people can study the same facts and focus on completely different things.
Thomas Kuhn
Frames and inconvenient facts
A frame tells you which questions to ask and which evidence to take seriously. That is why two analysts can read the same filing and reach very different conclusions.
Pay attention to what does not fit
You believe commodity producers cannot sustain high returns, yet one firm does so across several cycles. It may be temporary—or your original view may be missing a real advantage.
The useful habit is simple: when a fact does not fit your view, do not explain it away too quickly.
What is my current view making hard to notice?
Act 04
You cannot know everything. Decide what is enough.
The world may be complicated. Your decision rule does not always need to be.
Herbert Simon
Bounded rationality
Perfect optimisation assumes that you can see every option and consequence. In practice, time, information, and attention are limited. A good process must include a point at which you stop searching and decide.
Satisficing
Set a clear bar and choose an option that clears it. In hiring, define the non-negotiables—competence, reliability, learning ability, and compensation fit—then hire when a candidate meets them well enough.
Search cost
If another week of research has only a 5% chance of changing a small purchase decision, continued analysis may be less rational than acting and learning.
Adaptive strategy
Under uncertainty, a choice that is good enough, reversible, and informative can beat a polished five-year plan. Pilot five stores, learn, and update before planning a national rollout.
How much uncertainty is actually reducible before I need to act?
Gerd Gigerenzer
When a simple rule is enough
A decision rule does not need to copy all the complexity of the world. In the right setting, a few reliable signals can be clearer and more robust than a large model.
Less can be more
A 40-variable credit model may look sophisticated. If repayment history, income stability, and debt burden capture most default risk, the simpler rule may generalise better and fail more transparently.
The task matters
Understanding, prediction, and decision are different jobs. You may need a detailed model to understand churn. To decide which support tickets need immediate attention, three clear rules may work better.
First understand enough of the situation to avoid a crude answer. Then ask whether more detail will actually change the choice. If it will not, stop.
What is the simplest rule that captures enough of what matters for this decision?
Act 05
Use probabilities, not declarations.
A belief should change when the evidence changes.
Bayes
Put a number on the belief
Start with an estimate, look at the new evidence, and revise the estimate. You do not need the equation to use the habit.
A company has a 30% chance of losing its largest customer within two years. The customer begins testing a rival: 30% → 45%. A contract extension arrives: 45% → 35%. Updating is not indecision; it is the point.
Diagnostic evidence
“The company raised price 15%” is weak evidence of pricing power if the entire industry raised price. “It raised price 15%, competitors did not, and volume still grew” is much harder to explain without genuine pricing power.
How much more likely is this evidence under my hypothesis than under the alternatives?
Base rates
Check what usually happens
The story explains why this case feels special. The base rate tells you what usually happens in similar cases. A brilliant restaurant concept or charismatic turnaround CEO can improve the odds, but should not erase them.
What happens in the relevant reference class before I tell myself why this case is different?
Kahneman & Tversky
Four common probability mistakes
- RepresentativenessResemblance is mistaken for probability.
- AvailabilityVivid examples feel statistically common.
- AnchoringA starting number contaminates later estimates.
- Base-rate neglectThe specific story overwhelms the statistical pattern.
Is this shortcut a bug here—or a useful heuristic?
Philip Tetlock
Make forecasts you can check
Replace “confident” with a number. Define the event, record the forecast, update it, and later see whether you were right.
“AI will transform banking” cannot be scored. A dated forecast about the five largest private banks, a measurable service-interaction threshold, and a definition of human escalation can.
Before the evidence arrives, write down what would change your mind.
What probability am I assigning, and what evidence would move it by 10–15 points?
Worked example · investing
A “cheap” company, seen six ways.
Company X trades at 12× earnings while peers trade at 25×. That makes it look cheap. It does not tell us whether it is a good investment.
It trades at 12×.
That is a useful fact, but not yet an investment case.
Look beyond the multiple.
The P/E says nothing about debt, cyclicality, reinvestment, accounting quality, or whether today’s earnings will last.
Find out why it is cheap.
Lower quality can lead to churn, lower volume, poor utilisation, weaker margins, and less money to reinvest. Cheapness may be a symptom, not an opportunity.
What story fits the facts?
It could be undervalued, in structural decline, or simply near the bottom of a cycle. Each explanation requires different evidence.
Know what would make you pass.
For example: avoid highly leveraged cyclical companies with weak interest coverage, however cheap they look.
Make the thesis testable.
Start with the success rate of similar turnarounds, estimate the odds, and write down what evidence would raise or lower them.
Worked example · product management
Daily active users have stopped growing.
Before launching notifications or campaigns, work out what has actually changed.
DAU has stopped growing.
That tells us what happened, not why.
What changed inside DAU?
Check new-user activation, retained users, usage frequency, churn, and seasonality before choosing a remedy.
Find the weak link.
A poor first session can lower retention. That means less content and fewer referrals, which makes the product weaker for the next group of users.
What kind of problem is this?
It may be a marketing problem, an activation problem, or simply a product people do not need every day. Each calls for a different test.
Focus on two numbers.
Use first-session completion and seven-day retention. Leave the rest aside unless it can change the roadmap.
Try one reversible change.
Pilot a clearer onboarding path for one cohort. Decide in advance how much retention must improve before you roll it out.
Act 06
Put the ideas to work.
No single approach is enough. The value comes from using them together on the same decision.
Understand first Do not simplify the situation too early.
↔Decide eventually Do not demand a complete theory before acting.
Change the frame Sometimes the model itself is wrong.
↔Update the odds Sometimes the model is fine and the probability changed.
Use the map A useful model must leave things out.
↔Check the gaps Know what was removed and whether it matters.
Look one layer deeper There is usually more to understand.
↔Know when to stop If more detail will not change the choice, act.
A practical workflow
Twelve steps for a real decision
- 01Define the problem
Write the question clearly. Expose assumptions built into its wording.
- 02Separate map from territory
List the categories, metrics, and models in use; note what each omits.
- 03Trace the system
Map what builds up, what flows in and out, where feedback appears, and where delays sit.
- 04Map relationships
Find variables whose meaning depends on each other rather than standing alone.
- 05State your starting view
Write down your frame and try at least one competing explanation.
- 06Find the cause
Trace the process that produced the visible result.
- 07Set a search limit
Identify which unknowns can actually change the decision.
- 08Test a simple rule
Ask whether a few robust cues can make the decision reliably.
- 09Check the base rate
Find out what usually happens in similar cases before getting lost in this story.
- 10Assign a probability
Use a number rather than “likely,” “confident,” or “bullish.”
- 11Search for anomalies
Find the strongest fact your model struggles to explain.
- 12Update, decide, review
Move with the evidence; act when analysis has low value; later score the forecast.
Transfer test
“Country X will become a manufacturing superpower.”
Turn the slogan into questions you can answer.
- Define success. Export share? Manufacturing value added? Employment? Technological sophistication?
- State the story. China replacement, demographic dividend, friend-shoring, or domestic-market scale?
- Trace what is required. Power, logistics, labour, capital, education, regulation, suppliers, currency, and scale.
- Compare costs properly. “Cheap wages” mean little without productivity, energy, logistics, quality, and training costs.
- Look within the country. Regions and institutions may differ enough to produce very different outcomes.
- Say what would prove you wrong. Identify the few variables that could kill the thesis and write a dated forecast.
Keep these
The 12 questions worth memorising
- What generated this?
- Do I understand it, or have I only named it?
- What did my model leave out?
- What system would naturally produce this behaviour?
- Where are the feedback loops and delays?
- What relationships create the pattern?
- What does my framework make hard to see?
- What can I realistically know before I have to act?
- What is the simplest rule that captures enough?
- What is the base rate?
- What probability am I actually assigning?
- What evidence would materially change that number?
The practical takeaway
You do not need to memorise the thinkers. Use these five checks.
- Name the claim. What are you calling this, and what does that label leave out?
- Explain the mechanism. What process, relationships, incentives, and feedback loops produced it?
- Challenge the frame. What competing explanation would direct your attention somewhere else?
- Set the bar for action. Which unknowns can actually change the decision, and when will you stop researching?
- Put a number on it. What probability are you assigning, and what evidence would change it?
The integrated mental stack
A good thinker can say:
This is the model I am using. This is what it captures. This is what it leaves out. This is how confident I am. This is what would make me change my mind.