IIO409 | Dr. Mauricio Zambrano-Bigiarini

How to Use ChatGPT to Learn Hydrological Modelling (I): set-up, prompts and trustworthy answers

Sun, 16 Aug 2026 00:00:00 +0000

Estimated reading time: 16-21 minutes

Before we begin: learning in an age of rapidly changing AI

Preparing university students for professional practice has become more difficult for one additional reason: we do not know exactly what AI tools will be capable of by the time today’s students graduate. Generative AI is evolving quickly and is already changing how we search for information, write, analyse data, program and solve problems. The educational challenge, therefore, is not merely to adopt a new technology, but to decide what to use it for and which student capabilities we actually want to strengthen.

Recent discussion highlights both opportunities and risks. Generative AI can widen access to explanations, personalise support and accelerate some tasks, but it can also encourage shallow learning, dependence and a false sense of mastery when it is used to produce polished answers without the necessary reasoning. In higher education, this has led to growing emphasis on critical thinking, adaptability, curiosity and lifelong learning, as well as on pedagogical frameworks that focus on purposeful integration rather than technology alone.

That is the purpose of this three-part tutorial series. It is not designed to teach you how to obtain quick answers. It is designed to help you use generative AI to build knowledge, question explanations, detect errors, practise reasoning, work more efficiently and receive feedback without outsourcing understanding. In IIO409, ChatGPT can help you write a function in R, propose a calibration strategy or explain GLUE in seconds. But you still need to be the person who can interpret a new result, modify a procedure and defend every modelling decision.

The sequence is deliberate. Part I focuses on setting up ChatGPT and obtaining useful, verifiable answers. Part II uses it as a Socratic tutor for conceptual understanding and critical thinking. Part III uses it as a base R programming companion for practical work and vibe coding without sacrificing reproducibility, validation or understanding.

ChatGPT can be genuinely useful for learning hydrological modelling, but its usefulness depends largely on how you use it. If you type only “Explain hydrological model calibration”, you may receive a correct general answer, but one that is still far too broad to be genuinely helpful in IIO409. If, instead, you define who you are, what you are learning, what your problem is, what prior knowledge you already have and what kind of answer you need, the conversation becomes much closer to a useful tutorial.

In IIO409 this matters especially. The course is not simply about running a rainfall-runoff model. You are expected to learn how to formulate a conceptual model, choose data and models, analyse sensitivity, calibrate and verify, interpret uncertainty, regionalise and discuss climate-change impacts. In other words, you must learn to make and defend modelling decisions.

1. Before you ask: set up the context

One of the simplest and most effective ways to improve ChatGPT’s replies is to give it stable information about yourself and about the course.

You do not need to write a biography. You need to provide the context that materially changes the answer.

A good initial context for an IIO409 student might be:

I am a Civil Engineering student in Chile taking IIO409 Hydrological Modelling.

I have already completed an R course and I can work with data.frame objects,
matrices, functions, plots and packages.
In this course I need to learn rainfall-runoff modelling, hydrometeorological data,
sensitivity analysis, calibration, verification, uncertainty, regionalisation
and climate change.

I work in base R. I do not want to use tidyverse.

When you help me:
- prioritise conceptual understanding and hydrological reasoning;
- do not invent references, R functions or results;
- separate facts, assumptions and inferences;
- if important information is missing, ask me before answering;
- for R code, deliver simple, reproducible and verifiable solutions;
- explain why each important step works;
- tell me how to check whether the result is correct.

You can store part of this information in Custom Instructions. Custom Instructions are available on both free and paid ChatGPT accounts, although the exact features and limits may vary by plan.

A particularly useful instruction is:

Before answering a technical task, check whether any missing information could materially change the answer.
If so, ask me up to three focused questions before continuing.

This simple rule reduces one of the most common failure modes: ChatGPT silently filling in missing information that you never provided.

Why does this help?

A language model does not automatically know your learning objective, your software ecosystem or the constraints of the task. When context is missing, it generates a plausible continuation. In hydrological modelling, a plausible assumption can still be methodologically wrong.

For example, if you ask:

How should I calibrate my model?

at least these questions are missing:

Which model?
Which variable are you calibrating against?
What is the purpose of the model?
What warm-up period are you using?
What are the calibration and verification periods?
What objective function do you want to optimise?
Which parameters are calibratable?
What ranges are physically acceptable?

A confident reply to such an underspecified question should make you cautious rather than reassured.

2. The R-C-T-F structure: Role + Context + Task + Format

There is no single magic formula for writing prompts. For this course, however, it is useful to think in four parts.

R: Role

State the perspective from which you want the answer.

Examples:

Act as a university tutor with expertise in hydrological modelling.

Act as a very critical methodological reviewer.

Act as a Socratic tutor. Do not give me the answer immediately.

The role does not magically turn ChatGPT into a true expert. It helps to guide the style and depth of the answer.

C: Context

Provide the information needed to understand the problem.

For example:

I am working on a pluvial catchment in south-central Chile.
I have daily precipitation, daily temperature and observed daily discharge.
I want to assess a conceptual rainfall-runoff model.

If you are working on the course project, add:

catchment;
period;
temporal resolution;
available data;
model;
aim of the analysis;
decisions already taken;
constraints.

T: Task

State exactly what ChatGPT must do.

A weak task:

Tell me about KGE.

A better one:

Explain which error components KGE attempts to summarise and why two simulations
with the same KGE can still behave very differently hydrologically.

Even better:

After explaining it, give me two hypothetical simulations with the same KGE but
different hydrological errors and ask me which one would be preferable for low flows.
Do not answer that last question until I reply.

F: Format

Tell ChatGPT how you want the answer delivered.

Answer in this order:
1. conceptual explanation;
2. hydrological example;
3. common mistake;
4. one question to check whether I understood.
Maximum 500 words.

For programming:

Provide:
1. strategy;
2. a base R function;
3. a small reproducible example;
4. checks of the result;
5. possible errors.
Do not use tidyverse.

3. A long prompt is not necessarily a good prompt

A common mistake is to assume that a prompt must be huge.

What matters is not length itself, but the removal of relevant ambiguity.

A good conversation can also be iterative:

Me: I need to learn global sensitivity analysis for IIO409.
Before explaining it, ask me what I already know.

ChatGPT: What do you understand by the sensitivity of a parameter?

Me: I think it measures how much the output changes when the parameter changes.

ChatGPT: Good. What limitation does that definition have if you vary one parameter while all others stay fixed?

This kind of exchange often produces more learning than asking for an encyclopaedic explanation of Sobol indices.

OpenAI’s Study Mode was designed around a similar logic: interactive questions, scaffolded explanations and checks of understanding. The underlying principle is simple and powerful: if you want to learn, do not ask only for content; ask to be questioned.

4. The instruction you should use for ambiguous problems

A very useful general rule is to tell ChatGPT explicitly that it should ask questions before solving when key information is missing.

For IIO409, I recommend making this a standard instruction:

Do not assume data, methods or objectives that I have not given you.
If missing information could change the solution, ask me before answering.
Ask at most three questions, prioritising the ones with the greatest methodological impact.

For instance:

My simulation obtained KGE = 0.76. Is my model good?

A methodologically serious response should ask, among other things:

good for what purpose?
is this calibration or verification?
which KGE components produce that value?
how does the model behave for the processes or flow ranges that matter?

Hydrological modelling is not about classifying isolated metrics as “good” or “bad”.

5. ChatGPT can still be wrong even when it sounds confident

A fluent reply is not evidence of correctness.

ChatGPT may:

invent a reference;
attribute an equation to the wrong author;
use a non-existent R function;
confuse function arguments;
assume units;
invent numerical values;
misread a figure;
suggest a methodological strategy that is not appropriate for your aim.

That is why you need a verification stage.

Rule 1: separate facts from proposals

Ask:

Separate your answer into:
- facts that can be checked;
- assumptions you are making;
- interpretation;
- recommendation.

Rule 2: request sources when they really matter

For each important scientific claim, provide a verifiable reference.
Do not invent DOI numbers. If you are not sufficiently sure of a reference, say so.

Then check the article yourself.

Do not cite ChatGPT as a substitute for a scientific source.

Rule 3: when dealing with R, run the code

Do not ask only:

Is this code correct?

Run it.

Check:

stopifnot(length(sim) == length(obs))
stopifnot(all(is.finite(sim)))
stopifnot(!anyNA(sim))

Inspect plots. Test the function on a small case with a known behaviour. Check dates and units.

Rule 4: ask for counter-examples

Give me a case in which this recommendation would be wrong.

This is particularly useful in modelling.

Rule 5: ask it to criticise itself

Now act as a critical reviewer.
Identify the three weakest points in your previous answer.

This does not guarantee correctness, but it often reveals hidden assumptions.

6. Do not ask it to think for you: ask it to help you think

If your aim is learning, avoid prompts like:

Solve this exercise.

Prefer:

I want to solve this exercise myself.
Do not give me the final solution.
Ask me one question at a time and only give a hint if I get stuck.

Another useful variation is:

First ask me to explain, in my own words, what I think will happen.
Then challenge my explanation.
Only at the end show me a reference solution.

This completely changes ChatGPT’s function: from answer generator to reasoning trainer.

7. Free or Plus for IIO409?

This question matters, especially when a monthly subscription would be a significant cost for the student’s household.

My recommendation for this course

I do not think students need ChatGPT Plus in order to pass IIO409 or to learn hydrological modelling well.

The free tier is often enough for:

studying concepts;
asking practice questions;
checking short fragments of R;
preparing discussion points;
using Study Mode;
rehearsing oral explanations.

A paid plan can be useful when you work intensively with longer conversations, files, code and repeated tool use. But this is mainly a matter of convenience and productivity, not educational necessity.

When might a paid plan be worth it?

It may make sense if:

you use ChatGPT heavily for programming and debugging;
you regularly work with uploaded files;
you analyse longer documents;
the free-tier limits interrupt your work often;
you are in a particularly intensive phase of the project.

When would I not pay for it?

I would not pay only to:

ask for definitions;
summarise notes;
generate practice questions;
review short code;
brainstorm study ideas.

For those tasks, the quality of the study method usually matters much more than the type of plan.

An equity consideration

No essential learning activity in the course should depend on a student’s ability to pay for a subscription. If two students have the same knowledge and one cannot afford a paid plan, that should not create an academic disadvantage.

Treat a paid plan, therefore, as an optional productivity tool, not as a requirement.

8. Five prompts you can save today

Prompt 1: Socratic tutor

Act as a Socratic tutor for hydrological modelling for a Civil Engineering student.
I want to learn [TOPIC].

First ask me what I currently understand.
Then ask me one question at a time.
Do not give me the solution before I answer.
When I make a mistake, identify the conceptual error and give me a hint.
At the end, give me a slightly different problem to test transfer.

Prompt 2: conceptual checker

I am going to explain [CONCEPT] in my own words.
Do not rewrite it immediately.

Assess my explanation in terms of:
1. correct ideas;
2. incomplete ideas;
3. mistakes;
4. implicit assumptions;
5. one question that reveals whether I really understand it.

Prompt 3: methodological critic

Act as a critical reviewer of a hydrological modelling study.
Do not try to be nice: look for methodological weaknesses.

I will describe my methodology.
Identify problems that could affect:
- data consistency;
- warm-up;
- calibration;
- verification;
- objective function;
- identifiability;
- uncertainty;
- interpretation.

Prioritise the problems by impact and explain how to check each one.

Prompt 4: programming companion

Help me solve this problem in base R, without tidyverse.

Before writing code:
1. summarise the goal;
2. identify inputs and outputs;
3. ask me for any crucial missing information.

Then deliver a minimal reproducible solution.
Include automatic checks and explain how to verify the result.
Do not invent functions or package arguments.

Prompt 5: hallucination detector

Review your previous answer as if you were an auditor.
Flag every claim that depends on:
- a reference;
- a software function;
- a number;
- a date;
- an assumption that I did not provide.

For each one, tell me how I can verify it independently.

9. Checklist before you trust an answer

Before using a reply from ChatGPT in IIO409, ask yourself:

Do I understand why the answer makes sense?
Have I identified its assumptions?
Can I explain it without copying its wording?
Can I think of at least one case in which it would not hold?
Have I checked the references and functions?
Have I run the code?
Have I checked units and dates?
Could I defend this decision in front of my lecturer?

If the answer to several of these is “no”, you are not finished yet.

10. Final idea: using ChatGPT well means asking better questions

The aim is not to memorise a set of “magic prompts”.

The aim is to develop a habit:

define the problem more clearly, make assumptions explicit, ask for evidence, test the result and refine the solution.

That is exactly the same habit you need in good hydrological modelling.

In Part II we will apply this logic to the theory component of IIO409: how to use ChatGPT to study concepts, read articles critically, diagnose modelling cases and prepare for assessments in which memorising definitions is not enough.

How to Use ChatGPT to Learn Hydrological Modelling (II): conceptual understanding and critical thinking

Sun, 16 Aug 2026 00:00:00 +0000

Estimated reading time: 18-22 minutes

In Part I we looked at setting up ChatGPT, giving it context, writing clearer prompts and checking its answers. Now the goal changes.

Here ChatGPT is not an encyclopaedia. It becomes a tutor, examiner, critical interlocutor and case generator.

The theory component in IIO409 is designed precisely to reduce the value of memorising definitions and to increase the value of understanding, interpreting, diagnosing and arguing:

\[ Theory = 0.40\,ECT + 0.20\,LCA + 0.25\,DMC + 0.15\,PC \]

where ECT stands for Conceptual Theory Assessments, LCA for Critical Reading of Scientific Articles, DMC for Modelling Diagnosis through Case Studies, and PC for Scientific Participation in Class.

The implication is important: if you use ChatGPT only to obtain ready-made answers, you are using it in the least useful way for this assessment design.

1. Your goal is not to remember: it is to build a mental model

Suppose you are studying equifinality.

You can ask:

What is equifinality?

You will get a definition.

But a definition is not the same thing as understanding.

Try this sequence instead:

Explain equifinality with an example from a conceptual rainfall-runoff model.

Then:

Now give me two parameter sets that could produce similar hydrographs
but different internal states. I do not need real numbers; I want to understand the mechanism.

Then:

Why might this problem remain hidden if I look only at NSE or KGE?

Then:

Now give me one plausible but incorrect statement about equifinality.
Do not tell me why it is wrong until I try to criticise it.

Finally:

Change the problem: now the aim is not total discharge reproduction, but low-flow analysis.
Does the importance of equifinality change? Ask me first what I think.

This turns a static definition into a structure of reasoning.

2. The Socratic ladder

OpenAI’s Study Mode is based on interactive questioning, scaffolded explanations and checks of understanding. Even without any special mode, you can reproduce that logic in your own prompting.

Use five levels.

Level 1: explain

Explain the concept with an example from a Chilean catchment.

Level 2: retrieve

Stop the explanation and ask me five questions, one at a time.
Do not show the answer until I reply.

Level 3: challenge

Look for weaknesses in my replies.
Do not correct only the outcome: identify the conceptual error.

Level 4: transfer

Change one important condition of the problem and ask whether my reasoning still holds.

Level 5: defend

Act as a lecturer in an oral defence.
Question my assumptions and ask for a justification of each decision.

If you can answer successfully at level 5, your understanding is probably much stronger than if you can merely repeat a definition.

3. ECT (40%): Conceptual Theory Assessments

ECT tasks will prioritise interpretation and reasoning.

To prepare well, you should not simply ask ChatGPT for “exam questions”. Instead, ask it for the type of cognitive work the assessment values.

3.1 Interpreting results

Prompt:

Generate a short hydrological modelling case with:
- performance in calibration and verification;
- bias;
- one efficiency metric;
- one visible feature of the hydrograph.

Ask me three interpretation questions.
Do not give the solutions until I answer.

Then ask:

Evaluate my reply by separating:
- observation;
- inference;
- conclusion that is not yet justified.

This distinction is extremely useful. Many student errors occur when an observation is turned immediately into a causal claim.

For example:

“The model underestimates peaks.”

is an observation.

“Precipitation must be underestimated.”

is a hypothesis, not a necessary conclusion.

3.2 Comparing methodological alternatives

Try:

Compare two calibration strategies:
A) optimise NSE;
B) optimise a metric that gives greater weight to low flows.

Do not tell me which one is better.
Set up a use case and ask me which I would choose and why.
Then change the use case.

The important idea here is that a methodological decision depends on the purpose of the model.

3.3 Working with scenarios

I have a model calibrated in a wet period.
Create a dry-period verification scenario.
Ask me what problems I would expect and what diagnostics I would perform.

Then:

Give me an alternative explanation to mine.

Learning to produce alternative explanations is essential if you want to avoid overconfident conclusions.

3.4 Finding assumptions

Ask:

Give me an apparently correct methodology for calibrating a conceptual model,
but hide five implicit assumptions.
Do not flag them.
My task will be to identify them.

Then compare your list with ChatGPT’s and discuss any differences.

4. How to study each main content block in IIO409

4.1 Hydrological models and conceptual models

Do not memorise only classifications.

You should be able to answer:

Which process is each store trying to represent?
Which processes are absent?
Which observations could discriminate between two structures?
Which parts of the discharge behaviour depend on which model components?

Prompt:

Describe a catchment with climate, seasonality, snow/no snow, vegetation,
geology and discharge regime.

Do not choose a model for me.
Ask me what processes I would include in my conceptual model and why.
Then critique my representation.

A more demanding exercise:

Propose two different conceptual models that could reproduce the same discharge behaviour.
Ask me what extra data I would use to discriminate between them.

4.2 Hydrometeorological data

The question is not only “What data exist?”, but:

Are these data suitable for the modelling question?

Try:

Give me a hypothetical data set for a Chilean catchment with five quality problems:
missing dates, inconsistent units, bias, station change or temporal-resolution problems.

Do not tell me the problems. Give me clues and ask me to diagnose them.

Or:

Compare an in situ source, a satellite product and a reanalysis for precipitation.
Build the comparison in terms of:
- resolution;
- representativeness;
- biases;
- coverage;
- implications for calibration.

Then give me one case where the apparently most precise option is not the best.

4.3 Sensitivity: from “which parameter matters” to “why”

For Sobol analysis you should understand at least the conceptual difference between first-order effects and interactions.

Prompt:

Generate hypothetical Sobol results for four parameters.
Include first-order and total indices.

Do not interpret the table.
Ask me:
1. which parameter dominates;
2. which one shows interaction;
3. what this implies for calibration;
4. what I cannot conclude from these indices.

A deeper question:

Can a weakly sensitive parameter still be physically important?
Make me argue both for and against.

This helps you separate sensitivity from “physical importance”, which are not the same thing.

4.4 Calibration and verification

Study calibration as a problem of inference and decision, not merely “finding the best parameter values”.

Prompt:

Give me a case in which a calibration with PSO obtains an excellent objective value,
but the model is still not adequate.

Include enough clues for me to diagnose the problem.
Do not give me the solution yet.

You should be able to discuss:

objective function;
warm-up period;
dependence on the calibration period;
parameter bounds;
overfitting;
stability across periods;
multiple optima;
equifinality;
the relation between purpose and metric.

Make ChatGPT challenge you:

I am going to defend this claim:
"The parameter set with the best KGE is the true parameter set of the catchment".

Argue with me as a lecturer.
Do not accept claims without justification.

4.5 Uncertainty and GLUE

An uncertainty band does not speak for itself.

Prompt:

Create a conceptual GLUE example with:
- 10,000 simulations;
- an acceptance threshold;
- a percentage of behavioural simulations;
- a predictive band.

Ask me to interpret what each element means.
Then ask me which sources of uncertainty are NOT necessarily represented.

Another important question:

If I increase or reduce the GLUE threshold, what effects should I expect?
Do not answer: make me predict first.

4.6 Regionalisation

Prompt:

I have an ungauged catchment and three potential donor catchments.
Create climatic, topographic and hydrological-signature information for the four.

My task will be to choose a regionalisation strategy.
Then question my similarity criterion.

You need to understand that geographical proximity and hydrological similarity are not synonyms.

4.7 Climate change

Here the key is to think in terms of a modelling chain.

Prompt:

Build a conceptual chain from climate scenarios to future discharge:
climate model -> downscaling/bias correction -> meteorological variables -> hydrological model -> indicators.

Hide three sources of uncertainty.
Ask me to identify them and explain how they propagate.

Then:

Now change just one methodological decision in the chain.
Ask me how the conclusions might change.

5. LCA (20%): Critical Reading of Scientific Articles

ChatGPT can summarise papers. That is probably one of the least demanding ways to use it.

For LCA you need to go further.

Step 1: read first yourself

Before asking ChatGPT anything, identify:

the problem;
the question;
the hypothesis;
the data;
the method;
the results;
the main conclusion.

Then use ChatGPT to compare with your interpretation.

I am going to give you a paper.
Do not summarise it first.

Ask me these questions one at a time:
1. What is the scientific question?
2. What is the hypothesis?
3. What evidence could falsify it?
4. Which methodological decisions are critical?
5. Which limitations affect the conclusion?

Step 2: separate result from interpretation

Build a table with:
- observed result;
- authors' interpretation;
- another plausible interpretation;
- evidence needed to distinguish between them.

Step 3: search for assumptions

Identify five methodological assumptions in the paper.
For each one, indicate which conclusion might change if the assumption fails.

Step 4: formulate a new question

Based only on the paper, propose five possible follow-up research questions.
Order them from a small extension to a genuinely new conceptual question.

Then decide yourself which one is scientifically meaningful.

Step 5: prepare an oral presentation

Do not ask:

Make my slides.

Instead:

Help me identify the three essential scientific messages.
Ask me first what I think is most important before proposing them.

The presentation should be a consequence of your understanding, not a substitute for it.

6. DMC (25%): Modelling Diagnosis through Case Studies

This component is especially suitable for studying with ChatGPT because it can generate endless variations of a case.

Base prompt:

Act as an IIO409 lecturer.

Create a realistic rainfall-runoff modelling case with at least four methodological problems.
The problems may involve data, warm-up, calibration, verification,
objective function, sensitivity, parameters, uncertainty or interpretation.

Do not mark the errors.
Give me only the case and then ask:
1. which problems I detect;
2. which one is most serious;
3. what consequences it would have;
4. how I would correct it.

After you answer:

Do not give me your full solution yet.
Tell me only which important problem I failed to detect.

Finally:

Now provide your full diagnosis and compare it with mine.

Make the cases progressively harder

Beginner level:

calibration and verification in the same period;
incorrect units;
no warm-up period.

Intermediate level:

objective function inconsistent with the modelling aim;
parameters at the bounds;
climate-change analysis with an inconsistent baseline.

Advanced level:

good aggregate metric but structured errors;
parameter compensation;
inconsistency between internal states and discharge;
parametric uncertainty presented as total uncertainty.

7. PC (15%): Scientific Participation in Class

Participation is not the same thing as talking often. It means making a contribution that improves the discussion.

ChatGPT can help before class.

Suppose the topic is parameter identifiability.

Tomorrow we will discuss parameter identifiability.

Give me:
- three conceptual questions that could open a discussion;
- two debatable statements;
- one example in which high sensitivity does not guarantee identifiability;
- one counter-argument to each statement.

Do not memorise the questions. Choose one that you actually understand and rephrase it in your own words.

You can also use ChatGPT to prepare hypotheses:

Give me five possible explanations for a parameter that changes strongly between two periods.
Classify them as:
- data;
- structure;
- optimisation;
- non-stationarity.

A scientifically useful classroom contribution might be:

“If the parameter changes between periods, how do we distinguish weak identifiability from structural compensation within the model?”

That connects concepts and opens discussion.

8. Feynman technique with ChatGPT

A good test of understanding is to explain a concept in your own words.

Prompt:

I am going to explain GLUE as if I were teaching it to a classmate who is just beginning.

Do not interrupt me.
When I finish:
1. identify what is correct;
2. find ambiguities;
3. detect important missing ideas;
4. ask me one question that my explanation cannot yet answer.

Then rewrite the explanation yourself.

Do not ask ChatGPT to rewrite it for you.

9. The “predict before you ask” method

This is one of the best study strategies for hydrological modelling.

Before consulting ChatGPT, write your own prediction:

If I increase a parameter associated with storage, I expect that...

Then ask:

Evaluate my prediction.
Do not focus only on whether the direction is correct.
Discuss under which conditions it could fail.

This prevents the passive habit of allowing ChatGPT to think first every time.

10. Build your own bank of mistakes

Each time you answer something incorrectly, record:

Concept:
My answer:
Error:
Why I got it wrong:
Correct principle:
New example:

After several weeks, ask:

Here is my error log.
Do not explain the topics again.
Identify patterns and create five new questions targeted specifically at my weaknesses.

This turns ChatGPT into an adaptive practice tool.

11. Master prompt for preparing an ECT

Act as a Civil Engineering lecturer in Hydrological Modelling.

Topic: [TOPIC].

I want to prepare for a conceptual assessment.
Do not test memorisation of definitions.

Generate five questions with increasing difficulty:
1. interpretation;
2. methodological comparison;
3. cause-effect;
4. diagnosis;
5. transfer to a new scenario.

Ask me one question at a time.
Wait for my answer.
Assess the reasoning, not only the conclusion.
When my answer is incomplete, use a hint before giving the solution.
At the end, identify my two main conceptual weaknesses.

12. Master prompt for DMC

Create a realistic hydrological diagnosis case for IIO409.

Include information about:
- data;
- model;
- warm-up;
- calibration;
- verification;
- objective function;
- parameters;
- results;
- uncertainty.

Introduce between three and five methodological problems of different severity.

Do not label them.
Ask me to:
1. identify them;
2. prioritise them;
3. explain the consequences;
4. propose corrections;
5. defend one alternative.

Then question my diagnosis as in an oral defence.

13. What not to do when studying theory with ChatGPT

Avoid turning your studying into:

Summary -> passive reading
Flashcards -> repetition without context
Automatic answers -> false confidence
AI-generated essay -> no practice of reasoning
Unquestioned explanations -> learning mistakes

The key question is always:

Can I apply this idea in a new case without ChatGPT telling me what to do?

If the answer is no, you are still studying too passively.

14. Before an assessment

Try this 45-minute routine:

10 min: retrieval

Without ChatGPT, write down everything you remember about the topic.

15 min: Socratic questioning

Use ChatGPT to identify gaps.

10 min: new case

Ask for a problem you have not seen before.

5 min: oral explanation

Explain it out loud without looking.

5 min: error log

Write down exactly what you did not understand.

This is far more useful for IIO409 than asking ChatGPT for “a short summary for the test”.

15. The goal: learn to think like a modeller

A good modeller is not defined by the ability to recall the longest definition.

A good modeller asks:

What does this model represent?
What does it not represent?
What evidence supports this decision?
What alternatives exist?
What assumption am I making?
How might I be wrong?
Which observation would distinguish between competing explanations?
Does the conclusion depend on the purpose?
What uncertainty remains?

ChatGPT can help you practise those questions again and again.

But only if you use it deliberately: make it question you, contradict you, change the conditions and demand that you justify your choices.

In Part III we will take the same logic into the practical side of the course: base R, vibe coding, reproducibility, debugging, sensitivity, calibration, uncertainty and preparation for the practical challenges.

How to Use ChatGPT to Learn Hydrological Modelling (III): base R, vibe coding and reproducible projects

Sun, 16 Aug 2026 00:00:00 +0000

Estimated reading time: 20-25 minutes

In this third part we will use ChatGPT as a programming companion for solving hydrological-modelling problems in base R, without tidyverse.

We assume that you have already taken an R course and know the fundamentals: objects, vectors, matrices, data.frame objects, functions, conditionals, loops, basic plotting and the use of packages.

The goal here is not to teach R from the beginning. It is to learn a new workflow:

describe -> generate -> run -> check -> interpret -> modify -> document

This is often called vibe coding: conversational programming with AI support. It can speed up work considerably, but it can also produce convincing-looking code that is conceptually wrong.

In IIO409 the rule should be:

AI may write code. You still remain responsible for knowing what problem the code solves, why it solves it in that way, and how to check whether the result is correct.

1. What the practical component really assesses

The practical component is:

\[ Practical = 0.25\,PRE + 0.10\,DTM + 0.35\,DPI + 0.10\,PFG + 0.20\,DTI \]

where:

PRE = Reproducible Team Project;
DTM = Technical Modelling Dossier;
DPI = Individual Practical Challenges;
PFG = Final Group Presentation;
DTI = Individual Technical Defence.

In addition, to pass the practical component you must achieve at least 4.0 in the combined average of the individual components DPI and DTI.

This completely changes how it makes sense to use ChatGPT.

If ChatGPT writes a perfect script but you cannot modify it during an individual challenge or explain a modelling decision in the defence, then the code did not help you learn.

2. Responsible vibe coding

A good workflow has seven steps.

Step 1: specify the problem

Do not begin with:

Write me code to calibrate a model.

First define:

inputs;
outputs;
units;
temporal resolution;
model;
objective;
constraints;
success criterion.

Example:

I have daily precipitation [mm/day], mean temperature [°C] and discharge [m3/s]
for a catchment of 850 km2.

I want a base R function that:
1. aligns the series by date;
2. detects missing values;
3. converts discharge to mm/day;
4. returns a data.frame ready for modelling.

Do not use tidyverse.
Before programming, tell me what additional information you need.

Step 2: ask for a minimal solution

Do not ask for 300 lines of code.

Implement the minimum working version first.
Use base R functions.
Do not add optimisation or plotting yet.

Small solutions are easier to verify.

Step 3: run it immediately

Never read a long script and assume it works.

Copy the minimum version, run it and inspect:

str(x)
head(x)
summary(x)

Check sizes:

dim(x)
length(x)

Check missing values:

colSums(is.na(x))

Step 4: inspect the meaning, not only the object

A vector can have the right length and still be conceptually wrong.

Ask:

Are dates aligned?
Are the units correct?
Is the time ordering correct?
Are there duplicates?
Are the values physically plausible?
Have missing values been handled explicitly?

Step 5: test

Ask:

Now create three simple tests that will fail if:
- dates are unsorted;
- discharge is negative;
- the lengths do not match.

For example:

stopifnot(!is.unsorted(dat$date))
stopifnot(all(dat$q >= 0 | is.na(dat$q)))
stopifnot(nrow(dat) == length(dat$date))

Step 6: explain

Explain each decision in the code that could affect the hydrological result.
Do not describe trivial syntax.

This distinction matters.

read.csv() is syntax.

Converting m3/s to a daily depth over the catchment is a quantitative decision that you must understand.

Step 7: refactor

Only afterwards:

Turn the solution into reusable functions.
Add argument validation.
Keep base R and minimal dependencies.

3. Core R skills you still need

ChatGPT can help with syntax, but you must still recognise and understand the following structures.

3.1 Paths and project structure

Avoid:

setwd("C:/Users/my_name/Desktop/Project")

Prefer a project directory and relative paths.

file.path("data", "precipitation.csv")

Ask ChatGPT:

Review this project and identify absolute paths that harm reproducibility.
Suggest replacements using file.path().

3.2 Reading and writing data

You should be comfortable with:

dat <- read.csv("data/input.csv")
write.csv(dat, "results/output.csv", row.names = FALSE)

saveRDS(dat, "results/data_processed.rds")
dat2 <- readRDS("results/data_processed.rds")

Always ask:

What information might be lost if I save this object as CSV rather than RDS?

3.3 Dates

dat$date <- as.Date(dat$date)

For date-time values:

dat$datetime <- as.POSIXct(
 dat$datetime,
 tz = "UTC",
 format = "%Y-%m-%d %H:%M:%S"
)

Date errors can ruin a calibration without necessarily triggering a coding error.

Ask:

Give me tests that help verify whether two hydrological series really represent the same days.

3.4 Subsetting

idx <- dat$date >= as.Date("2000-01-01") &
 dat$date <= as.Date("2010-12-31")

cal <- dat[idx, ]

You should be able to define explicit periods for:

warm-up;
calibration;
verification.

3.5 Missing data

Do not replace NA automatically.

First inspect:

colSums(is.na(dat))

Then decide what to do.

Ask:

I have 2% missing precipitation values.
Do not recommend interpolation immediately.
Ask me about duration, temporal distribution and modelling purpose before suggesting a treatment.

3.6 Functions

The basic unit of a reproducible project should be a clear function.

run_experiment <- function(par, input) {
 stopifnot(is.numeric(par))
 stopifnot(is.data.frame(input))

 # run model
 sim <- ...

 sim
}

Ask ChatGPT to avoid monolithic scripts:

Split this script only when there is a clear functional responsibility.
Do not create unnecessary functions.

3.7 `apply`, `lapply` and loops

You should understand both styles.

out <- lapply(files, read.csv)

or:

out <- vector("list", length(files))

for (i in seq_along(files)) {
 out[[i]] <- read.csv(files[i])
}

There is no need to replace all loops with apply functions. The criterion is readability and correctness.

3.8 Base graphics

For diagnostics:

plot(obs, sim,
 xlab = "Observed",
 ylab = "Simulated")

abline(0, 1)

Time series:

plot(dat$date, dat$obs, type = "l")
lines(dat$date, dat$sim)

Residuals:

res <- dat$sim - dat$obs
plot(dat$date, res, type = "h")
abline(h = 0)

Ask ChatGPT:

Do not suggest a plot just because it looks nice.
Tell me what hydrological diagnostic question each plot helps answer.

3.9 Randomness control

For sensitivity, calibration or uncertainty analyses:

set.seed(1234)

The seed should be recorded whenever it matters.

3.10 Error handling

if (!all(required %in% names(dat))) {
 stop("Missing required columns.")
}

For operations that may fail:

ans <- tryCatch(
 run_model(par),
 error = function(e) NA
)

But do not hide errors indiscriminately.

Ask:

In what situations would tryCatch() hide a methodological problem that should actually stop the analysis?

3.11 Packages

Use explicit namespace calls when they improve clarity:

hydroGOF::KGE(sim, obs)

Before trusting a function suggested by AI, check the documentation:

help(package = "hydroGOF")
?hydroGOF::KGE

or consult the official package documentation.

ChatGPT can invent function names or arguments. The documentation is the authority.

3.12 Recording the environment

At the end of an analysis:

sessionInfo()

Save it.

Also record:

R version;
packages;
seed;
parameters;
periods;
input files;
execution date when relevant.

4. Structure of a reproducible project

A possible structure is:

IIO409_project/
|
|-- 00_data/
| |-- raw/
| `-- processed/
|
|-- 01_scripts/
| |-- 01_read_data.R
| |-- 02_qc.R
| |-- 03_model.R
| |-- 04_sensitivity.R
| |-- 05_calibration.R
| |-- 06_verification.R
| `-- 07_uncertainty.R
|
|-- 02_results/
|-- 03_figures/
|-- 04_docs/
|-- config.R
`-- README.md

ChatGPT can help you audit this structure:

Act as a reproducibility auditor.
I will paste the file tree and the main scripts.
Identify:
- absolute paths;
- manual steps;
- objects created outside scripts;
- parameters written in multiple places;
- undocumented dependencies;
- result files that cannot be rebuilt.

5. Centralised configuration

Avoid repeating key parameters across many scripts.

cfg <- list(
 warmup_start = as.Date("2000-01-01"),
 calibration_start = as.Date("2001-01-01"),
 calibration_end = as.Date("2010-12-31"),
 validation_start = as.Date("2011-01-01"),
 validation_end = as.Date("2020-12-31"),
 seed = 1234L
)

Then:

source("config.R")

This improves traceability.

6. Working with hydrometeorological data

A sensible workflow is:

read -> validate -> sort -> align -> convert units -> diagnose -> model

Prompt:

I have these columns:
date, p_mm, t_c, q_m3s.

Before modelling, create a quality-control checklist.
For each check:
- explain why it matters hydrologically;
- provide base R code;
- indicate which action should NOT be automated without inspection.

Examples of checks:

stopifnot(!anyDuplicated(dat$date))
stopifnot(!is.unsorted(dat$date))

summary(dat$p_mm)
summary(dat$t_c)
summary(dat$q_m3s)

Negative precipitation:

which(dat$p_mm < 0)

Negative discharge:

which(dat$q_m3s < 0)

Do not delete suspicious values automatically. Investigate them.

7. Converting discharge to a depth

This is a good exercise in both concept and programming.

Ask:

Do not give me the formula immediately.
Make me derive how to convert mean daily discharge [m3/s] into runoff depth [mm/day]
for a catchment with area A [km2].
Check the dimensional consistency of each step.

Then implement it yourself:

q_to_mm_day <- function(q_m3s, area_km2) {
 stopifnot(area_km2 > 0)

 volume_m3_day <- q_m3s * 86400
 area_m2 <- area_km2 * 1e6

 (volume_m3_day / area_m2) * 1000
}

Test it dimensionally and with simple cases.

8. Implementing the hydrological model

For TUWmodel, GR4J or any other model used in the course, avoid asking AI to re-implement everything from scratch if a trusted function or lecturer-provided code already exists.

A better use is:

Here is the documentation of the function we are going to use.
Do not invent arguments outside this documentation.

Explain:
1. inputs;
2. outputs;
3. the meaning of each parameter;
4. units;
5. relevant internal states;
6. three checks to perform before running it.

Then:

Write a base R wrapper to run the model and return:
- the simulation;
- metrics;
- parameters;
- the period.
Include argument validation.

Your wrapper should make the project’s modelling decisions explicit.

9. Sensitivity analysis

Do not begin by asking only:

Do Sobol.

First define:

parameters;
ranges;
output of interest;
sample size;
metric;
seed;
computational cost.

Prompt:

I want to perform a global Sobol sensitivity analysis.

Before writing code, build with me a decision table with:
parameter, range, justification, output, sample size, seed and expected cost.

Do not choose values arbitrarily without marking them as assumptions.

After you run it:

Here are my first-order and total indices.
Do not interpret them immediately.
Ask me four questions to check whether I understand interaction and dominance.

10. Calibration with PSO

AI can help you configure the experiment, but it should not silently decide:

bounds;
objective function;
period;
number of particles;
stopping rule.

Create an explicit objective function.

Skeleton:

objective_function <- function(par, input, obs) {
 sim <- run_model(par, input)

 if (length(sim) != length(obs)) {
 stop("sim and obs have different lengths")
 }

 score <- ...

 score
}

Prompt:

Review this objective function.
Do not change the code yet.

Look for:
- information leakage;
- NA values;
- length mismatches;
- max/min orientation;
- wrong period;
- warm-up dependence;
- units.

If you use hydroPSO, check the documentation of the version actually installed.

Do not trust an argument merely because ChatGPT suggested it confidently.

11. Calibration is not validation

One useful check is:

stopifnot(max(cal_dates) < min(val_dates))

This is only one example, but it captures the idea: encode important decisions as tests.

Ask:

Give me five programmatic tests that reduce the risk of accidentally using
validation data during calibration.

12. Metrics and diagnosis

Do not rely on a single metric.

You may calculate metrics with appropriate packages or verified functions and complement them with diagnostic plots.

Ask:

I have NSE, KGE and PBIAS for calibration and verification.
Do not classify the model yet.
Tell me what additional information I should inspect for three different aims:
1. water balance;
2. low flows;
3. flood peaks.

The same simulation can be useful for one aim and inadequate for another.

13. GLUE and uncertainty

A typical computational workflow involves:

generating parameter sets;
running the model;
calculating a performance measure;
applying a behavioural threshold;
analysing the distribution of results;
building diagnostics.

Ask ChatGPT to work in blocks:

Do not write the whole GLUE analysis in one script.

First design the functions we need and their interfaces.
Then we will implement and test them one by one.

Conceptual example:

evaluate_par <- function(par, input, obs) {
 sim <- run_model(par, input)
 score <- ...
 c(score = score)
}

Before running thousands of simulations, test five.

set.seed(1234)

Then ask:

Which problems might not appear with five simulations but could appear with 100,000?

This introduces issues such as memory, execution time, single-run failures and result storage.

14. Regionalisation

For regionalisation you will probably combine catchment attributes and hydrological signatures.

You should be able to handle basic matrix and data.frame operations.

Example:

attrs <- data.frame(
 basin = c("A", "B", "C"),
 p_mean = c(1200, 950, 1300),
 elev = c(850, 620, 910)
)

d <- dist(scale(attrs[, c("p_mean", "elev")]))
as.matrix(d)

Critical question:

What is the limitation of using scale() and Euclidean distance if attributes have
different distributions, relevance or collinearity?

Programming should lead into methodological discussion.

15. Climate change

Here reproducibility becomes essential because the modelling chain is long.

You can ask:

Design a file structure and a set of functions for a workflow:
climate data -> preprocessing -> bias correction/downscaling -> hydrological model
-> indicators -> period comparison.

Do not write the code yet.
First identify which metadata must be kept to reproduce each step.

Keep separate:

original data;
processed data;
scenarios;
outputs;
indicators.

Never overwrite the original inputs silently.

16. Debugging: provide the full error context

Do not write:

It does not work.

Provide:

1. the goal;
2. the minimal code;
3. the full error message;
4. the expected result;
5. the obtained result;
6. the structure of the objects involved.

Prompt:

This is a minimal reproducible example.
Do not rewrite everything.

1. identify the most likely cause;
2. explain which part of the error message supports it;
3. propose the smallest possible change;
4. give me a test that verifies the fix.

The “smallest possible change” prevents unnecessary rewrites of the whole project.

17. When the code runs but the result is wrong

This is the most dangerous situation.

Ask:

The code runs without errors, but the hydrograph looks suspicious.
Give me an ordered diagnostic strategy from the simplest checks
to structural problems in the model.
Do not assume the issue is calibration.

A sound diagnosis should consider:

data;
time alignment;
units;
warm-up;
PET;
parameters;
objective function;
implementation;
conceptual structure.

18. Ask for tests before you optimise

Prompt:

Here is my run_model() function.
Before optimising anything, create a small set of tests.

I want to check:
- arguments;
- dimensions;
- NA handling;
- reproducibility;
- behaviour with out-of-range parameters;
- unit consistency.

You do not need a huge testing framework to benefit from this.

19. PRE (25%): Reproducible Team Project

Use ChatGPT as an auditor, not as an invisible author.

Once a week:

Audit the reproducibility of our current project.

I will give you:
- the file tree;
- README;
- config;
- scripts.

Look for steps that only work because we know manually what to do.

Another helpful review:

Could a third person run this project from scratch?
List everything they would still have to guess.

That prompt usually reveals many weaknesses.

20. DTM (10%): Technical Modelling Dossier

The dossier should explain decisions.

A simple table helps:

Decision	Choice	Alternative	Justification	Consequence
Warm-up	…	…	…	…
Objective function	…	…	…	…
Parameter range	…	…	…	…
GLUE threshold	…	…	…	…

ChatGPT can challenge the table:

Act as a reviewer.
For each row in this table, formulate one question that would force the team
to justify the decision with evidence or hydrological reasoning.

Do not ask it to invent the justification for you.

21. DPI (35%): Individual Practical Challenges

This is the most important practical component.

Prepare by practising variations, not by repeating the same script.

Here is an exercise that I have already solved.
Create a variation that changes two important conditions.
Do not only change the numbers.

I want it to force me to modify the code and reinterpret the result.

Examples:

a different period;
missing data;
another objective function;
a parameter at its bound;
a change of temporal resolution;
a new catchment;
a different hydrological aim.

To practise properly:

Do not give me code for the first 15 minutes.
Only answer conceptual questions and review my plan.

22. PFG (10%): Final Group Presentation

ChatGPT can help you remove information, not only add it.

We have 12 results.
Our scientific message is [MESSAGE].

Help us select the four results that are strictly necessary.
For each result we exclude, explain why it is not essential.

Then:

Formulate five difficult questions that a lecturer might ask after the presentation.

23. DTI (20%): Individual Technical Defence

This is probably one of the best ways to use ChatGPT before assessment.

Act as an oral examiner for IIO409.

I will describe our project.
Ask me individual questions about:
- decisions;
- parameters;
- sensitivity;
- calibration;
- performance;
- uncertainty;
- limitations;
- alternative scenarios.

Ask one question at a time.
If my answer is superficial, go deeper.
Do not praise me or give hints too early.

Then:

Change one decision in the project and ask what consequences I would expect.

That is exactly the kind of reasoning you will need in the defence.

24. Git and version control

For a team project, Git is strongly recommended.

Minimum skills:

git status
git add
git commit
git pull
git push

ChatGPT can help interpret errors, but do not execute destructive commands that you do not understand.

Prompt:

Explain this Git conflict.
Do not give me a resolution command yet.
First explain what happened and what information I could lose with each alternative.

25. Minimum README

Your README.md should explain:

1. objective;
2. data;
3. structure;
4. requirements;
5. order of execution;
6. generated outputs;
7. configurable decisions;
8. authors/contact.

Prompt:

Audit this README from the perspective of a person who has never seen the project.
What information is still missing to reproduce it?

26. A template prompt for almost any R problem

Save this one:

Act as a scientific programmer with expertise in R and hydrological modelling.

CONTEXT
[describe data, model, units, period and objective]

TASK
[what you need to achieve]

CONSTRAINTS
- base R; no tidyverse.
- Minimal dependencies.
- Do not invent package functions or arguments.
- Reproducible code.
- Validate inputs.
- Use set.seed() if there is randomness.
- Do not hide important errors.

PROCESS
Before writing code:
1. summarise the problem;
2. identify crucial missing information;
3. propose a minimal strategy.

Then:
4. deliver the code;
5. create a small example;
6. add checks;
7. explain how to validate the result hydrologically;
8. identify possible failure modes.

27. The most important prompt for learning code written by AI

After you receive any script, ask:

Now remove the finished solution from your explanation.

Make me reconstruct this code block by block.
Ask me what each block should do before showing it.

If you cannot reconstruct it, you probably have not learned it.

28. Final rules for vibe coding in IIO409

Never run 10,000 simulations before testing 5.
Never calibrate before checking data, units and dates.
Never use a function suggested by AI without checking the documentation.
Never accept a metric without looking at the hydrological behaviour.
Never confuse code that runs with methodology that is sound.
Never submit a script that you cannot modify.
Never present a decision that you cannot defend.
Keep seeds, configuration, versions and metadata.
Separate raw data, processed data and results.
Use ChatGPT to test and question your work, not merely to produce it.

The professional skill we want is not simply the ability to write R faster.

It is the ability to build a modelling chain whose logic, code, data, results and limitations you can explain and reproduce.

If you use ChatGPT in this way, AI can speed up mechanical tasks while you concentrate effort on what really matters: thinking like a hydrological modeller.

IIO409 | Dr. Mauricio Zambrano-Bigiarini

How to Use ChatGPT to Learn Hydrological Modelling (I): set-up, prompts and trustworthy answers

Before we begin: learning in an age of rapidly changing AI

1. Before you ask: set up the context

Why does this help?

2. The R-C-T-F structure: Role + Context + Task + Format

R: Role

C: Context

T: Task

F: Format

3. A long prompt is not necessarily a good prompt

4. The instruction you should use for ambiguous problems

5. ChatGPT can still be wrong even when it sounds confident

Rule 1: separate facts from proposals

Rule 2: request sources when they really matter

Rule 3: when dealing with R, run the code

Rule 4: ask for counter-examples

Rule 5: ask it to criticise itself

6. Do not ask it to think for you: ask it to help you think

7. Free or Plus for IIO409?

My recommendation for this course

When might a paid plan be worth it?

When would I not pay for it?

An equity consideration

8. Five prompts you can save today

Prompt 1: Socratic tutor

Prompt 2: conceptual checker

Prompt 3: methodological critic

Prompt 4: programming companion

Prompt 5: hallucination detector

9. Checklist before you trust an answer

10. Final idea: using ChatGPT well means asking better questions

Recommended references and further reading

How to Use ChatGPT to Learn Hydrological Modelling (II): conceptual understanding and critical thinking

1. Your goal is not to remember: it is to build a mental model

2. The Socratic ladder

Level 1: explain

Level 2: retrieve

Level 3: challenge

Level 4: transfer

Level 5: defend

3. ECT (40%): Conceptual Theory Assessments

3.1 Interpreting results

3.2 Comparing methodological alternatives

3.3 Working with scenarios

3.4 Finding assumptions

4. How to study each main content block in IIO409

4.1 Hydrological models and conceptual models

4.2 Hydrometeorological data

4.3 Sensitivity: from “which parameter matters” to “why”

4.4 Calibration and verification

4.5 Uncertainty and GLUE

4.6 Regionalisation

4.7 Climate change

5. LCA (20%): Critical Reading of Scientific Articles

Step 1: read first yourself

Step 2: separate result from interpretation

Step 3: search for assumptions

Step 4: formulate a new question

Step 5: prepare an oral presentation

6. DMC (25%): Modelling Diagnosis through Case Studies

Make the cases progressively harder

7. PC (15%): Scientific Participation in Class

8. Feynman technique with ChatGPT

9. The “predict before you ask” method

10. Build your own bank of mistakes

11. Master prompt for preparing an ECT

12. Master prompt for DMC

13. What not to do when studying theory with ChatGPT

14. Before an assessment

10 min: retrieval

15 min: Socratic questioning

10 min: new case

5 min: oral explanation

5 min: error log

15. The goal: learn to think like a modeller

Recommended references and further reading

How to Use ChatGPT to Learn Hydrological Modelling (III): base R, vibe coding and reproducible projects

1. What the practical component really assesses

2. Responsible vibe coding

3.7 `apply`, `lapply` and loops