Simulation_Improvement_Tips.docx

Scientific Inquiry with AI • Creating AI-Generated Simulations • Simulation Improvement Playbook Page 1
Simulation Improvement Playbook
A reusable checklist for making any simulation correct, explorable, and
worth keeping
Scientific Inquiry with AI
Additional Support File — V1
Created by LLNL Summer 2026 STEM Education Research Team
Student Team: Ramina Amino, Jahanvi Chamria, Arya Ferozy, Bryanna Gonzalez, Tai
Le, Zedikiah McAdams, John Navarra, Joshua Sarabia, Abdurrahman Raza
Faculty Team: Praveen Pathak, David Rakestraw, David Strubbe, Brian Utter
V1

Scientific Inquiry with AI • Creating AI-Generated Simulations • Simulation Improvement Playbook Page 2
How to use this. Once your simulation runs and passes its validation checks, walk this list to
make it genuinely useful. It's organized by category. Each suggestion is a Do paired with the
Don't it replaces, tagged by priority, with a copy-paste prompt to make the change. You'll
reuse this on every simulation you build, whatever the topic, science or not, not just this one.
Tier Meaning
MUST Do these or the simulation isn't trustworthy or usable. Non-negotiable.
SHOULD Do these to make it a real instrument for inquiry, not just a demo.
NICE Polish that pays off when you reuse or share it.
1. Scientific Validity & Trust
Tier Do / Don't Prompt to get there
MUST Do: Use the full governing model.
Don't: Let the AI silently substitute a
simplifying approximation (e.g.
small-angle, ideal-gas, linearization)
without telling you.
Use the full governing rules for [phenomenon].
Do not apply any simplifying approximation
unless I ask. State in a comment any
assumption you made.
MUST Do: Show a live numeric readout of the
quantity your validation depends on.
Don't: Hide the result behind animation,
so you can only eyeball it.
Display a live numeric readout of [period / total
energy / conserved sum / key metric], updating
as the simulation runs.
MUST Do: Make a known/limiting case
reachable from the controls.
Don't: Bake in settings where you can't
recover a case with a known answer.
Make sure I can set [the parameters] to the case
where the answer is known to be [value], so I
can check it.
SHOU
LD
Do: Surface a conserved quantity or
fixed total so drift is visible.
Don't: Assume conservation holds
without ever displaying it.
Compute and plot [the conserved quantity / fixed
total] over time so I can see whether it stays
constant.
V1

Scientific Inquiry with AI • Creating AI-Generated Simulations • Simulation Improvement Playbook Page 3
2. Controls & Interactivity
Tier Do / Don't Prompt to get there
MUST Do: Expose the real parameters as
sliders or inputs, with units on every
label.
Don't: Hard-code magic numbers a user
can't change.
Expose [parameters] as labeled sliders with
units and sensible ranges. Let me also type
exact values.
MUST Do: Add a reset button that returns to a
known starting state.
Don't: Force a page reload to start over.
Add a Reset button that restores the initial state
and clears any plots.
SHOU
LD
Do: Let inputs reach the extremes where
the model should fail.
Don't: Clamp ranges to only the
well-behaved regime.
Extend the slider ranges so I can push
[parameter] to extreme values where the model
may break.
NICE Do: Add play / pause / step and a speed
control.
Don't: Lock the animation to one speed
you can't follow.
Add play, pause, single-step, and a speed
control for the animation.
3. Visual Clarity & Readout
Tier Do / Don't Prompt to get there
MUST Do: Label axes, units, and what each
visual element represents.
Don't: Leave an unlabeled curve or blob
the viewer has to guess at.
Label all axes with quantities and units, and add
a short legend for every plotted element.
SHOU
LD
Do: Plot the output over time or over the
swept parameter, not just an animation.
Don't: Rely on animation alone, which is
hard to measure.
Add a graph of [output] versus [time / swept
parameter] alongside the animation.
NICE Do: Mark the predicted value or known
case on the plot.
Don't: Make the user compare against a
number in their head.
Draw a reference line on the plot at the
predicted value [value] so I can compare at a
glance.
V1

Scientific Inquiry with AI • Creating AI-Generated Simulations • Simulation Improvement Playbook Page 4
4. Reusability Across Topics
Tier Do / Don't Prompt to get there
MUST Do: Keep it one self-contained HTML file
you download, not just an in-app
Artifact/Canvas preview.
Don't: Leave it as a live view inside the
AI app, or split it across files or require a
build step or API key.
Keep everything in a single self-contained HTML
file with inline JS and CSS, no external libraries,
runs offline. Give me the full file to download.
SHOU
LD
Do: Print the assumptions and the
domain of validity on screen.
Don't: Leave the model's limits implicit so
future-you forgets them.
Add an on-screen note listing the key
assumptions and the range where this model is
valid.
SHOU
LD
Do: Add a one-line header: what this
models and where it breaks.
Don't: Hand yourself an unlabeled black
box you won't understand next month.
Add a one-line title and subtitle stating what this
simulates and where it stops being accurate.
NICE Do: Provide preset scenario buttons for
common cases.
Don't: Make every demo a manual slider
hunt.
Add preset buttons that load [a few named
scenarios] with one click.
Why “downloaded HTML file,” not “Artifact / Canvas / Gem”? An Artifact (Claude) or
Canvas (ChatGPT, Gemini) is the live preview beside the chat, handy for building, but it lives
in that AI account, may need a login, and can change or disappear. (Gemini Gems and
custom GPTs are saved chatbot personalities, not a simulation file.) A single downloaded
.html file runs offline anywhere, looks the same for everyone, and is yours to keep, re-open,
and submit. Build in the live view if you like, then have the AI output the whole thing as one
self-contained HTML file and download it.
Companion to this lesson: Damped_Pendulum_Simulation.html is the reference simulation for
“Building to Understand: Creating AI-Generated Simulations as a Mode of Scientific Inquiry,”
and Pendulum_Simulation_Prompt_and_Build_Log.docx shows every Do below applied against
a real build, with its outcome recorded.
V1

Scientific Inquiry with AI • Creating AI-Generated Simulations • Simulation Improvement Playbook Page 5
5. Robustness & Reliability
Tier Do / Don't Prompt to get there
MUST Do: Handle bad or extreme input
gracefully (no crashes, no blank screen).
Don't: Let a divide-by-zero or NaN
silently freeze the simulation.
Guard against invalid or extreme inputs: prevent
NaN/Infinity, and show a clear message instead
of crashing.
SHOU
LD
Do: Use a stable, small-enough time step
(or an adaptive method) so results don't
depend on it.
Don't: Pick a step size that makes the
result drift or blow up.
Use a numerically stable integration step (or an
adaptive method); confirm the result doesn't
change if I halve the step.
NICE Do: Show the current numeric error or
step size when relevant.
Don't: Hide numerical settings that affect
accuracy.
Display the integration step size and let me
change it, so I can see its effect on accuracy.
6. Pedagogical Usefulness (Learning)
Tier Do / Don't Prompt to get there
SHOU
LD
Do: Let the user predict, then reveal, e.g.
a ‘show answer’ or target-matching
mode.
Don't: Give away the result before the
user commits to a guess.
Add a mode where I set a prediction first, then
reveal the actual result and the difference.
SHOU
LD
Do: Highlight the non-obvious behavior
(the threshold, resonance, saturation,
tipping point) when it appears.
Don't: Bury the interesting regime among
ordinary ones.
Visually flag when [the interesting behavior]
occurs as I sweep [parameter].
NICE Do: Let the user export their data or a
snapshot for a report.
Don't: Trap the results inside the page.
Add a button to export the current data as CSV
and to save a snapshot image of the plot.
V1

Scientific Inquiry with AI • Creating AI-Generated Simulations • Simulation Improvement Playbook Page 6
Remember the order: validate first, improve second. None of this polish matters if the
underlying model is wrong, make it correct, then make it useful.
One-Glance Checklist
Before you call a simulation done, confirm:
● MUST — full model (no hidden approximations) · live readout of your check quantity · a
reachable known/limiting case · labeled controls with units · reset · graceful on bad input ·
one self-contained HTML file you can download.
● SHOULD — a conserved quantity or fixed total shown · a plot vs. time or swept parameter ·
inputs reach the failure regime · assumptions & validity on screen · stable time step ·
predict-then-reveal.
● NICE — play/pause/step · reference line at the predicted value · preset scenarios ·
data/image export.
V1