Most “AI prompt” lists are vague templates that produce vague outputs. This one is different: every prompt below is specific enough to copy, paste, and run right now — for chemistry, biology, physics, and Earth science. Includes what makes each prompt work, and what fails without that structure.
- Vague prompts (“help me with my science experiment”) return vague output. Specific constraints produce specific, usable science guidance.
- The most effective prompts name the specific variables, ask for prediction with reasoning, and request a follow-up question the student should investigate next.
- AI can’t run the experiment for you — but it’s genuinely good at predicting outcomes, explaining mechanisms, flagging safety concerns, and generating testable hypotheses.
- Safety warning: never use AI predictions as a substitute for proper lab safety procedures. Treat AI output as a starting hypothesis, not a verified result.
The honest reason AI works well for science experiments is unglamorous: it’s fast at generating plausible hypotheses and explaining mechanisms, and students are often stuck at exactly those two points. You know you want to test something involving pH and plant growth, but you don’t know which variables to isolate or why the mechanism would produce the effect you expect. AI bridges that gap in under a minute.
What AI can’t do: run the experiment. Verify its own outputs. Guarantee safety. Replace the observation that comes from actually mixing the compounds and watching what happens. The prompts below are starting points, not substitutes.
Ask any LLM “help me design a science experiment” and you’ll get back a generic scientific method outline you could have found in a fifth-grade textbook. The output is technically correct and completely useless because it doesn’t know what you’re trying to investigate, what materials you have, or what grade level you’re working at.
The fix is specificity. Every prompt that works has the same structure underneath it: subject + specific variable + constraint + reasoning requirement. The constraint is what most guides skip — it’s the instruction that forces the AI to stay within the bounds of what’s actually testable rather than describing experiments that require lab equipment you don’t have.
Chemistry prompts benefit most from specificity about concentration, temperature, and available materials. “What happens when I mix X and Y” is the most common chemistry question — and also the weakest version of it. The prompts below push harder.
Use this when you want to predict what happens in a chemical reaction before running it — especially useful for checking safety concerns and understanding the mechanism first.
I'm running a chemistry experiment with [REACTANT 1] and [REACTANT 2] under the following conditions: temperature [X°C], concentration [X mol/L], closed or open container. Predict: (1) the products of this reaction and the balanced equation, (2) whether this reaction is endothermic or exothermic and why, (3) any safety hazards I should know before starting, and (4) one variable I could change to test a different outcome. I have access to: [list available materials, e.g., standard school lab glassware, no fume hood].
Why it works: The materials constraint prevents the AI from suggesting experiments requiring equipment you don’t have. Asking for the balanced equation forces mechanistic reasoning rather than vague description. The “one variable to change” request generates your next experiment automatically.
Acid-base chemistry is one of the most common school experiments. This prompt generates a full experimental design, not just a description of what pH is.
I want to test how [specific factor: e.g., dilution, temperature, or added salt] affects the pH of [specific solution, e.g., lemon juice / household bleach / bicarbonate solution]. Design a step-by-step experiment I can run with litmus paper or a pH meter. Include: (1) independent variable, dependent variable, and two control variables I must keep constant, (2) expected results and the chemical reason for them, (3) a data table I should fill in during testing, and (4) what a surprising result would look like and what it might mean.
Why it works: Requesting the explicit variable identification forces the AI to structure the experiment correctly rather than describing pH in general terms. The “surprising result” question is the one most guides skip — it trains students to think about falsification before they start.
Paper chromatography is a classic low-equipment experiment. This prompt pushes beyond “what will I see” into “why will I see it.”
I'm doing a paper chromatography experiment using [solvent: water / rubbing alcohol / saltwater] and [pigment source: black felt-tip marker / spinach leaves / food coloring]. Before I run it, predict: (1) how many pigment bands I'm likely to see and what colors, (2) which pigment will travel farthest up the paper and why, in terms of polarity and molecular interaction, (3) how my results would change if I switched from water to rubbing alcohol as the solvent, and (4) one question about the results that I should try to answer by modifying the experiment after my first run.
Why it works: Asking for a mechanism explanation (polarity, molecular interaction) prevents the AI from giving a purely descriptive answer. The solvent-switch question teaches the key concept — that solvent polarity controls separation — without requiring you to run a second experiment first.
AI output is a hypothesis, not a verified safety clearance. For any experiment involving heat, concentrated acids or bases, or unfamiliar compound mixtures — consult your school’s safety guidelines or a qualified instructor before proceeding. AI language models are not chemistry safety databases and can produce confident-sounding incorrect predictions about reaction hazards.
Biology experiments often have longer timescales and messier variables than chemistry experiments. The prompts below are designed around the constraint that real biology happens over days or weeks, not one lab session — and that living systems don’t follow instructions as neatly as chemical reactions do.
One of the most accessible biology experiments. The problem is that students usually design them badly — testing too many variables at once, or not accounting for natural variation between individual plants.
I want to test how [specific variable: light color / soil pH / watering frequency / fertilizer type] affects the growth of [specific plant species, e.g., radish seedlings / bean sprouts / lettuce]. My experiment will run for [X days] and I have [X number of plants] to work with. Help me design it properly: (1) how many plants per group do I need for the result to be meaningful, and why, (2) what exactly should I measure (height, leaf count, root length, or something else) and how often, (3) what natural variation between plants might make my results misleading, and how to account for it, and (4) what result would tell me my hypothesis was wrong.
Why it works: The replication question (how many plants per group) is where almost every student experiment fails — they test one plant per condition and draw conclusions from it. Asking the AI to address natural variation and falsification forces a methodologically sound design.
Catalase, amylase, or pectinase in school biology. The classic question is “how does temperature affect enzyme activity” — but students rarely know what the mechanism is that produces the result they observe.
I'm testing the effect of [temperature / pH / substrate concentration] on the activity of [enzyme: catalase from liver / amylase from saliva / pectinase]. I'll measure enzyme activity by [method: bubble production / clearing of starch with iodine / juice yield from fruit]. Before I run the experiment: (1) draw out in words what the activity vs. [variable] curve should look like, and explain what happens at the molecular level at each key point on that curve, (2) predict which temperature / pH will produce peak activity and why, (3) what would denaturing look like in my measurements, and (4) one common mistake students make when running this experiment that would cause incorrect results.
Why it works: Asking for a verbal curve description forces the AI to articulate the bell-curve shape of enzyme activity — and explain the mechanism at each stage. The common-mistake question is genuinely useful; AI is good at cataloguing methodological errors because they appear frequently in educational literature.
The classic potato-in-salt-water experiment, made rigorous.
I'm doing an osmosis experiment using potato cylinders in salt solutions of different concentrations: [0%, 0.5%, 1%, 2%, 5% NaCl]. Before I run it, predict: (1) at which concentration the potato mass will increase, stay the same, and decrease — and the specific biological reason for each, (2) what the approximate isotonic point (no mass change) is for potato cells and why, (3) what my mass-change vs. concentration graph should look like, and (4) why two potato cylinders cut from the same potato might give slightly different results even if the procedure is identical.
Why it works: Asking for the isotonic point specifically forces the AI to reason about cell solute concentration, not just describe the concept of osmosis. The “same potato, different results” question teaches biological variability — one of the hardest concepts for students to accept.
Agar plate experiments testing which household substances inhibit bacterial growth. Common experiment, widely misunderstood.
I'm testing the antimicrobial effect of [substances: honey / garlic extract / tea tree oil / hand sanitizer] against [bacterial source: bread mold spores / E. coli culture / bacteria from unwashed hands] on agar plates. Help me design the experiment and interpret it: (1) what a clear zone of inhibition means, and what its absence means — including the possibility that the substance works but not by killing bacteria, (2) how to measure the inhibition zone in a way that gives comparable numbers across substances, (3) one confounding variable that could make one substance look more effective than it actually is, and (4) why my results might differ from published studies using the same substances.
Why it works: The “absence of zone” question is critical — students often conclude that no zone means no effect, when the substance might be bacteriostatic rather than bactericidal. The confounding variable question generates real critical thinking.
Physics experiments have the advantage of quantitative predictions — you can usually calculate an expected result before running the experiment and compare it to what you observe. The prompts below lean into this: every one asks for a predicted number, not just a predicted direction.
The classic pendulum experiment — with the math that most student guides skip over.
I'm investigating what affects the period of a simple pendulum. I'll test three variables: string length (25cm, 50cm, 75cm, 100cm), bob mass (50g, 100g, 200g), and starting angle (10°, 20°, 30°, 45°). Before I run it: (1) calculate the predicted period for each string length using the formula T = 2π√(L/g), showing the calculation, (2) predict which of the three variables will have no measurable effect on the period, and explain the physics reason, (3) at what angle will my results start to deviate significantly from the simple pendulum formula, and why, and (4) what is a realistic source of error in my timing method if I'm using a stopwatch?
Why it works: Asking for the calculation output first gives you a benchmark to compare your experimental results against — which is the whole point of physics. The “no measurable effect” question tests whether the student understands what the formula actually says. Most students are surprised that mass doesn’t matter.
Rolling a ball off a table and predicting where it lands. Simple setup, real kinematics.
I'm designing a projectile motion experiment using a ramp and a ball that rolls off the edge of a table [height: X meters]. I'll vary the ramp height to change the launch speed. Before running it: (1) walk me through how to calculate the predicted landing distance for a given launch speed — include the formula and a worked example with [specific ramp height], (2) predict how the landing distance changes if I double the table height (keeping launch speed the same), (3) what air resistance actually does to my results at the ball speeds I'll be using, and whether I need to account for it, and (4) how I would measure the launch speed of the ball without specialized equipment.
Why it works: The air resistance question matters because every textbook ignores it. At the speeds a ball rolls off a school lab table, air resistance is genuinely negligible — but students don’t know that unless they’re told explicitly. Getting the AI to reason through it builds real physical intuition.
Testing Ohm’s Law and how resistors in series vs. parallel behave differently.
I have a circuit with a battery ([voltage: X V]) and resistors of [X Ω, Y Ω, Z Ω]. First, calculate: (1) the total resistance and current if these are wired in series, (2) the total resistance and current if they're in parallel — show the formula for parallel resistance and the calculation, (3) in which configuration will the battery drain faster and why, and (4) if one resistor in a parallel circuit fails (goes to infinite resistance), what happens to the current through the other resistors? What if it fails short-circuit (goes to zero resistance)?
Why it works: The failure mode questions at the end are what separates this from a textbook exercise. Students who understand how series and parallel circuits fail under fault conditions actually understand the concepts — not just the formula.
Tuning forks, standing waves, resonance. Harder to design well than it looks.
I'm investigating how the length of an air column affects the resonant frequency of sound, using a tube closed at one end and a series of tuning forks (frequencies: [list available Hz values]). Help me design and predict: (1) for a tube of length [X cm], calculate the fundamental resonant frequency — include the formula and show the relationship between tube length and wavelength, (2) what I should hear (and what I should feel) when the tube length matches the resonant condition for a given tuning fork, (3) how my results would differ for an open-ended tube versus a closed-ended tube of the same length, and (4) what the second and third harmonics would be for my fundamental frequency.
Earth science experiments often involve real-world data rather than controlled lab conditions. The prompts below are designed for that context: using publicly available datasets, modeling climate systems, and interpreting geological evidence. They work well for both classroom lab settings and independent projects.
Testing how slope angle, vegetation cover, and soil type affect erosion rate. Can be done with a simple ramp, soil, and water.
I'm designing an experiment to test how [slope angle / vegetation cover / soil type] affects the rate of soil erosion, using a plastic tray ramp, potting soil, and a measured amount of water poured at a constant rate. Before I start: (1) which of these three variables has the largest real-world effect on erosion, based on current soil science, and what's the evidence, (2) how should I measure "erosion rate" — what exactly should I collect and weigh or measure, (3) what's a realistic range of slope angles for my experiment that would show a meaningful difference in results, and (4) what does a control condition look like for this experiment?
Why it works: The “which variable matters most in real life” question anchors the experiment in actual environmental science rather than isolated lab curiosity. Students who know slope angle is generally more significant than soil type in erosion research are doing science, not just running a procedure.
Simulating ocean acidification with vinegar solutions and seashells or eggshells. One of the most direct ways to observe a real environmental process in a classroom.
I'm simulating ocean acidification by placing [seashells / chalk / eggshells] in solutions of different acidity: [water only / 1% vinegar / 5% vinegar / 10% vinegar]. Before I run it: (1) what is the actual pH of each solution, and how does it compare to the current ocean pH (~8.1) and predicted future ocean pH (~7.8), (2) what chemical reaction causes the shell material to dissolve, and write the equation, (3) what I should observe over 24 hours versus 72 hours — will the rate change over time, and why, and (4) what organisms in real oceans are most threatened by the process I'm simulating, and why those species specifically.
Why it works: Connecting the vinegar pH to actual ocean pH values gives the experiment real-world calibration. Asking about rate change over time (yes — it slows as the solution saturates with dissolved shell material) produces a more sophisticated observation task than just “measure what dissolves.”
Testing tap water, bottled water, or local water sources for pH, hardness, nitrates, or dissolved oxygen using test kits.
I'm testing [tap water / river water / rainwater / bottled water] for [pH / water hardness / nitrate levels / dissolved oxygen] using a [test kit / pH meter / chemical test strips]. Before I collect samples: (1) what range of values is considered safe or normal for each parameter I'm testing, and what standard sets those limits, (2) what environmental or infrastructure factors in my area could cause my tap water to differ from the recommended range, (3) how should I collect and store my water samples to avoid contaminating the results before testing, and (4) if my results show elevated nitrates, what are the three most likely sources in my local area?
Using publicly available weather data (NOAA, Weather Underground, local station records) to investigate a local climate trend. This is a data-analysis project rather than a physical experiment, which makes AI especially useful.
I want to analyze [30 years of temperature data / annual precipitation records / storm frequency data] for [my city/region] from [NOAA / Weather Underground / NASA GISS] to investigate whether [specific trend, e.g., summer temperatures have increased / frost-free season has lengthened / extreme rainfall events have become more frequent]. Help me design the analysis: (1) what specific metric should I calculate to test my hypothesis — not just average temperature, but something more specific and meaningful, (2) what statistical method is appropriate for detecting a trend in climate data over 30 years, at my level (I'm a high school student / undergraduate), (3) what natural climate cycles (El Niño, PDO, AMO) could produce a trend in my data that isn't a long-term climate signal, and how I could check for them, and (4) where to find and download the specific dataset I need, with the exact search terms or URL.
Why it works: Climate data projects are hard to design well because students often compare two data points and call it a trend. Asking for a specific metric and a statistical method appropriate to the grade level produces an analysis framework that’s actually defensible. The natural cycles question is genuinely advanced — and genuinely important.
Quick Reference: Which Prompt for Which Situation
| # | Prompt | Domain | Best for |
|---|---|---|---|
| #01 | Reaction prediction | Chem | Pre-lab safety + mechanism review |
| #02 | pH experiment design | Chem | Variable identification + data table |
| #03 | Chromatography | Chem | Understanding polarity + solvent effects |
| #04 | Plant growth | Bio | Multi-week experiment design + replication |
| #05 | Enzyme activity | Bio | Temperature/pH effect + denaturation |
| #06 | Osmosis | Bio | Quantitative prediction + biological variation |
| #07 | Microbial growth | Bio | Antibiotic/antimicrobial fair projects |
| #08 | Pendulum | Physics | Variables that don’t matter + formula verification |
| #09 | Projectile motion | Physics | Kinematics calculation + air resistance judgment |
| #10 | Circuit resistance | Physics | Ohm’s Law + series vs. parallel |
| #11 | Sound waves | Physics | Resonance + harmonics |
| #12 | Soil erosion | Earth | Environmental variables + control design |
| #13 | Ocean acidification | Earth | Real-world calibration + chemical equation |
| #14 | Water quality | Earth | Local environmental investigation |
| #15 | Weather data analysis | Earth | Climate trend projects, advanced level |
All prompts tested on GPT-4o and Claude 3.5 Sonnet, April 2026. Results will vary by subject and grade level — treat all AI output as a starting hypothesis, not a final answer.
What Makes These Prompts Work — and What Doesn’t
There’s a pattern across every effective science experiment prompt, and it’s worth naming directly: they all force reasoning, not retrieval. The weak version of “what happens when I mix baking soda and vinegar” is a retrieval question — the AI looks up the answer and tells you. The strong version asks for the balanced equation, the thermodynamics, the safety hazards, and the next variable to test. That requires the AI to reason through the mechanism, not just recall a fact.
The second pattern: every good prompt includes a constraint. Materials available, time available, grade level, equipment available. Without constraints, AI will design experiments requiring equipment you don’t have. With them, it designs within your actual situation.
At the end of any prompt on this list, add: “Flag any part of your answer where you’re uncertain or where I should verify with a primary source.” This prompt addition reliably reduces confident-sounding wrong answers. AI language models don’t naturally flag their own uncertainty — you have to ask for it. More prompt frameworks: bestprompt.art.
What AI Can’t Do — and Shouldn’t Be Asked To
Direct observation. The AI has never seen a flame test. It doesn’t know what copper chloride actually looks like burning — emerald green, brilliant, a little startling if you’ve never seen it. No prompt will give you that. The experiment is still your job.
Verified safety clearance. AI language models can identify known hazards with reasonable reliability, but they also hallucinate, they don’t know your specific lab setup, and they’re not chemical safety databases. Use them for initial hazard screening only — then verify with your instructor or your school’s chemical safety guidelines before running anything unfamiliar.
Statistical validity. AI can suggest what to measure and how often. It cannot tell you whether your specific dataset, collected from your specific experiment, is statistically significant. That requires running the statistics yourself, on your actual numbers.
The experiment is still yours. The AI made the design sharper. What happens in the lab when you actually run it — that’s the part no prompt can substitute for.
References and further reading
- National Science Teaching Association. (2022). Science and Engineering Practices in the NGSS. ngss.nsta.org
- Leviathan, Y., Kalman, M., & Matias, Y. (2025). Prompt repetition improves non-reasoning LLMs. arXiv:2512.14982. arxiv.org/abs/2512.14982
- NOAA National Centers for Environmental Information. Climate data online. ncdc.noaa.gov/cdo-web
- Atticus Project. CUAD Dataset (for prompt constraint methodology). atticusprojectai.org/cuad
- EPA. (2023). Secondary Drinking Water Standards: Guidance for Nuisance Chemicals. epa.gov/sdwa
- Wharton GAIL. (2025, December). Playing Pretend: Expert Personas Don’t Improve Factual Accuracy. gail.wharton.upenn.edu (on why task-specific instructions outperform persona prompts for factual content)
- BestPrompt.art. AI prompt frameworks and templates. bestprompt.art
