Comprehensive Physics Prompts

Getting Real Results from AI in Physics Workflows

I spent about three weeks last fall trying to use raw AI models to generate physics problem sets for my senior-level classical mechanics course. The results were awful. Generic textbook problems with hand-wavy explanations and occasional dimensional analysis errors that slipped through. I figured the issue wasn't the model itself but the prompt structure, so I started mapping out what actually worked versus what produced garbage output. That mapping became what people are now calling Comprehensive Physics Prompts. They are structured prompt templates designed specifically for physics content generation, problem formulation, derivation requests, and conceptual explanation. Unlike a general-purpose prompt that says "explain Newton's laws," a physics-specific prompt tells the model what level you need, what mathematical formalism to use, whether to include free-body diagrams or stick to algebraic derivation, and how to handle edge cases like non-inertial reference frames. The core idea is that physics AI output is extremely sensitive to how you frame the problem. A vague prompt produces a vague and often wrong answer. A tightly scoped prompt with explicit constraints produces something you can actually use in a lecture or research note.

How to Build Effective Physics Prompts

Start by identifying the exact output you need. Are you looking for a worked derivation with every step shown, a multiple-choice question with distractor explanations, a computational simulation script in Python or MATLAB, or a conceptual explanation aimed at a specific audience level? Each of these requires a different prompt architecture. Level specification is the single most important variable. A prompt that asks for a derivation of Lagrangian mechanics without specifying the audience will default to something between undergraduate junior level and graduate level depending on the model. I learned this the hard way when a model derived the Euler-Lagrange equation assuming familiarity with the calculus of variations but then skipped the boundary condition argument, which confused every student in my class who had only seen basic single-variable calculus. I now always include the prerequisite knowledge level in my prompts.

Prompt Structure That Works

Here is the template I use as a baseline: State the physics domain and specific topic. Define the target audience and their known prerequisites. Specify the required output format. Include any constraints on mathematical formalism, numerical precision, or assumptions. Request that the model identify and state any approximations being made. Ask it to flag where physical intuition diverges from the mathematical result. A concrete example. I prompted for a problem involving a damped harmonic oscillator driven at resonance. My prompt specified: undergraduate physics major level, include the differential equation setup, show the particular solution using complex exponential method, give the phase relationship between driving force and displacement, and calculate the quality factor Q for damping ratios less than 0.1. I also asked the model to explicitly note where the steady-state approximation breaks down. The output was accurate through the transient regime analysis but missed the fact that at exact resonance the amplitude grows linearly with time before saturation effects kick in. That gap in the answer is exactly why I now always include a sanity-check request in my prompts.

Common Pitfalls and How to Avoid Them

Pitfall one: dimensional inconsistency. Models occasionally drop constants or mishandle unit conversions, especially when working through multi-step derivations. I found that explicitly asking the model to verify dimensions at each major step catches roughly 80 percent of these errors. The remaining 20 percent require manual inspection of the final expression. Pitfall two: false confidence in numerical outputs. When a prompt asks for a numerical result, the model may produce a plausible-looking number that is wrong because it used an incorrect formula internally. I now always ask for the formula before the numerical answer. This small request changed my error rate from about one mistake per three problems to roughly one mistake per ten problems. Pitfall three: conflating similar concepts. Models sometimes confuse related but distinct ideas, such as treating thermal conductivity and diffusivity as interchangeable, or mixing up group velocity with phase velocity in wave mechanics. I address this by adding a constraint that asks the model to distinguish between commonly confused terms when both are relevant to the problem.

A Specific Edge Case I Encountered

Last semester I generated a prompt about relativistic collision kinematics where a particle collides with a stationary target and both products are identical mass. The model gave a correct answer for the threshold energy calculation in the lab frame but presented it using the wrong Lorentz invariant. It used the sum of the rest energies squared instead of the Mandelstam variable s evaluated at threshold. The numerical answer was correct because the algebra happened to work out, but the derivation was physically wrong. Students who tried to apply that method to a non-symmetric collision would get the wrong answer. My workaround was to add an explicit requirement that the model identify which Lorentz invariant it is using and state why that invariant applies. This forces the model to commit to a specific approach and makes errors in the physical reasoning visible rather than hidden behind a correct number. I also started cross-referencing every relativistic result against the standard textbook formulas for threshold energy before sharing them with students.

When Comprehensive Physics Prompts Fail Completely

They do not handle genuinely open-ended research questions well. If you ask for an analysis of a physical system that has no standard textbook treatment, the model will hallucinate details from analogous systems. I tried using prompts to generate discussion material about quantum entanglement in topological superconductors and got back a mix of standard BCS theory and incorrect claims about Majorana fermion localization lengths that sounded convincing but were wrong. The model has no access to current arXiv papers and cannot verify claims against primary literature. Another failure mode is highly visual or geometric reasoning. Prompts asking for detailed spatial descriptions of field configurations, such as the electric field topology near a charged ellipsoid, tend to produce qualitatively correct statements but quantitatively wrong results. I stopped relying on the model for anything requiring precise spatial visualization and switched to generating those descriptions manually from analytical solutions. For situations like these, I recommend combining AI-generated prompts with a verification step using established references. Gradshteyn and Ryzhik for integrals, Jackson for electrodynamics problems, Landau and Lifshitz for advanced mechanics. A prompt can generate a first draft quickly, but manual verification against authoritative sources is still necessary for anything you plan to publish or use in graded coursework.

Practical Workflow Recommendation

Use a comprehensive physics prompt to generate a first-pass draft. Run a dimensional analysis check on every equation. Verify boundary conditions match the problem statement. Check limiting cases where the physics should reduce to something you already know. If the prompt is for educational material, have a peer or colleague spot-check at least one problem for accuracy before distributing it. This pipeline typically takes about 20 minutes per problem set compared to the 2 to 3 hours I was spending writing problems from scratch before I started using structured prompts. The real value of Comprehensive Physics Prompts is not that they replace your understanding of physics. They replace the tedious scaffolding work that surrounds it. The prompt gets you from blank page to rough draft fast. You still need to verify, refine, and contextualize the output yourself. That part cannot be automated away yet.