Appendix B — Prompt Engineering (Student Guide)

Prompt engineering is an art and you can get better at it by relying on simple patterns for getting useful help from LLMs. The goal is to get useful, verifiable help—without outsourcing your learning.

Foundations

  1. Be explicit. State the goal, audience, tone, and output contract (bullets? code? a table with specific columns?).
  2. Structure the task. Give headings or a checklist. The format matters more than fancy wording.
  3. Ask for a brief rationale. Request a 3–5 bullet “why this approach” summary (not a full hidden transcript).
  4. Plan first, then act. Ask the model to outline steps / assumptions before producing the answer.
  5. Avoid conflicting instructions. If you include exceptions, say exactly which rule wins.
  6. Iterate with a rubric. Tell the model how you’ll grade the result (e.g., Accuracy, Reproducibility, Clarity, Citations) and ask it to self-check against that rubric before finalizing.
  7. Constrain scope. Set limits: dataset size, time, number of alternatives, or libraries allowed.
  8. Provide context. Paste the paragraph, code, or figure description you’re working on; don’t ask in a vacuum.
  9. Prefer “show, then tell.” Ask for the result and a short verification checklist you can run.

Agentic “toggles” (when needed)

  • Reasoning effort: If the model offers options (fast/medium/high), pick higher effort for complex tasks; otherwise add “Take the time to check edge cases and failure modes.”
  • Parallelization: For independent subtasks (e.g., summarize 3 papers), allow parallel work; forbid it when steps depend on each other.

Prompt scaffolds you can copy

A. Reading → retrieval prompts

You are helping me study. From the excerpt below, write 5 Q→A pairs that test:
- definitions,
- claims + evidence,
- assumptions.
Each Q should be specific and answerable from the text. Keep answers ≤ 2 sentences.

[PASTE EXCERPT]

B. Code review for reproducibility

Role: fussy lab TA.
Task: review my R script for reproducibility problems.
Output: a checklist with concrete fixes. Include file paths, seeds, package versions.

[PASTE CODE]

C. Outline critique (scientific report)

Audience: soil science peers.
Goal: critique my outline for clarity & evidence.
1) List missing background or definitions.
2) Flag claims lacking data/figures.
3) Suggest stronger topic sentences.

[PASTE OUTLINE]

D. Geospatial analysis guardrails

You are a geospatial QA buddy.
Given my analysis plan, identify:
- CRS pitfalls,
- raster/vector mismatches,
- resampling/aggregation gotchas,
- units and NA handling.
Return a stepwise preflight checklist.

[PASTE PLAN]

Verification beats vibes

  • Numerical outputs: replicate in R; show code & seeds.
  • Citations: follow links and read them; don’t accept un-verified references.
  • Claims: compare against course texts or primary sources.

Document your use

Every assignment submission should include an AI Use Log (see Writing & AI). Prompts & models matter: different instructions can change outcomes.

References

  • AI Daily Brief: “11 GPT-5 prompting techniques…” (transcript provided in class).