AI projects · Use · weekend

Prompt wording vs. answer quality experiment

These projects are not aligned to any state standards or curriculum. They are generally about learning, understanding, using, and evaluating AI.

Ages 11-14

Overview

Small wording changes can change AI answers a lot. You will run a controlled prompt experiment: same task, three prompt styles, a fixed grading rubric, and a comparison that separates "sounded nice" from "met the rubric."

Age band

Ages 11–14 (middle school).

Time needed

Weekend (about 3–5 hours).

AI goal

Use — practice prompting as an experiment, not as magic words.

Materials

Free tools only; no paid accounts required.

  • A free browser chat assistant.
  • A rubric sheet with 4 yes/no checks for the task you choose.
  • A results table: prompt version × trial × rubric score (0–4).

Pick a concrete task with a right-enough answer, such as:

  • Explain how a lever works for a 5th grader, including a labeled example.
  • Give three healthy after-school snacks with one reason each and no brand ads.
  • Summarize the water cycle in ≤80 words including evaporation and condensation.

Steps

  1. Write the rubric first (before prompting). Example for the lever task: (a) defines lever, (b) names fulcrum, (c) gives a real-world example, (d) stays age-appropriate. Each yes = 1 point.
  2. Design three prompt versions for the same task:
    • V1 Minimal: one sentence request
    • V2 Structured: bullet requirements matching the rubric
    • V3 Structured + check: same as V2, plus "score yourself against these four checks and fix gaps"
  3. Run three trials per version (9 total). New chat each time if the tool keeps history. Paste prompts exactly; do not hand-edit mid-answer.
  4. Score blindly if possible. Print answers with version labels hidden; score with the rubric; then reveal labels.
  5. Summarize. For each version, compute mean score = (sum of points) / 3. Note the range (min–max).
  6. Conclude from data. Which version scored higher in your nine trials? Write one reason that fits the rubric design (for example, V2 asked for each required piece). Avoid claiming the "best prompt ever."

Ethics and safety note

Do not use prompts that ask for harmful instructions. Do not paste other students' private work into the chat. If the tool refuses a prompt, record "refused" and move on.

What done looks like

CriterionNot yetDone
Rubric-firstScored by vibes after seeing answersRubric written before any prompt
DesignOne prompt onlyV1/V2/V3 defined and saved
TrialsOne shot each3 trials × 3 versions with scores
Summary"V3 is magic"Means + ranges + limited claim

Adaptations

Classroom

Class shares one rubric; groups test different tools or different tasks.

Homeschool

Read answers aloud and score together to make the rubric concrete.

Weekend

Two trials per version if time is short; keep three versions.

Research

Swap rubric order vs prompt order in a second experiment to show bias when rubrics are written after seeing answers.

Science fair

Display prompt texts and a bar chart of mean scores with error bars as min–max.

All AI projects

Prompt wording vs. answer quality experiment · Schoolhouse