Unit 10: Capstone - Build a Useful AI Tool

Lesson at a glance

| Item | Detail | | --------------------- | ------------------------------------------------------------------------------------ | | Suggested length | 4 × 60 minutes (project work + presentation) | | Recommended placement | Week 10 of AI Fluency | | Prerequisite | Units 1–9 complete | | Materials | Whatever each student needs for their chosen build path; presentation slot for demos |

Safety: Capstones must follow the AI Use Agreement. Reject project ideas that involve identifiable peers' data without consent, voice/face cloning of real people, or anything that violates the school's policies. Approve every capstone topic before students start building.

Standards & credential alignment

• AI4K12 Big Ideas: All five - students integrate.
• CSTA K-12: 3A-AP-13, 3A-AP-22, 3A-IC-25, 3A-IC-28.
• ISTE Students: 1.4 Innovative Designer, 1.5 Computational Thinker, 1.6 Creative Communicator.
• NIST AI RMF: All four functions in miniature.

Learning objectives

By the end of the capstone, students can:

1. Scope an AI project to something they can actually finish in 4 days.
2. Pick the right tool tier (frontier, local, RAG, agent, multimodal) for their problem and defend the choice.
3. Build, iterate, and ship a working artifact (config + prompts + demo).
4. Disclose, verify, and contribute throughout the build, with documentation.
5. Present their tool to peers in a 5-minute demo with Q&A.
6. Reflect critically on what worked, what didn't, and what they'd do next with more time.

The four build paths

Pick one. Each is scoped to be finishable in four class periods plus modest homework.

Path A - Local Study Buddy (Unit 6 + Unit 4)

Build a custom local LLM (Ollama Modelfile or LM Studio config) tuned as a tutor for one of your other classes.

Deliverables:

• The Modelfile / config (with system prompt).
• 5 prompts you tested and the model's responses.
• A 1-page write-up of your iteration history (prompt v1 → v2 → v3 + why).
• 3-minute live demo: student plays the tutee; class plays evaluators.

Path B - RAG over Your Notes (Unit 7)

Use a NotebookLM-style tool (Google's NotebookLM, ChatGPT's Custom GPT with file uploads, Claude Projects, or an open-source RAG stack) to build a study companion that answers from your class notes, not from training data.

Deliverables:

• The corpus you uploaded (10–30 pages of your own notes/textbook excerpts you have rights to).
• 10 questions you asked and the cited answers.
• A side-by-side comparison: same question with no RAG vs. with RAG. Note hallucinations the bare model produced that RAG fixed.
• 3-minute demo + Q&A.

Path C - Prompt Library Product (Units 3 + 4)

Curate a polished, themed prompt library that solves one repeated problem in your life. Examples: college-essay coach, ACT/SAT drill bot, debate-prep helper, code-review buddy.

Deliverables:

• ≥ 8 polished prompts in a structured doc (purpose, prompt text, expected output, example output, use notes).
• A "how to use this library" intro page for a hypothetical user.
• Demo: peer uses your library on a real problem in real time.

Path D - Multimodal Workflow (Unit 8)

Design a workflow that uses image, voice, or video AI for a creative project (with consent and full attribution).

Examples:

• Storyboard a short film with text-to-image; record narration; assemble.
• Generate album art for a fictional album; describe the production process.
• Build an "image-to-recipe" or "image-to-summary" workflow with a vision-LLM.

Deliverables:

• The artifact (storyboard, video, image set, etc.).
• Full prompt + tool log.
• Disclosure note covering what AI did and what you did.
• 3-minute demo.

Pacing - Day 1 (60 min): Pitch and approve

| Time | Segment | Notes | | ----------- | ---------------------------------------- | ----------------------------------------- | | 0:00 – 0:15 | Mini-lesson - what makes a good capstone | Scoped, useful, demoable, ethical. | | 0:15 – 0:35 | Activity - pitch sheet | Each student fills out a 1-page pitch. | | 0:35 – 0:55 | Conferences | Teacher approves or redirects each pitch. | | 0:55 – 1:00 | Set milestones | |

Day 1 - Pitch sheet template

Each student fills out:

• Path: A / B / C / D
• Problem in one sentence: "I want a tool that …"
• Why this is the right path: (1–2 sentences)
• Tools I'll use: (specific models / apps)
• What "done" looks like: the actual artifact, in concrete terms
• Risks / what could go wrong: be specific
• Disclosure plan: what AI did, what I did, how I'll verify

The teacher signs the sheet to approve. Students who can't write a clear "done" definition aren't ready to start - get them there first.

Pacing - Days 2 and 3 (60 min each): Build

| Time | Segment | Notes | | ----------- | ---------- | ---------------------------------------------------------------- | | 0:00 – 0:05 | Stand-up | Each student names yesterday's progress, today's plan, blockers. | | 0:05 – 0:55 | Build time | Teacher circulates; quick 1:1 unblocks. | | 0:55 – 1:00 | Wrap-up | Update milestones; flag at-risk students. |

Day 2 ends with a working rough cut of the artifact. Day 3 ends with a polished cut and rehearsed demo. Anything still incomplete after Day 3 stand-up gets scoped down - students ship a smaller working thing rather than a bigger broken thing.

Pacing - Day 4 (60 min): Demo day

| Time | Segment | Notes | | ----------- | ------------------------------ | ------------------------------------------------------------------- | | 0:00 – 0:50 | Demos (3 min each + 1 min Q&A) | Time-box hard. Hat with names if order is needed. | | 0:50 – 0:55 | Audience choice award | Vote. | | 0:55 – 1:00 | Course close | Teacher's closing remarks. Hand-shake, real ink on the certificate. |

The closing remarks (5 min)

A version of this - make it your own:

"Ten weeks ago, you didn't know what a token was. Today you've built a working AI tool. You learned how LLMs work, you learned how to drive them, you learned how to install one on your own machine and run it with the wifi off, you learned how to fight prompt injection, you learned how to use this stuff without lying to anyone - including yourself.

The companies that hire people who can do this are paying $80K to start. Some of you will work for them. Some of you will start the next one. Either way, you signed an agreement on the first day to use this tool like an adult, and from what I've seen, you have. That's the credential."

Capstone rubric (40 points total)

| Dimension | Points | Look-fors | | ------------------------- | ------ | ----------------------------------------------------------------------------------------------- | | Problem clarity | 5 | The problem is real, specific, and finishable. | | Tool fit | 5 | The right tier (frontier vs. local vs. RAG vs. agent vs. multimodal) for the problem, defended. | | Build quality | 10 | The artifact works. The demo runs without rescue. | | Iteration evidence | 5 | Prompt or config v1 → v2 → v3 with rationale. | | Disclosure & verification | 5 | Clear, honest AI Use Note. Facts verified. | | Ethics | 5 | Consent, privacy, copyright handled correctly. | | Demo & Q&A | 5 | 3-minute demo runs clean; can answer follow-up questions. |

Pass bar: 30/40 to receive course completion. Below 30 → re-cut and re-demo within one week.

What's next, after the course

The point of this course is not "I learned about AI." The point is I can use AI well, professionally, and ethically, without being fooled by it. Every student leaves with:

• A signed AI Use Agreement on file.
• A personal prompt library (10+ entries).
• A custom local-LLM persona on a USB stick or cloud config.
• A capstone artifact they can show in a college essay or job interview.
• The vocabulary and reflexes to learn the next thing as the field evolves.

The field will change. The disclose / verify / contribute test won't.