A deeper guide to how large language models respond to instructions, written so you can practice deliberately—not guess—and a visual catalog of tools across chat, code, research, media, and ML (free tiers, freemium, and paid called out).
Chat apps hide the plumbing: your words become a prefix; the model predicts likely continuations. Products also inject hidden instructions and tools (browser, code). Same words in two apps → different answers—plan for that.
Everything below is a checklist. If a slot is empty, the model guesses—and burns tokens guessing.
Five slots in one prompt
1Role
Stance: tutor, skeptic, linter, translator.
2Audience
Who reads the output—you, a marker, investors.
3Task
Single outcome: outline, bugfix, rubric, diagram.
4Inputs
Paste quotes, code, data—trim noise.
5Format
Bullets, table, JSON, word cap, language.
Core vocabulary
Token
Chunks the model reads—often subwords. Price and limits are counted in tokens, not “words.”
Context window
Max prefix length at once. Long uploads + long replies compete; older stuff falls off the cliff.
System vs user
In APIs, system rules sit separate from chat turns. In consumer UIs, fake it: “Rules for this thread: …”
Grounding
Tying answers to text you provide or sources the tool retrieves—still verify citations yourself.
Three modes
Zero-shot
Instructions only—no examples. Fast for common tasks (“summarize in two sentences”).
Few-shot
Two or three input→output pairs first; teaches shape, tone, edge cases—great for JSON keys or rubrics.
Chain-of-thought
Ask for reasoning before the verdict (“think stepwise; final answer last line”). Helps math and debugging.
Build one prompt in five moves
1
Name the artifact
One sentence: slide outline, pytest file, risk table—not “help me with AI.”
2
Paste the smallest proof
Quote, traceback, dataset snippet—prove context without dumping whole PDFs.
3
Hard constraints
Language, length cap, banned moves (“do not rewrite unrelated files”).
4
Run once, skim
Check structure before facts—easier to steer tone than undo hallucinations.
5
Delta prompts
Follow-ups: “compress 30%,” “swap example to robotics,” “add a critique section.”
Copy-ready template
ROLE: You are [expertise + attitude].AUDIENCE: The reader is [level / role].TASK: Produce [single deliverable].INPUT: """ pasted material """CONSTRAINTS: [length, tone, cite-or-say-unknown, language].FORMAT: [bullets | table columns | code block language].
Symptom → lever
RAMBLEAdd max words / max bullets; demand outline-first.
MATHNotation conventions; ask for derivation policy per exam rules.
ESSAYSUse AI for outlines and counterarguments—not final prose if prohibited.
GROUP WORKSplit tasks and timelines; do not ghost-write teammates’ sections.
Integrity: Follow syllabus AI rules. Disclosure beats guessing; using AI to bypass assessment where forbidden is still misconduct.
FREE / OPEN FREEMIUM PAID OR SUBSCRIPTION
Product names are not “types of AI.” Brands such as Lovable, Emergent, or similar “AI builder” startups are applications that combine ordinary ingredients—usually large language models, templates, cloud hosting, and payments—not a separate scientific category alongside chatbots or vision models. Use this list by what you need (chat, code help, images, video, research), not by hype labels. Multi-output hubs sit last before labs: AI builders (multi).
Chat & general assistants
Front doors for questions, drafting, and multimodal experiments. Paid tiers usually unlock stronger models, higher limits, or priority access—pricing changes often.
AI search with apps and model choice; paid plans expand limits.
Coding, agents & IDEs
Where prompts meet repositories: autocomplete, chat in-editor, and terminal-aware agents. Student GitHub accounts often unlock free Copilot—verify on GitHub Education.
AI builders — websites, images & video in one place
Two buckets below: creative suites (graphics + video + light publishing) and prompt-to-app tools (describe a product, get code + hosting). Both still stack ordinary LLMs and media models—compare credits, exports, and terms before class work.
Creative suites & marketing sites
Templates plus AI for posts, clips, and simple published pages.
Template video maker with AI script/image assists—quick class-friendly clips.
Prompt-to-app & prompt-to-site (full-stack style)
Natural-language builders that scaffold UI, sometimes databases and deploy—same underlying models as chat + coding tools; pricing and data-handling rules differ by vendor. Includes Lovable, Emergent, and similar products students hear about by name.
Visual site builder with AI copy/layout assists on paid tiers—professional publishing workflows.
Hands-on mini projects
Each task uses one website or app: chat, research, code, websites, slides, images, short videos, and more. Follow the steps in order. At the end, save a screenshot or write a few sentences. Skip any lab if you cannot sign in; pick another.
The AI answers using files you upload—not random web guesses.
STEPS
Create a notebook. Upload 1–2 PDFs you already have (class notes, a chapter, syllabus).
Ask: “Make a simple outline from only these files.”
Ask: “Give me 5 practice questions and short answers from only these files.”
Hand in: Screenshot of the outline or questions, plus one sentence: Did anything sound like it was not in your PDFs?
Google Colab: run AI code in the cloud
~30 MIN · MEDIUM
Google Colab → File → New notebook (Google account).
Run a tiny text-AI example without installing Python on your PC.
STEPS
Open a new notebook. Runtime → Change runtime type → Python 3 → Save.
First cell: paste and run !pip install -q transformers torch sentencepiece (wait until it finishes).
New cell: paste and run this (downloads a small model the first time—may take a minute): from transformers import pipeline
sentiment = pipeline("sentiment-analysis")
print(sentiment("I enjoy this class."))
print(sentiment("This homework was confusing."))
Change the two strings to two sentences you wrote yourself.
Hand in: Screenshot showing the printed output under your two sentences. One sentence: fast or slow?
See how the same tool behaves with normal vs silly inputs.
STEPS
Pick one demo with buttons or a box to type in.
Try a normal input (real photo, clear sentence).
Try a silly input (blank, joke, or wrong language).
Write down what the demo printed each time.
Hand in: Three lines: Input 1 + result · Input 2 + result · What surprised you?
One-page website from a prompt
~25 MIN · EASY
Use a general chat AI (ChatGPT, Claude, or Gemini). You are practicing prompt → HTML/CSS, not a separate product category.
Large language models can emit markup from instructions; you copy the code into a file and open it locally—same skill whether the UI is a chat box or an IDE.
STEPS
Open a new chat. Ask: “Give me one complete HTML file (with embedded CSS) for a single-page site for our school coding club: big title, short welcome, 3 sections (meetings, projects, contact), footer with a fake email. Readable fonts, simple layout.” Ask it to put everything in one code block.
Copy the code into Notepad or VS Code. Save as club.html. Double-click the file or drag it into Chrome/Edge/Firefox to preview.
Second message: “Same page but main accent color blue and welcome text max two sentences—return the full updated HTML file.” Replace your file and refresh the browser.
Hand in: Screenshot of the page in the browser + copy-paste your first prompt + one sentence: did any tag look missing or broken?
Say who the slides are for and what must appear—then fix once.
STEPS
Ask for “about 5 slides for classmates about [pick a hobby or school topic]. Friendly tone. Include a title slide and one slide with 3 bullet risks or tips.”
Second message: “Shorten every title. Slide 3 must have exactly 3 bullets.”
Export or screenshot the deck.
Hand in: Link or screenshot. One sentence: which slide looked boring—what would you ask next time?
Describe the picture clearly: words on the poster, style, colors.
STEPS
Type a fake club event: exact title words, date, “flat cartoon style, big text.”
If the words look blurry, ask again with “very large readable text, simple background.”
Download the image you like best.
Hand in: Final image + copy-paste the prompt you used last.
Short video from a text prompt
~20 MIN · EASY–MEDIUM
Pick one that offers text → video (free credits change often): Runway, Pika, or Canva (Magic Media / video tools). If one asks for payment, try another.
Learn to steer motion like you steer images: subject, style, length, camera feel.
STEPS
Create an account if needed. Find “generate video” or “text to video.”
Prompt example: “5-second clip, cartoon cat wearing tiny sunglasses walking on a sunny sidewalk, side view, bright colors, no text on screen.”
If the site asks for aspect ratio, pick landscape or vertical once and keep it.
Generate. If you get a second free try, change one word (e.g. “rainy” instead of “sunny”) and compare.
Hand in: Download link or screenshot of the player + your final prompt. Two sentences: wait time, and did the motion match what you imagined?