AI Foundations

AI went from a novelty to an everyday tool faster than almost any technology before it. This course is a calm, practical map: what the main tools are, what each is good at, and how to use them with good judgment — no hype, no jargon.

You don't need to be technical. By the end you'll be able to pick the right tool for a task, write clearer prompts, and understand the new wave of 'agentic' tools that can take actions for you.

Chat assistants are the front door to AI. You type a request; they respond. The skill is in how you ask: give context, say what good output looks like, and iterate.

• Claude Desktop — a desktop assistant that can also connect to your files and tools (see the Connectors lesson).
• ChatGPT — a general assistant strong at writing, reasoning, and — increasingly — images.

Rule of thumb: treat the assistant like a sharp, fast intern. Be specific, review its work, and never paste secrets or private data you wouldn't want stored.

Modern assistants can generate and edit images from a description. The craft is describing subject, style, composition, and lighting clearly, then refining.

• ChatGPT image generation — good for illustrations, mockups, and edits.
• Gemini (and its image model nicknamed 'Nano Banana') — strong, fast image creation and editing.

Always keep cultural and religious imagery respectful and accurate — AI can get sacred details wrong, so verify before sharing.

There isn't one 'best' AI — different models lead at different things and the leaderboard changes monthly. A good habit in 2026: match the model to the task (writing vs. coding vs. images vs. cheap-and-fast).

OpenRouter is a single doorway to many models from different companies — handy for trying options without separate accounts. Start cheap and small; reach for a frontier model only when the task is hard.

The newest shift is agentic tools that don't just answer — they take steps to complete a goal (research, write code, edit files). OpenAI's Codex-style tools can act like a tireless assistant working through a task list.

Power comes with care: give clear goals, review what they did, and keep a human in the loop for anything that matters.

Connectors (often built on the Model Context Protocol, MCP) let an assistant securely reach your files, calendar, or apps — so it can work with your real information instead of guessing.

Security first: only connect what you need, review permissions, and never grant write access you can't undo. This same care is how we built this very university.

Every AI model processes language in chunks called tokens. A token is roughly 4 characters — when you send text to Claude or ChatGPT, it counts tokens, not words. This matters because models have a context window: a maximum number of tokens they can 'see' at once.

Think of it like a reader's memory. A model with a 200k token context window can hold roughly 150,000 words — that's a full novel plus a conversation. But a smaller window (8k tokens) fits maybe a few pages. When you paste a 50-page document and ask a question, a small context window might miss important sections.

Why This Matters for Your Work

When you're asking an AI to research something, draft parchar materials, or analyze a long Gurbani passage, the context window affects quality. Claude and newer GPT models have large windows (100k+), but older or cheaper models don't. If you upload a 100-page manuscript and the model only 'sees' the first 30 pages, it can't give you a full analysis.

Tokens also affect cost. Providers charge per token (sometimes per 1,000 tokens). A 5,000-token prompt costs more than a 500-token one. Learning to write tighter, clearer prompts saves money and gets faster responses.

Practical: How to Measure Your Prompts

Use token counters (Claude has one built in; ChatGPT will show token count in Plus and Pro accounts). Before you paste a huge document, ask: 'Is this request bigger than my model's context window?' If yes, split it into smaller pieces or use a model with a larger window.

For parchar work — drafting talks on Amritsar's history, analyzing a Raag's emotional arc — you rarely need more than 20k tokens. But for research that pulls together dozens of sources, a large window is worth the extra cost.

• Tokens are the currency of AI language — manage them like bandwidth.
• Context windows set the ceiling for what an AI can process at once — know your model's limits.
• Cheaper prompts come from tighter writing and clearer requests.

AI models are powerful, but they have real blind spots. They're trained on data from the internet — which is skewed toward English, the Global North, certain worldviews. For Sikh knowledge, this is critical: models may not know recent scholarship, may confuse Sikhi with Hinduism or Buddhism, may not understand the spiritual weight of a Shabad without context.

A model can also sound confident while being wrong. This is called hallucination (covered in earlier lessons), but there's a deeper issue: structural bias. If a model was trained partly on texts written by authors unsympathetic to Sikhi, it may carry those biases forward. When you ask about the Khalsa or Amrit, the model might reflect colonial-era or brahminical framings if it wasn't explicitly trained to recognize and counter them.

When NOT to Use AI

Never use AI to generate original Gurbani quotes. Never ask it to make up historical facts. Don't use it to replace reading primary sources — Gurbani, Rehit Maryada, or verified histories. Don't ask AI to make decisions for you about your own spiritual practice, guiding someone's naam simran, or resolving a personal crisis.

AI also shouldn't be your only source when stakes are high. If you're teaching, preparing for an exam, or writing something that will shape how people understand Sikhi, verify AI's claims against real sources. Cite those sources, not the AI.

Limits in Gurbani and Theology

Gurbani is poetry — nuance, meter, and spiritual depth don't always survive paraphrase. A model can summarize a Shabad's themes but may flatten its lyricism. For serious gurbani work, read the actual text, listen to santhyas (traditional recitations), and consult human scholars.

Theology — especially questions about hukam, moh (attachment), or spiritual experience — requires lived understanding. Use AI to organize your thoughts, draft arguments, find sources. But the final wisdom comes from your own practice and community.

• Bias in training data means models may misrepresent non-Western traditions — verify everything for Sikhi-specific work.
• Confidence ≠ accuracy — even smooth answers can be wrong.
• Some work requires human judgment, lived experience, and primary sources — keep AI in its lane.

Good prompts are specific, kind, and clear. They tell the AI: what is your role, what is the context, what does success look like. For Gurbani work and parchar, this is an art.

A bad prompt: 'Explain the meaning of Gurbani.' Too vague. The model will default to a generic summary.

A better prompt: 'I'm writing a 15-minute kirtan talk for high school students new to Sikhi. Explain the concept of Hukam (divine order) from Gurbani, using 2-3 examples that show how it connects to their daily choices. Avoid jargon; use conversational English.'

This prompt gives the AI: audience (high schoolers new to Sikhi), purpose (kirtan talk), format (15 minutes), concept (Hukam), scope (2-3 examples), tone (conversational). The AI can now give you something useful.

Prompts for Different Tasks

Translation help: 'Translate this line from Gurmukhi to English, keeping the poetic rhythm. Explain one image or metaphor that's hard to translate.' (Give the actual line.)

Research summaries: 'Find [three sources] on the Khalsa's role in [historical period]. Summarize what each says in one paragraph, then tell me what's agreed and what's debated.'

Draft editing: 'This is a draft parchar talk on [topic]. Is the argument clear? Where does it get confusing? What would you add for a Canadian Sikh audience, age [25-45]?'

Fact-checking: 'Is this claim true? If yes, tell me the source. If no, what's the right fact? [Claim].' Then verify the AI's answer yourself.

The Iterate Loop

Your first response from AI is rarely final. Ask follow-ups: 'Make it shorter, that was preachy,' or 'Add more of this, cut that part.' Treat the AI like a co-worker. Show it what you're actually building (a website, a talk, a lesson) and ask for revision rounds.

• Specific prompts beat generic ones — tell the AI who, what, why, and what success looks like.
• Context is your friend — explain the background, the audience, the stakes.
• Iterate — first answers are drafts, not finished work.

AI is a co-pilot for language work. It can help you translate, parse Gurmukhi grammar, check whether your Punjabi is idiomatic, and generate practice materials — but it's not a substitute for a teacher.

Gurmukhi script recognition is improving: Claude and ChatGPT can usually read Gurmukhi text if you paste it or image it. You can ask: 'Transliterate this line into Roman Punjabi,' or 'What does this word mean in the context of Gurbani?' The AI will give you a starting point.

Translation with AI

For Gurbani, never ask AI alone to translate. Instead, use it to propose translations, then check against scholarly ones (SGGS-USA, Talib Singh's works, or published santhyas). Ask the AI: 'Here's my translation of this Shabad. Does it capture the meaning? What am I missing?' This way you learn and catch errors.

For modern Punjabi — a letter, a speech, a social media post — AI is much more reliable. You can draft in English, ask Claude to render it in conversational Punjabi, and refine. Punjabi Sikhs in your sangat can review the final version.

Grammar and Writing Checks

Gurmukhi grammar is intricate (case endings, word order, the ਖ਼ and ਖ distinction). AI can explain these, but not perfectly. When learning Gurmukhi script, use AI to check your transliterations and parsing, but verify against a grammar guide or a human speaker.

For Punjabi writing — whether you're composing an ਅਰਦਾਸ note, a thank-you letter, or parchar material — ask AI to: 'Check this for grammar and idioms. Is this how a native speaker would phrase it?' Then ask a trusted Punjabi speaker to review.

Limits

AI struggles with Gurbani's metaphorical depth and spiritual layers. It can parse the surface meaning but misses the 'Shabad experience' — the raag, the santhya, the silence after. Use AI for homework help, translation brainstorms, and script practice. For deep Gurbani study, sit with a teacher or a santhya recording.

• AI can propose translations and parse grammar — but always verify against scholarly or native sources.
• Use AI as a learning tool, not a shortcut — the struggle to understand deepens the knowledge.
• Script and transliteration are AI's strength; spiritual interpretation is not.

The power of AI isn't in any single tool — it's in how you chain them together. A workflow is a sequence: 'First I use Claude for research, then I organize in a spreadsheet, then ChatGPT helps edit.' For Sikh parchar, scholarship, or personal projects, building your own workflow saves time and keeps quality high.

Sample Workflow: Preparing a Parchar Talk

Step 1 (Research): Ask Claude, 'Outline a 20-minute talk on [topic]. What should I cover? What sources exist?' Step 2 (Draft): 'Write a 5-minute intro that hooks a [demographic] audience.' Step 3 (Refine): Paste the draft and ask, 'Make this conversational. Where's the argument weak? What's unclear?' Step 4 (Check): Fact-check any claims against real sources. Step 5 (Practice): Use a recording tool (not AI) to hear yourself say it aloud. Revise based on how it lands.

Sample Workflow: Writing Academic or Gurbani-Focused Scholarship

Step 1: Ask AI to find sources on your topic (it won't find everything, but it's a starting point). Step 2: Use AI to organize your notes — 'I have notes on [themes]. Group them by argument.' Step 3: Draft sections — 'Write a paragraph on [this angle] for an academic audience unfamiliar with Sikhi.' Step 4: Have AI help structure logic — 'Does this argument flow? Are there gaps?' Step 5: Copy-edit for clarity and tone. Step 6: Have a human expert read it.

Tools to Chain Together

Most workflows use 2-3 tools: Claude for research and drafting, ChatGPT for a second opinion, Perplexity (web search AI) for fact-checking, a spreadsheet for organizing. Don't use AI for everything — some steps are faster by hand (reading a source, listening to a Shabad, talking to a mentor).

Keep a Prompt Library

Save prompts that work. Keep a note of: 'For parchar prep, ask AI to [specific structure]' or 'For gurbani translation help, give AI [this format].' Over time, you'll notice which prompts yield the best results for your work. Reuse and tweak them.

• Workflows chain tools together in a sequence — plan before you start.
• Each step has an AI tool, a human step, or both — know which is which.
• Prompt libraries save time — collect what works for you.

AI is moving beyond 'you ask, AI answers.' The next wave is agents — AI systems that can take actions for you. An agent can draft an email, send it, read the reply, and adjust. It can research a question across multiple sources, synthesize findings, and hand you a memo. For Sikh work, this opens new possibilities and new risks.

What Are Agents?

An agent is an AI given access to tools and autonomy. Instead of giving it a prompt and waiting for text, you give it a goal: 'Research this historical figure and write a 500-word biography.' The agent then: searches the web, reads sources, compares them, writes, and presents the draft — often without you asking each step.

Tools an agent might use: web search, calendar access, email, file storage, spreadsheets, publishing platforms. Think of it as hiring an intern who can work unsupervised for stretches.

Opportunities for Parchar and Scholarship

An agent could: monitor Sikh news and compile a weekly summary. Draft multiple versions of a parchar talk and present them for your review. Organize research notes across dozens of sources. Translate a new article from Punjabi news. Manage a social media calendar for your parchar initiative. These tasks eat time today; agents could handle them in the background.

Real Risks

Agents are less transparent. You don't watch every step, so errors hide more easily. If an agent sends an email or posts something, you need to review before it goes live — but you might forget. Agents amplify both quality and mistakes.

For Sikh content, the stakes are higher. An agent that misrepresents Gurbani or perpetuates bias does it at scale. A parchar agent that gets Sikh history wrong reaches hundreds. These tools are coming; the discipline to use them well is on us.

Not Yet Ready for Unsupervised Work

Don't set an agent loose on Gurbani work, scholarship, or parchar without strong guardrails. Right now, an agent summarizing a Shabad might miss the spiritual layer. One writing parchar might flatten nuance. Review its work carefully. As tools improve, you'll trust them more — but today, agents are for tasks where clarity, not poetry, matters.

• Agents are AI with access to tools and light autonomy — powerful, but less transparent than chat assistants.
• Good uses: research synthesis, content organization, scheduling, monitoring. Bad uses: unsupervised spiritual or scholarly work.
• Agents are coming; building judgment now prepares you for later tools.