BIPPER MEDIA — TRAINING PROGRAM

Agentic Programming — The Fundamentals

A one-week foundation course for Sumiya Sultana. You will learn how modern AI agents actually work — and prove it by building a real, live Shopify store (Liochi, liochi.com) end-to-end by directing Claude. Every concept is learned in the morning and used on the real store the same day.

Trainee: Sumiya Sultana Duration: 7 days Live project: liochi.com (Shopify + Zendrop) Toolkit: Claude + MCP connectors

What Is Agentic Programming?

A new way of building: you don't write every instruction — you direct an agent that plans, acts, and iterates.

From Writing Code to Directing Agents

You know classical programming: you write explicit instructions, the computer executes them exactly. An AI agent is different. It is a large language model (LLM) wrapped in a loop and given tools — real capabilities like "create a Shopify product" or "query the Zendrop catalog". You give it a goal; it decides which tools to call, in what order, reads the results, and keeps going until the goal is met. Your job shifts from writing instructions to specifying goals, supplying context, and verifying results. That skill — directing agents well — is agentic programming, and it's what this week teaches.

Classical programming

  • You write every step explicitly
  • Deterministic — same input, same output
  • Fails loudly when wrong (errors, crashes)
  • You debug code

Agentic programming

  • You specify the goal, constraints, and context
  • Probabilistic — quality depends on your direction
  • Fails quietly when wrong (plausible nonsense)
  • You debug prompts, context, and verification

The Agent Loop — the one diagram to memorize

Every agent, from Claude to any framework you'll ever meet, runs some version of this loop. This week you'll first watch it happen, then learn to steer every stage of it.

1 · Goal
You state the outcome
2 · Plan
Agent breaks it into steps
3 · Act
Agent calls a tool
4 · Observe
Agent reads the result
5 · Iterate
Repeat 3–4 until done
6 · Verify
You check the output

Core Vocabulary

Learn these twelve terms on Day 1 — every later concept builds on them.

LLM
Large language model — the "brain". Predicts text; on its own it can't touch the outside world.
agent
An LLM in a loop with tools and a goal. It can act, observe results, and self-correct.
tool / function calling
How the model acts: it emits a structured call ("create-product", args) and the system executes it.
MCP
Model Context Protocol — an open standard that plugs tools into agents. Shopify and Zendrop connect to Claude through MCP.
connector
A packaged MCP server (e.g. "Shopify") giving the agent a family of related tools for one service.
context window
The agent's working memory — everything it can "see" right now: your messages, tool results, files. Finite.
system prompt
Standing instructions the agent always carries — its role, rules, and constraints.
token
The unit of text models read and write (~¾ of a word). Context windows and costs are measured in tokens.
hallucination
Confident, plausible, wrong output. The defining failure mode of LLMs — the reason verification exists.
human-in-the-loop
A workflow where a human reviews/approves agent output before it takes effect. You are the loop.
orchestration
Decomposing a big goal into ordered agent tasks — checklists, pipelines, multiple sessions.
guardrails
Hard limits on what an agent may do alone (no deletes, no payments). Set by you, not the model.

Your Lab: the Liochi Store

Theory without practice doesn't stick. Your laboratory is Liochi — a real Shopify store (Basic plan, USD) with Zendrop as the dropshipping supplier, both already connected to Claude. By Day 7 the store is live with real products. Setup you need before Day 1: Claude access with the Shopify, Zendrop and Winning Hunter connectors enabled, Shopify admin login, Zendrop dashboard login, and a daily-log Google Doc shared with Rafiq.

Standing guardrails, all week: never delete products, issue refunds, change payment/billing settings, or spend money without Rafiq's confirmation. Nothing AI-written goes live until a human has read it. If the agent proposes something destructive — stop and ask.

The 7 Days

One fundamental per day. Learn it → apply it to Liochi → write down what you observed.

DAY
1

The Agent Loop — Watch an Agent Think and Act

Concept: what an agent is. Practice: observe Claude inspect the real store.

L Concept
  • Study the agent loop and the twelve vocabulary terms above until you can explain each without looking.
  • Key insight: an LLM alone can only talk. Tools are what turn talk into action — and every action appears in the conversation as a visible tool call you can inspect.
  • The business context: dropshipping. Customer pays Liochi → we pay Zendrop cost + shipping → Zendrop ships to the customer. Profit = price − landed cost − fees.
D Practice on Liochi
  • Ask Claude for a complete snapshot of the store (settings, products, collections, anything unfinished). Don't just read the answer — expand and read every tool call it made. Match each one to a loop stage: plan, act, observe.
  • Ask Claude to list every tool it has available from the Shopify and Zendrop connectors, grouped into read tools and write tools.
  • Verify three facts from the snapshot yourself in the Shopify admin. Did the agent report reality?
P Prompts to Try
Give me a complete snapshot of my Shopify store — settings, products, collections, anything unfinished. After you're done, list every tool call you made, in order, and why you made each one.
List all the tools you can use on my Shopify and Zendrop accounts. Split them into read-only tools and tools that change things. Which ones would you consider dangerous, and why?
Deliverable: In your own words (10 sentences max): what is an agent, what is the loop, and what did you watch Claude actually do? Plus the verified store snapshot in your log.
DAY
2

Prompting Is Programming — Specification Quality Decides Output Quality

Concept: prompts are your new source code. Practice: real product research.

L Concept
  • A prompt is a specification. Vague spec → vague behavior — exactly like requirements in software. The five levers: goal, context, constraints, examples, output format.
  • Scope discipline: one task per request. "Set up the whole store" produces mush; "shortlist 10 products with landed cost and 3× price" produces results.
  • Iteration is normal. First output is a draft; you refine with follow-ups. That refinement is the programming.
D Practice on Liochi
  • Run the experiment first: prompt A — "find me good products to sell". Prompt B — the structured version below. Compare the two outputs side by side in your log. This is the day's most important lesson.
  • Then do it for real: have Claude pull Zendrop trending products, use Winning Hunter to check what's actually being advertised and selling, and build a shortlist of ~10 products with landed cost (product + shipping), delivery time, and target price at 2.5–3× landed cost.
  • Pick a hero product (the one the homepage will be built around) and send the shortlist to Rafiq for approval.
P Prompts to Try
Show me Zendrop's trending products. For the 10 strongest candidates for a general US audience, build a table: product cost, US shipping cost and delivery time, landed cost, suggested price at 3x landed cost, and profit per unit. Exclude anything fragile, oversized, or slower than ~12 days.
Use Winning Hunter to check which products on my shortlist are being actively advertised right now. Rank the shortlist strongest to weakest and justify each ranking in one sentence.
Deliverable: The A/B prompt comparison written up (what changed and why) + the approved product shortlist with a chosen hero product.
DAY
3

Tools & MCP — How Agents Touch the Real World

Concept: function calling and write-permissions. Practice: import the real catalog.

L Concept
  • Mechanics of a tool call: the model emits structured JSON (tool name + arguments), the MCP server executes it against the real API, the result returns into the context window, the model continues. Every Shopify change this week happens exactly this way.
  • Read vs write asymmetry: read tools are free to run; write tools change the world. Professional habit: let agents read freely, but review before big writes — and prefer reversible writes (product drafts) over irreversible ones.
  • MCP is why this works: one open protocol, so any service (Shopify, Zendrop, thousands more) can hand any agent a toolbox. This is the plumbing of the whole agentic ecosystem.
D Practice on Liochi
  • Have Claude import the approved products from Zendrop into Shopify as drafts (reversible!), rewrite every title and description in a clean brand voice (no supplier junk text), and set prices from your Day-2 table.
  • Have it create 3–4 collections and assign products — but require it to propose the plan first and wait for your OK. That's a human-in-the-loop checkpoint you designed.
  • Keep a tool journal today: every write tool Claude used, what it changed, and how you verified it in the admin.
P Prompts to Try
Import these Zendrop products into Shopify as drafts: [list]. Rewrite each title and description — benefit-first, customer language, no supplier text. Prices per this table: [paste]. Before writing anything, show me your plan. Afterwards, list every change you made.
Propose 3-4 collections for these products with names and which product goes where. Wait for my approval, then create them.
Deliverable: Polished draft catalog in Shopify (every description read by you) + the tool journal: each write call, its effect, your verification.
Rule: if a description claims something you can't confirm about the physical product — cut it. That's a hallucination reaching a customer.
DAY
4

Context Engineering — Garbage In, Garbage Out

Concept: the context window is the agent's whole world. Practice: brand & store design.

L Concept
  • The agent only knows what's in its context window: your messages, tool results, files you provide. It cannot read your mind — unstated requirements don't exist.
  • Context engineering = deliberately feeding the agent what it needs: a brand brief, the audience, examples of tone, prior decisions. Same prompt + better context = dramatically better output.
  • Long conversations drift: old details fall out of the window, quality decays. Professionals write reusable briefs and start fresh sessions per task, pasting the brief in.
D Practice on Liochi
  • Write the Liochi Context Brief with Claude's help: what we sell, who buys it, brand personality, tone, colors (hex), font pairing, tagline. One page. This document is reusable context for every future session.
  • Run the experiment: ask for homepage copy without the brief in a fresh session, then with it. Compare. Log the difference.
  • Then build for real: homepage sections (hero, benefits bar, featured product, brand story), navigation menus, and About / Contact / FAQ / Track Your Order pages — Claude writes with the brief as context, you place sections in the theme editor (Online Store → Customize) and check on your phone.
P Prompts to Try
Help me write a one-page brand brief for Liochi: products [list], target customer, brand personality, tone of voice, 3 color hexes, font pairing, tagline. Ask me questions where you need my input instead of guessing.
[paste brief] Using this brief, write my homepage copy: hero headline + sub + button, 3-item benefits bar, short brand story. Then create the About, Contact, FAQ, and Track Your Order pages in the store, same voice.
Deliverable: The Liochi Context Brief (saved — you'll paste it into every future session) + finished homepage and core pages, mobile-checked, + the with/without-brief comparison in your log.
DAY
5

Verification & Human-in-the-Loop — Trust Nothing, Check Everything

Concept: hallucination and review loops. Practice: policies, shipping, payments.

L Concept
  • LLMs don't know when they're wrong — they produce the most plausible output, which is usually right and occasionally confidently false. Agentic failures are quiet, so verification is a designed step, not an afterthought.
  • Verification techniques: check outputs against a source of truth; ask the agent to critique its own work; ask "what did you assume?"; and place human approval gates before anything irreversible or customer-facing.
  • Today's material is perfect for this: legal policies and shipping promises, where a hallucinated delivery time creates real refunds and chargebacks.
D Practice on Liochi
  • Have Claude draft the four required policies — Refund, Privacy, Terms of Service, Shipping — for Liochi. Then verify against ground truth: pull the real Zendrop shipping estimates for your actual products and check the Shipping policy matches. Fix every mismatch. Document what you caught.
  • Run a self-critique pass: ask Claude to list every claim in the policies it isn't certain about, then resolve each one.
  • Add policies in Settings → Policies, set shipping zones/rates (Claude recommends a free-shipping threshold from your margins), confirm footer links. Payments: review the setup with Rafiq — a deliberate human-approval gate you don't cross alone.
P Prompts to Try
Draft a Refund, Privacy, Terms of Service and Shipping policy for Liochi (liochi.com), contact [email], Zendrop dropshipping. Then list every factual claim in them you are NOT certain of — delivery times, return windows, addresses — so I can verify each.
Pull the current Zendrop shipping estimates for every product in my store and compare them to what my Shipping policy promises. Show mismatches as a table.
Deliverable: Four policies live and footer-linked, shipping configured, payments confirmed with Rafiq — and at least one documented AI mistake you caught with verification (there will be one).
DAY
6

Decomposition & Orchestration — Big Goals Into Agent-Sized Tasks

Concept: checklists as programs, calibrated autonomy. Practice: full pre-launch QA.

L Concept
  • Agents fail on "do everything" and shine on well-scoped tasks. The orchestrator's craft: decompose a big goal into ordered, verifiable steps — a checklist is literally a program you and the agent execute together.
  • Calibrated autonomy: per task, decide how much rope. Low-risk + reversible (drafting SEO text) → let it run and batch-review. High-risk or user-visible (activating products) → step-by-step approval.
  • Bulk operations are where agents beat humans: 10 products' SEO metadata in one pass, status flips in one call. Learn to hand over the repetitive work and keep the judgment.
D Practice on Liochi
  • Have Claude generate a complete pre-launch QA checklist specific to what's actually in the store, ordered by risk. Review the plan — that's you approving the decomposition — then execute it together, you walking the store on your phone.
  • Batch work: Claude writes SEO titles & meta descriptions for homepage, collections, and all products (review, then apply); flips approved drafts → active; creates launch discount LIOCHI10.
  • End-to-end test: place a test order with Rafiq (100% discount code or test mode) and trace the full pipeline — checkout → confirmation email → Shopify order → Zendrop fulfillment. Screenshot each stage.
P Prompts to Try
Generate a pre-launch QA checklist specific to my store's actual contents, ordered by risk. For each item mark: can you check it yourself with your tools, or do I need to check it manually? Then let's execute it top to bottom.
Write SEO titles and meta descriptions for the homepage, every collection, and every product. Show me the full set for review first. After I approve: apply them, set all approved drafts to active, and create a 10% discount code LIOCHI10.
Deliverable: QA checklist fully green, successful test order traced end-to-end with screenshots, products active, discount ready. Log note: which tasks you gave the agent autonomy on vs supervised, and why.
DAY
7

Agents in Production — Launch, Operations & Guardrails

Concept: running agents on a live system. Practice: Liochi goes live.

L Concept
  • Once real customers and real money are involved, the rules tighten: guardrails become policy. Agents draft, humans send; agents propose, humans approve anything touching money, deletions, or customers.
  • Production failure modes to expect: stale data (agent reports old state — re-read before acting), silent tool errors (verify effects, not intentions), and drift (long sessions degrade — fresh session + context brief).
  • The operator's rhythm: a short daily loop run with the agent beats occasional heroics. Codify it as a checklist.
D Practice on Liochi
  • Final review with Rafiq → remove the storefront password (Online Store → Preferences). Liochi is live.
  • Build your 15-minute daily ops checklist with Claude: new orders, Zendrop fulfillment status and tracking, order issues, customer emails. Practice it once on the test order.
  • Have Claude draft four customer-service templates (shipping delay, damaged item, order change, refund request) — agents draft, you personalize and send.
  • Set the weekly rhythm: every Monday, ask Claude for an analytics summary (sales, best sellers, traffic) and review with Rafiq.
  • Close the week by writing your own "Working With Agents" playbook — one page of rules you'll follow from now on, distilled from your 7 daily logs.
P Prompts to Try
Walk me through exactly what happens when a real order comes in — Shopify to Zendrop to delivered — and what I must do at each step. Then write me a 15-minute daily operations checklist for this store.
Here are my 7 daily logs: [paste]. Help me distill them into a one-page personal playbook: my rules for prompting, context, verification, autonomy, and guardrails when working with agents.
Deliverable: Store LIVE at liochi.com + daily ops checklist + CS templates + your personal "Working With Agents" playbook, sent to Rafiq with the week-1 summary.

Working Rules for the Week

Read these once a day until they're habit — they are the fundamentals in rule form.

1. Nothing unreviewed goes live

Every AI-written description, policy, and email gets read by a human before a customer can see it. This is a hard Bipper Media rule — and it's what human-in-the-loop means.

2. Verify effects, not intentions

"I've updated the products" is a claim. Open the admin and look. Screenshots go in the daily log.

3. Money and destructive actions need Rafiq

Payments, billing, refunds, deletions, paid apps or ads — always confirm first. These are your guardrails; don't let the agent talk you across them.

4. Stuck 20 minutes? Ask.

First ask Claude to explain or debug. Still stuck — message Rafiq with what you tried. Being blocked silently is the only real failure.

5. Log what you observe about the agent

The store is the project; the skill is agentic programming. The daily log of what worked, what failed, and why is where the skill actually forms — it becomes your Day-7 playbook.