INITIALIZING BUNKROS IDENTITY LAB
LOC UNDERGROUND
SYS --:--:--
Learning / Prompt Engineering

Command, Do Not Ask.

The machine is not a mind reader. It's an eager, blindingly fast, completely literal intern. Give it a lazy instruction and it returns a lazy, generic result. This class stops you treating the prompt box like a search engine and teaches you to treat it like a script — with a four-block anatomy, few-shot examples, and chain-of-thought logic.

You'll be able to

Assign a persona

Constrain the AI's tone and vocabulary with role-play.

You'll be able to

Use the four blocks

Structure any prompt: Persona · Context · Task · Format.

You'll be able to

Show, not tell

Run few-shot prompting to force your exact style.

01 — Warm-up

Rate the prompts

Read each prompt and rate how effectively it will instruct the AI, 1 to 5 stars. Then see if you were right.

Self-Check · Rate each prompt
"Please write a nice blog post about coffee for me, thank you!"
"Act as a master barista. Write a 300-word blog post on the difference between Arabica and Robusta. Tone: passionate, slightly nerdy. Format: bullet points."
02 — Foundation

Direct the performance

Think of the AI as a skilled actor walking onto a set. Yell "Be sad!" and you get a clichéd performance. To get something brilliant, give the actor four things: a script, a backstory, a motivation, and strict boundaries to stand within. In the digital world that's prompt engineering: define the Persona (who the AI acts like), the Context (why it matters), the Task (the action), and the Format Constraint (how the answer must look).

Drag the slider below to see the same idea as a lazy prompt versus a directed one.

Lazy vs Directed — drag to compare
Directed · 40-word engineered prompt →

"Act as a senior brand strategist. We're launching a queer-owned bakery. Write 3 punchy taglines under 6 words each, warm and a little cheeky. Format as a numbered list." → sharp, on-brand, usable in seconds.

← Lazy · 6-word prompt

"Write something about my bakery." → a generic, grey wall of corporate text that says nothing and helps no one.

Pause & notice

Ever got a massive wall of text when you wanted a quick summary? The machine didn't fail — you forgot to set a format constraint.

Mini-recap

Stop talking to the AI like a friend. Speak to it like you're programming a machine with a strict set of rules.

03 — Visualise

Prompt anatomy

Every strong prompt snaps together from four blocks. Miss one and the AI fills the gap with a guess.

Block 1

Persona

Who the AI acts like. "Act as a senior graphic designer…" — this narrows its vast vocabulary to one specialist voice.

Block 2

Context

Why the task matters. "…creating a brand identity for a local bakery…" — grounds the answer in your situation.

Block 3

Task

The exact action. "…brainstorm 5 unique colour palettes…" — one clear imperative verb up front.

Block 4

Format

How it must look. "…as a Markdown table with hex codes." — the box the AI must stay inside.

04 — Interactive

The modular prompt builder

Pick one block from each category and watch the assembled prompt — and a simulated AI output — update live. Notice how each added constraint sharpens the result.

Sandbox · Stack your constraints
Persona
Context
Task
Format
Your engineered prompt
Pick one block from each category above…
Simulated AI output
Waiting for your constraints…
05 — Deep dive

Forcing logic

Logic

Chain of Thought

Models generate one token at a time, so they stumble on multi-step logic. Append "Think step-by-step and write out your reasoning before the final answer." This gives the model a runway to align its probabilities before it commits.

Style

Few-Shot Prompting

When you can't describe a tone, show it. Paste 2–3 examples of your own writing, then: "Write a new one using the exact same tone and formatting as the examples." Pattern-matching beats description.

A common misconception — politeness

The machine has no feelings. "Please", "Could you", "Thank you" waste tokens in your context window and dilute the command with conversational probabilities. Command the machine firmly.

Takeaway

Structure is the only language the AI understands. If you want a specific output, explicitly build the box you want it to stay inside.

06 — In the Wild

Maya pitches a venue

Maya organises a community drag show and mutual-aid fundraiser, and needs a persuasive proposal for cautious venue owners — but struggles with formal business writing.

Before

The lazy prompt

"Write an email asking a venue if we can host a drag show there." → an overly casual, rambling email that sounds unprofessional.

After

The four-block rewrite

Persona: a professional, persuasive event director. Context: a profitable, community-focused arts event. Task: pitch the venue owner on financial benefit and goodwill. Format: under 200 words, formal-but-warm, with 3 bulleted benefits.

Result: a flawless, compelling business pitch focused on the economics — the casual tone gone.
Why it matters

The AI has the vocabulary of a business executive — but it won't use it unless you assign it that persona. You must direct the performance.

07 — Research

Study the frontier

Thousands of developers publish their best prompt architectures for free. Go find a real system prompt — the hidden master instructions an AI gets before you ever type.

Search for

What to look up

"GitHub awesome prompt engineering" · "open-source system prompt library" · "Anthropic Claude prompt library".

Judge the source: prefer libraries hosted by the AI makers (OpenAI, Anthropic) or highly-starred GitHub repos. Warning: never pay for "prompt packs" advertised on social — the best techniques are published free.
Your findings
Saved locally on your device — never sent to a server.
08 — Assignment

The boss rewrite

You need a two-week extension on a project. The lazy prompt — "Write an email to my boss asking for more time" — will produce a poor email. Rewrite it as a prompt using all four blocks, with imperative verbs and zero filler.

Write your engineered prompt
Saved locally on your device.
Success rubric — tap to expand

Needs work: writes the actual email to the boss instead of a prompt for the AI.
Getting there: a prompt, but missing one or two blocks (e.g. no format constraint).
Solid: all four blocks, but still uses conversational filler.
Excellent: a strict imperative command with Persona, Context, Task and Format — no wasted words.

09 — Retrieval practice

Quiz — 5 questions

Question 1 · Medium
Which is an example of "few-shot prompting"?
A
Asking the AI to write a story in exactly 50 words.
B
Giving the AI three examples of your own writing and telling it to match that exact style.
C
Telling the AI to "think step-by-step."
Question 2 · Easy
Why omit "please" and "thank you" from prompts?
A
They waste tokens in the context window and dilute the strictness of the command.
B
They offend the AI's safety filters.
C
AI models respond negatively to human emotion.
Question 3 · Easy
What is the purpose of the "Persona" block?
A
To tell the AI your own name so it can personalise the chat.
B
To force the AI to adopt a specific vocabulary and viewpoint (e.g. "Act as a senior lawyer").
C
To change the colour of the chat interface.
Question 4 · Medium
An AI gets a complex maths puzzle wrong. Which addition most likely fixes it?
A
"Try again, but be smarter."
B
"Format this as a bulleted list."
C
"Think about this step-by-step and show your reasoning."
Question 5 · Easy · True / False
"If you omit the Format Constraint, the AI will refuse to answer." — True or False?
T
True
F
False
10 — Avoid these

Common mistakes

Trap

Blank-page syndrome

Treating the box like a Google search: "Marketing ideas for my business." Fix: keep a sticky note reading "Act As / Context / Task / Format" and paste it every time.

Trap

Conversational dilution

"Hey there, could you please help me write a quick email if you have time?" Fix: delete every word that isn't a command or constraint. "Write an email."

Trap

The moving goalpost

Arguing with a bad output across messages ("no, shorter… now longer…"). Fix: don't argue — edit the original prompt and regenerate fresh.

Trap

Assuming format comprehension

Asking for a "report" and getting an essay. Fix: never use generic format nouns — "Output as a Markdown table: Date, Status, Next Steps."

Debug your thinking — check each before you send
11 — Reflect

Directing vs collaborating

How does explicitly defining tone and persona change the AI's "personality"? By curating these personas, are you collaborating with the machine, or simply directing it?

Optional: if prompting requires strict imperative commands, how might speaking to machines for hours a day shape how we talk to humans? Does few-shot mimicking your own voice feel like an expansion of your capabilities, or a surrender of authenticity?

Private journal
Saved locally to your device — never sent to a server.
Well done

You've moved from passive consumer to active director — you have the syntax to command the machine. Next, we confront the ethical weight of these systems and the biases hidden in the code.

12 — Reference

Glossary — precise words

Prompt engineering
Structuring text instructions so an AI produces a specific, predictable output.
Persona
The character or profession assigned to an AI to constrain its vocabulary and tone.
Format constraint
A strict rule for the output's layout or length (bullets, JSON, 100 words).
Zero-shot prompting
Asking the AI to perform a task with no examples of the desired output.
Few-shot prompting
Giving the AI several examples to set a pattern before the final task.
Chain of thought
Telling the AI to "think step-by-step," which sharply improves logic and maths.
System prompt
The hidden master instructions a developer sets before any user interaction.
Prompt injection
A malicious prompt that tricks an AI into ignoring its rules or system prompt.