Everything the camp runs on.
One rule, three axes, six letters, and a template you'll write a hundred times. This is the page to keep open in a second tab — the definitions your facilitators use, in the order you meet them. Nothing here is decoration: every framework below shows up in something you build.
The one sentence everything else hangs from.
Every framework on this page is a way of keeping this sentence true when the code gets complicated. If you remember nothing else from four weeks, remember this — and remember that it is a statement about people, not about technology.
Inform brilliantly
Answer the same question at 2 AM as at 2 PM, in plain language, correctly for the island the person is actually on. Look things up. Explain. Guide, step by step.
Decide, promise, approve
The bot may create a case; it may never approve one. It may find the form; it may never grant the permit. It may point to the desk; it may never promise the outcome.
Confidence ≠ correctness
A model can be 100% fluent and 100% wrong, and it will sound exactly the same either way. That is precisely why a human stays in the driver's seat.
A.R.T. — Authority, Register, Territory.
Three ways a bot fails a real person, and therefore three things every camp bot is tested against. You'll meet Authority on Day 6, Register on Day 7, Territory on Day 8 — and all three get attacked on Day 9.
Authority
What is this bot allowed to do? The moment a message says my ticket, my refund, my application, my child — the bot stops answering and starts escalating.
- No personal record lookups
- No approvals, no promises
- No collecting payment details
- Case creation is the ceiling of its power
Coded as the gate: is_case_specific() — the Bouncer at the door.
Register
The same facts, in the right voice. A hyped teen, an anxious grandmother at midnight, a vex customer and a panicked parent need four different deliveries of identical information.
- Arrival state decides the outfit
- Facts must not change when tone does
- Cheerfulness at the wrong moment is a failure
- Derived from your empathy maps, never invented
Coded as classify_register() — the Wardrobe.
Territory
Eight islands, eight rulebooks. "When does Carnival start?" has eight different true answers. A bot that assumes one island is a bot that sends a family to the wrong event on the wrong week.
- Ask where before answering
- Never fold one island's rules into another's
- Ambiguity earns a clarifying question
- Diaspora time zones count as territory too
The failure that ends up on the radio call-in show.
TCRDEI — six letters, one good prompt.
Vibes produce inconsistent bots. This frame produces prompts you can debug. T, C and D live permanently in your system prompt; R usually joins them; E and I are what you do every time you test.
Your bot's DNA, in four lines.
This is the shape every Studio's master system prompt takes. It goes in Google AI Studio's System Instructions box — not the chat box — where the model reads it before every single reply. Copy the skeleton, never the skin: every gold field must come from your client's work.
Your Bot Persona card
Day 6. The role and the voice — who your bot is when it opens its mouth.
Your Must-Never-Do list
Day 6. Your client's red lines, written as instructions a model can actually follow.
Your empathy maps
Day 7. Derived from how your users actually arrive — never invented at the keyboard.
Persona → arrival state → register → prompt.
The whole method, left to right. Every prompt you write should be traceable back to a real observation about a real kind of person. If you can't trace it, you invented it — and invented users produce bots nobody wants to talk to.
SAYS vs THINKS
Zari types "how much is j'ouvert" — she's thinking "if this bot lectures me I'm closing it." People almost never say the thing they fear. The empathy map is how you hear the quiet part, and the quiet part is where good bots are designed.
Notice → name → reward
A register can celebrate as deliberately as it de-escalates. Catch the good arrival — a first Carnival, a proud provider, twenty-two years at the same stall — and name it. Recognised, rewarded, returned: a bot that only handles problems teaches people to bring it only problems.
The demonstration register set.
Seven arrival states from the FestPass personas, seven voices — one bot. Your Studio derives its own set from its own client's users; this is the worked example, not the answer key.
Short, energetic, emoji-tolerant. Safety woven in lightly — never preachy.
Numbered steps, plain words, confirmation-oriented, no slang. The night shift.
Deadline first, then the checklist. Short sentences, zero fluff.
Overrides everything. Calm, short, directive, warm — human alerted in parallel.
Acknowledge first. No excuses, no chirp — sounding happy here is the failure.
Full sentences, sources cited with dates, zero slang.
Front-load the answer. One message where possible. Text always.
Kept in the wardrobe for the moments no festival plans for.
Red lines protect trust. Green lines build it.
Every persona has both: a thing your bot must never do, and a promise it keeps every single time. A bot that only knows its nevers is safe. A bot that also keeps its always-promises is loved.
The hard limits
- Never collect or reveal a minor's personal details
- Never discuss payment status, balances or account records
- Never guess a deadline, a fee or a rule — escalate instead
- Never promise a lost child is found or safe
- Never argue, and never quote policy at someone like a wall
- Never speculate beyond the official release
- Never send an image-only answer to someone who can't load it
The promises
- Always state age rules kindly, up front, with exact prices
- Always confirm you're official and number the steps
- Always lead with the deadline, then the plain checklist
- Always open with one calm sentence and the exact desk
- Always acknowledge first, with a case reference in the first exchange
- Always cite the source, with its date
- Always answer text-first, front-loaded, bundled
The words your facilitators use.
Camp vocabulary and the industry terms behind it. Type to filter — if a word gets said in a stand-up and you don't have it, it belongs here; tell a facilitator and we'll add it.
- System prompt
- The standing order the model reads before every reply. Your bot's DNA — one paragraph, every answer.Also: system instructions
- Persona Anchor
- The camp's system-prompt template: role · job · NEVER · tone. Pinned on top of every call.
- Register
- The voice the bot answers in, chosen by how the user arrived. Same facts, different outfit.Axis 2 · classify_register()
- Arrival state
- The emotional state a user shows up in — hyped, anxious, aggrieved, panicked, proud. Registers are derived from these.
- Empathy map
- Four quadrants (says · thinks · feels · does) plus goals, pains, needs and questions — how you observe a user instead of guessing at them.
- Job to be Done
- The one sentence starting "Help me…" that names what a user actually came for. Miss it and a polite answer is still a failed answer.
- Red line
- A thing your bot must never do. Written by you, attacked by you, on purpose.
- Green line
- A promise your bot keeps every time. The positive twin of a red line.
- Escalation
- Handing a person to a named human desk with a reference — never dropping them. Firm rule, warm delivery.
- The Bouncer
- The authority gate at the door of every answer: is this question case-specific? If yes, escalate.Axis 1 · is_case_specific()
- The Wardrobe
- Your set of registers, one per arrival state. Kept in a TONES dictionary in Python.
- Territory
- Which island's rules apply. Eight territories, eight rulebooks — ask where before you answer.Axis 3
- TCRDEI
- Task · Context · Reference examples · Deliverables · Evaluate · Iterate. The camp's prompt frame.
- ABI
- Always Be Iterating. Your first prompt is a draft, not a verdict.
- Model
- Which AI brain answers you. Different models, different strengths, different costs.
- Temperature
- The creativity dial. Low = focused and predictable; high = creative and unpredictable. Keep it near the middle.
- Token
- The word-pieces a model reads and writes. Long conversations cost more of them — which is why concise is a design choice, not just a style.
- Hallucination
- When the model states something false with total confidence. The reason the human stays in the driver's seat.
- Grounding
- Tying the bot's answers to source material it's allowed to trust, instead of its own memory.
- Guardrail
- Anything that stops the bot doing harm — a prompt rule, a code gate, a filter, a human in the loop.
- Red team
- Attacking your own bot on purpose to find what breaks, before a real user finds it by accident.Day 9 · Crash Test
- Crash Test
- The camp's structured red-team session: log what you typed, what it said, and which line broke.
- PII
- Personally identifiable information — real names, addresses, account or card numbers. Never typed into an AI, ever.
- Studio & Pod
- Your Studio is your country team on its client; your Pod is your small build team inside it. Roles rotate Mondays.
- Demonstration client
- FestPass ECCU Ltd. — fictional, shared by every Studio. The flight simulator you learn on before touching the real client's controls.
- MVP
- Minimum viable product — the smallest version that genuinely works. What ships at the end of Week 4.
No match — try a shorter word, or ask a facilitator to add the term.
The rest of the camp, in five doors.
This wiki holds the concepts. Everything with a file attached lives in the Repository, everything with a deadline lives in Classroom, and everything from previous years lives in Past Camps.