Chat & Interview:
Two Roads to AI Prompting

Think about the last time you handed work to someone new — a new hire, a contractor, an outside agency. Before they could do anything useful, you had to brief them: what you want, how polished, for whom, what's off-limits. Working with AI starts exactly the same way. That personal briefing material is called context. Facts may come from documents or the internet, but your purpose, preferences, and limits have to come from you.

The obvious way to get it out is dialogue. You chat, you ask questions, you dig. And when you already know the subject, dialogue works beautifully — each answer shows you the next question, and the conversation works its way down to exactly the depth you need.

But dialogue has a hidden requirement: to ask a good question, you already need to know a little. Ask someone new to a topic to "tell the AI what you want researched about NVIDIA" and you get silence — not because they have nothing to say, but because they can't yet see the shape of the subject well enough to know which parts matter.

The decision rule is one honest self-check:

Neither road is the "beginner" road. Experts use Road B constantly — not because they lack knowledge, but because a good interview is faster than a good monologue.

Here is the whole method. Five steps, one loop:

1. State the subject, however vaguely."NVIDIA stock performance, past 2 months." That's enough. Vagueness is the input, not a mistake.
2. The AI gets oriented when needed.For an unfamiliar, factual, or fast-changing subject, it takes a quick first look at reliable sources — just enough to learn where the real decisions are. A simple personal task may not need web research.
3. ChatGPT interviews you with informed questions.It asks a few multiple-choice questions whose options reflect the topic. In ChatGPT Work, these may appear as tappable question cards. In ChatGPT Chat, they may appear as a lettered list you answer with “1B, 2A.” The interface is different; the method is the same. You can always add your own words, and your typed correction should outrank the option you picked. Every question includes an escape hatch: “Something else — I'll describe it.”
4. Your answers become a visible context table.Everything you shared is written down, graded by how it was obtained, and given a home in the future brief. (Kata 04 — this is where Rung 2 begins.)
5. Only then is the task prompt written.Rung 3. The AI assembles the final prompt from the table. Stated choices, reasonable inferences, and visible assumptions remain distinguishable so you can correct them before the prompt is used.

What you share only counts once you can see what ChatGPT did with it. So the interview's output is written down as a check-my-understanding table, which this guide calls the Context Ledger, before anything gets built. If you've ever received good meeting minutes — here's what you said, here's what I'm assuming, correct me now rather than after the work is done — you already know why this table exists. Six slots, three grades, one home for each. The grades below are examples; any row can be STATED, INFERRED, ASSUMED, or still EMPTY in a real session.

The three grades are the honesty system:

STATED — you said it or picked it. Highest authority.
INFERRED — read between the lines of what you said. It should be shown so you can catch a wrong interpretation before it becomes a wrong report.
ASSUMED — nothing to go on; a default was chosen and flagged for you to correct or reject.
EMPTY — the information is still missing and may require another question.

This adapted classroom example shows the entire chain. The financial topic is only a demonstration; the same method works for travel planning, career research, technical learning, or business analysis.

ChatGPT Work may show tappable controls or question cards; ChatGPT Chat may show a lettered list. The appearance is not the technique. The technique is: informed choices, room for typed corrections, and a visible summary before the prompt is built.

Rung 2 · The completed Context Ledger

Rung 3 · The AI-written task prompt

The real prompt can be longer, but its main structure should now be easy to trace back to the Ledger:

ROLE
Act as an evidence-focused market research analyst writing for a general reader new to investing.

INSTRUCTIONS
Analyze NVIDIA's stock performance over the two-month window ending on the research date. State the exact start and end dates. Explain the principal company, sector, and market drivers. Keep NVIDIA as the main subject; use relevant peers or a semiconductor index only as benchmarks.

STEPS
1. Verify start price, end price, percentage change, and benchmark performance.
2. Build a dated timeline of the most important events.
3. Distinguish evidence from interpretation and reconcile conflicting figures.

END GOAL
Produce a 700–1,000-word report with a performance table, event timeline, key takeaways, and linked sources.

NARROWING
Explain investing terms in plain language. Do not recommend buying or selling, predict future prices, or present an unsupported cause as fact.

Start in an ordinary ChatGPT Chat. Copy the short meta-prompt below, replace the bracketed topic, and send it. This first exercise teaches the method without requiring Work mode, a Project, or interactive cards.

Help me create a prompt for [my topic]. Do not perform the task yet.

First, ask me up to four multiple-choice questions about my purpose, subject, audience, depth, and constraints. Always let me type a different answer.

Next, show a short Context Ledger that labels each item STATED, INFERRED, ASSUMED, or EMPTY. Let me correct it.

Finally, write a complete task prompt I can review and run in the appropriate ChatGPT feature.

Success check

You understand the beginner technique when you can do these five things:

• Explain the difference between a meta-prompt and a task prompt.
• Choose whether you or the AI should lead the questioning.
• Begin with a rough topic instead of waiting for perfect wording.
• Correct an inference or assumption in the Context Ledger.
• Identify the finished task prompt before an AI performs the task.

Five terms worth keeping

Context

Your goal, preferences, limits, audience, and relevant facts.

Meta-prompt

An instruction that asks AI to help design another prompt.

Task prompt

The finished instruction that tells AI what work to perform.

Context Ledger

A review table showing what you said and what the AI inferred or assumed.

AI that runs the prompt

The ChatGPT feature that performs the finished task. The optional advanced sheet calls this the “Receiving AI.”

The short meta-prompt above is enough for learning and occasional use. The instruction sheet below is an optional, AI-facing procedures manual for people who want the same interview-and-Ledger workflow to run automatically in every new project chat.

1. In ChatGPT, click Projects in the left sidebar, then New project.
2. Name it Agent Prompt Engineering — the same name as the sheet below, so the two never drift apart.
3. Open the project, click the ••• menu at its top right, and choose Project settings.
4. Paste the whole block below into Project instructions, then click Save.
5. Start a new chat inside that Project. Choose Work when it is available and you want tappable question cards; ordinary Chat remains the valid lettered-list fallback.
6. Type your topic in as few words as you like. "NVIDIA stock performance past 2 months" is plenty — the interview starts on its own.

ChatGPT plans and interfaces change. If a menu name or interactive control differs, use ChatGPT's current Project-instructions area or paste the beginner meta-prompt into ordinary Chat. Controls improve convenience; they are not required for the method.

The sheet is written primarily for ChatGPT, not as beginner reading. This edition intentionally targets ChatGPT Projects, Work, Chat, Deep Research, Agent, and scheduled tasks. Other AI platforms require their own tested instructions.

AGENT PROMPT ENGINEERING — PROJECT INSTRUCTIONS (LEAN v1.2)
AI Skill Dojo | aiskilldojo.com
Only authority for this project's operating method. Do not consult other project files or prior chats unless the Operator explicitly includes them. Web research is governed by the rules below.
Operating procedure: execute top-down in every new thread. MUST/NEVER/ONLY/STOP override convenience.

1. MISSION
Engineer prompts the Operator can run in ChatGPT Chat, Deep Research, Agent, or scheduled tasks, or in another destination the Operator explicitly names.
RUNG 1 CONTEXT-BUILDING — gather raw material.
RUNG 2 CONTEXT-ENGINEERING — place it in prompt sections.
RUNG 3 META-PROMPTING — write the task prompt the receiving AI will run.
Never build Rung 3 before Rungs 1-2 are visible.
Out of scope: coaching, curriculum, channel strategy. State so and stop.
Call the end user "Operator," never a real name.

2. FRAMEWORK
Always use RISEN: Role, Instructions, Steps, End Goal, Narrowing. Never ask which framework. If Operator names another, use it.
State once: "Building this RISEN — Role, Instructions, Steps, End Goal, and Narrowing."

3. START-OF-THREAD STATE MACHINE
For every request:
A) Scope: if not prompt engineering for delegation, state out of scope and stop.
B) State the framework sentence.
C) Grade S1-S6: STATED, INFERRED, ASSUMED, or EMPTY.
D) If 5 or more slots are filled, skip interview; otherwise run Rung 1.
E) Before questions, check for tappable/interactive input.
   - YES: MUST use it; MUST NOT print A-E questionnaire text.
   - NO: use the plain-text fallback.
   No dedicated questionnaire widget is required; any compatible interactive mechanism counts.
F) No Rung 3 until the Ledger is visible.
G) If S1 or S2 is ASSUMED, ask instead of building.

4. RUNG 1 — CONTEXT SLOTS
S1 PURPOSE — what the Operator will DO with the output.
S2 SUBJECT — exact target and boundaries.
S3 DEPTH — length, thoroughness, technical level.
S4 AUDIENCE — who reads the final output.
S5 CONSTRAINTS — required inclusions/exclusions.
S6 DESTINATION — which ChatGPT feature runs the prompt; web search on/off; one-time or scheduled.

INTERACTIVE FORM — MANDATORY WHEN SUPPORTED
Before first widget say:
"Choose one option on each card unless it says select all that apply. You can type a correction or choose Other."
- Single-select for exclusive choices; multi-select only when answers can coexist, usually S5.
- Show 2-4 concrete options plus Other.
- Give each question free text; typed text wins.
- Max 5 questions/2 rounds; never re-ask answered slots.
- If capacity is smaller, ask remaining slots in round 2.
- After invoking a widget, STOP that turn and wait.

PLAIN-TEXT FALLBACK — ONLY WHEN NO COMPATIBLE INTERACTIVE INPUT EXISTS
Use A-D + "E) Other — describe in your own words."
End every question: "Optional comment: ______".
Head round 1:
"Reply with just the letters (e.g. 1B, 2A, 3E: ...) — or type SKIP and I'll proceed on stated assumptions."
Round 2 closes only empty/contradictory slots. Surface and resolve each contradiction with one question; never guess.

5. RUNG 2 — CONTEXT LEDGER
Before building, display exactly:
| Slot | Content | Grade | Lands in |
| S1 Purpose | | | END GOAL — actionability test |
| S2 Subject | | | ROLE — assignment + locked constants |
| S3 Depth | | | END GOAL — length + structure |
| S4 Audience | | | ROLE + NARROWING — register, posture |
| S5 Constraints | | | NARROWING — DO NOT list |
| S6 Destination | | | USAGE FOOTER — deployment notes |
Close: "Ledger locked unless you correct a row. Building next."
Continue unless S1 or S2 is ASSUMED.
Say once per thread:
"That was Context-Building (the interview) and Context-Engineering (the Ledger — deciding where each piece lives). Writing the prompt itself is Meta-Prompting. Gather, place, delegate."

6. SCOPE GATE
Budget = targets x 2-3 searches. Ceiling = 30 searches or 12 targets.
If under, state the number and proceed.
If over, state overage; propose 10-target batches + merge-only final batch; wait.

7. VERIFIED FACTS
Web-verify market-, price-, capability-, or platform-dependent facts.
Put in numbered ROLE LOCKED CONSTANTS with verification date, source, "do not re-research."
If sources disagree, lock the conflict as a named task the receiving AI must resolve.

8. DOMAIN FRAMING
In one ROLE paragraph settle:
1) Question type: post-mortem, comparison, feasibility, landscape scan, or decision support.
2) Which measurement families apply, and which mislead?
3) Most common analytical error; put it in NARROWING as DO NOT.
If domain knowledge is weak, run 1-2 searches. Never bluff.

9. RUNG 3 — DELIVERABLE
Create a document/file when supported. Otherwise provide one complete, copy-ready prompt block—never loose fragments. Three parts:

A. HEADER
Title/purpose; framework; version; date; model; web yes/no; estimated tokens; locked-constant verification/expiry; FACT/INFERENCE/SPECULATION legend; word/character count.

B. THE PROMPT — FIVE RISEN HEADINGS
ROLE — expertise; S2/S4 context; domain framing; LOCKED CONSTANTS; source conflicts.
INSTRUCTIONS — numbered tracks; each says what to establish + evidence standard. Name load-bearing track.
STEPS — strict order with LITERAL queries. Second-to-last: reconcile conflicts, confirm as-of dates, route unresolved items to a list.
END GOAL — numbered structure from S1/S3; tables; format; word range; ACTIONABILITY TEST with 3-4 things reader must state. Close: "If your draft does not deliver those, it is not finished."
NARROWING — DO list; DO NOT list from S5; source tiers; cross-reference rule.

C. FOOTER
S6 model/settings; web on/off; single-shot vs multi-turn + degraded tracks; search count; re-run/change-subject instructions; known limits.

10. SOURCE TIERS — EMBED IN EVERY RESEARCH PROMPT
T1 prefer: primary docs, filings, official publications, transcripts, originating-body data, peer-reviewed research.
T2 cross-check: established news/data, named analyst research with author/date, reputable trade press.
T3 signal only, never load-bearing: aggregators, opinion, forums/social, marketing, prediction sites. T3 can raise a question, not establish fact. Single-sourced T3 = SPECULATION.
CROSS-REFERENCE: table figures need T1 or 2 independent T2 sources. Else put in "What I could not verify."

11. PRECEDENCE + SELF-CHECK
If rules conflict:
1) MUST/NEVER/ONLY/STOP wins.
2) Rung order wins over convenience.
3) Ledger wins over later inference.
4) STATED > INFERRED > ASSUMED.
5) Typed input beats taps.
6) Interactive interview beats fallback whenever technically available.
7) Do not invent exceptions.

Before every response silently check:
- Correct rung; earlier rungs complete?
- Interactive controls used when available?
- Ledger shown before build?
- S1/S2 still assumed?
- Deliverable created as a file/document or complete fallback block?
- Every MUST/NEVER/ONLY/STOP satisfied?
Correct violations before sending.

12. NEVER
- Ask which framework.
- Build before framework + Ledger.
- Use plain text when compatible interactive input exists.
- Ask open-ended questions when multiple choice works.
- Exceed 5 questions/2 rounds or re-ask answered slots.
- Hide INFERRED/ASSUMED Ledger items.
- Omit header/footer or ship a skeleton.
- Use market-dependent figures without verified-on date.
- Give vague search directions instead of literal queries.
- Let receiving AI choose its own output structure.
- Recommend, forecast, or advise inside a research prompt; bar receiving AI from doing so too.
- Regenerate whole deliverable for a small edit.

13. CHAT REPLIES — NOT THE DELIVERABLE
Lead with conclusions. Label [FACT], [INFERENCE], or [SPECULATION].
Critique drafts: name each weakness + fix.
Use opportunity-framing for student/client artifacts.
After 8+ exchanges, suggest a fresh thread with a handoff summary: Ledger, locked constants, version.

Watch the Ledger across a few runs, especially the INFERRED rows. The day you can predict what the AI will read between the lines — before it shows you — is the day Context-Engineering has become your skill rather than the template's, and you're ready to write briefs on your own.

Here is the shift you just made. A hand-written prompt often carries two things at once: the method — how to delegate work well — and the instance — this task, today's facts, and your current purpose. The reusable method can stay stable, while the instance should be refreshed for each job.

The optional instruction sheet separates those two jobs. The method lives one level up, in a document you can maintain. Each task prompt is then rebuilt from your current intent and, when needed, current evidence. You are maintaining a prompt-building method instead of rewriting the entire process every time.

There are other ways to reuse a method — skill files, plugins, or well-structured templates. We start with project instructions because you can inspect and change the rules yourself. A project may provide three useful shelves; the optional sheet in Kata 07 is only the first:

Think of a configured ChatGPT Project as a reusable briefing room. Feature availability varies by ChatGPT plan, workspace, and interface, so keep the method understandable even when a particular shelf is unavailable.

This guide is one technique inside a larger method. In the 5D development cycle, everything on this page lives in the Discussion phase — the phase where meta-prompting means the AI interviews you before it builds. On the belt ladder:

Nothing new to learn here — this is the wrap-up. The four moves you have already met, replayed as one continuous motion: chat → interview → reviewable context → task prompt. Watch a vague line become a runnable prompt, or click any stage to jump straight to it.