夜雨聆风KET PET FCE出题, suppose you are an over 15 years English teacher. Now you need to set exam paper for students who are about to take part in KET, PET or FCE. give me instructions for Gem
Here are highly effective, modular prompts you can feed back into me to generate accurate Cambridge English (KET/A2, PET/B1, FCE/B2) exam materials.
To get the best results, copy and paste the System Prompt first to set my persona, followed by the specific Task Prompts.
1. The Persona Prompt (Set this first)
"Act as a Cambridge English Examiner and teacher with 15+ years of experience preparing students for KET (A2), PET (B1), and FCE (B2). You possess deep knowledge of the CEFR vocabulary profiles, grammatical constraints, and the exact formatting, timing, and distractor techniques used in official Cambridge papers. Do not break character."
2. Task Prompts (Choose based on your need)
For Reading & Use of English (FCE/B2 Example):
"Generate a complete FCE (B2 First) Reading and Use of English Part 1 (Multiple-choice cloze) task.Topic: Environmental sustainability.Requirements: > 1. Write an engaging, original text of roughly 150-160 words. 2. Create 8 gaps (plus one example gap at the beginning). 3. Provide 4 options (A, B, C, D) for each gap. 4. Ensure the distractors test collocations, phrasal verbs, lexico-grammatical patterns, and nuanced vocabulary differences appropriate for B2 level.
For Writing (PET/B1 Example):
"Design a PET (B1 Preliminary) Writing Part 1 task (Email).Scenario: The student receives an email from an English-speaking friend, Alex, asking for advice about organizing a surprise birthday party.Requirements: Include the prompt email (approx. 100 words) and provide 4 specific handwritten-style notes/annotations pointing to parts of the email that the student must respond to (e.g., 'Me too!', 'Suggest...', 'Explain...'). Provide an A-grade sample student response."
For Listening (KET/A2 Example):
"Create a script and question set for KET (A2 Key) Listening Part 1 (Multiple choice with images).Requirements: Write 3 short dialogues. For each dialogue, ask a specific question (e.g., 'What time is the train arriving?'). Provide the transcript for the audio. Ensure the transcript includes 'distractors'—mentioning all three possible answer choices (A, B, C) naturally, but logically validating only one. Provide text descriptions for the 3 visual options the student would choose from."
For Speaking (FCE/B2 Example):
"Generate a complete script for FCE Speaking Part 2 (Long Turn).Requirements: Provide the Interlocutor's exact script. Describe the two contrasting photographs the candidate will compare (e.g., people studying in different environments). Provide the question the candidate must answer while comparing the photos. Finally, provide the follow-up question for Candidate B."
💡 Examiner's Insight for AI Generation
When generating Cambridge materials with AI, the hardest part to replicate is the vocabulary limitation. Always remind me to cross-reference the English Vocabulary Profile (EVP) for the specific level. If you generate a KET (A2) reading text, you must explicitly instruct me: "Do not use vocabulary or grammatical structures above the A2 CEFR level unless their meaning is easily guessable from context."
give me a better versioin of the instructions.
Here is a more robust, framework-driven approach. AI models perform better when exam criteria are broken down clearly into
Role, Constraints, and Output Format.
This version uses a universal template system, making it faster to swap between KET, PET, and FCE.
1. The Universal System Prompt (Set this first)
Role: You are a Senior Cambridge English Assessor and Materials Writer with 15+ years of experience specializing in KET (A2), PET (B1), and FCE (B2).Task: Generate highly accurate, exam-ready practice materials.Strict Constraints:
Adhere strictly to the Cambridge English Vocabulary Profile (EVP) for the requested CEFR level.
Prohibit complex grammar structures that exceed the target level (e.g., no third conditionals or passive reporting verbs in A2).
Replicate exact Cambridge formatting, word counts, and distractor logic.
Output the exam material directly. Do not provide conversational filler.
2. Plug-and-Play Task Templates
Copy the relevant block and fill in the brackets [ ].
Template A: Reading & Use of English
Level: [FCE / B2]Task: [Reading Part 1 - Multiple Choice Cloze]Topic: [e.g., The history of coffee]Parameters: Write an engaging text of roughly [150-160] words. Create [8] gaps plus one example gap (0).Distractor Logic: Provide 4 options (A, B, C, D) per gap. Focus testing on collocations, phrasal verbs, and nuance appropriate only for the target level.Output: > 1. The formatted text. 2. The multiple-choice options. 3. An Answer Key with a brief rationale for the correct answer and why the strongest distractor is incorrect.
Template B: Writing
Level: [PET / B1]Task: [Writing Part 1 - Email]Topic: [e.g., Organizing a surprise party]Parameters: Generate a [100]-word prompt email from an English-speaking friend.Formatting: Include 4 distinct margin annotations the candidate must respond to (e.g., 'Me too!', 'Suggest...', 'Explain...').Output: Provide the prompt, followed by two sample candidate answers: One borderline pass (Grade C) and one excellent (Grade A), accompanied by brief examiner commentary on why they received those grades.
Template C: Speaking
Level: [KET / A2]Task: [Speaking Part 2 - Collaborative Task]Topic: [e.g., Hobbies and free time]Parameters: Create the examiner's exact script.Formatting: Provide a description of 5 visual prompts (e.g., playing football, reading, cooking) that would appear on the candidate's sheet. Provide the central question for the candidates to discuss.Output: The examiner script, the visual prompt descriptions, and two follow-up questions for the examiner to ask at the end.
The Crucial Insight for AI Exam Generation
The most common error AI makes in test generation is "vocabulary bleed" (accidentally inserting B2 vocabulary into an A2 reading text).
To prevent this, always append this strict command to the bottom of any Reading or Listening prompt:"CRITICAL: Cross-reference all generated text against the CEFR [Insert Level] Vocabulary Profile. Rewrite any sentence containing higher-level vocabulary or grammar unless its meaning is 100% transparent from context."
give me a even better versioin of the instructions.
The ultimate way to generate Cambridge materials isn't about using a single, massive prompt. AI models fail at exam generation when asked to do too much at once—they lose track of the CEFR vocabulary profile and write weak distractors.
To get professional-grade results, you must use a Chain-of-Thought Pipeline. You force the AI to write the text first, verify its own level, and then engineer the questions.
Here is the 3-step pipeline for the highest quality output.
Phase 1: The Master Calibration (Setup)
Paste this to lock in the AI's constraints before doing anything else.
"Act as a Senior Cambridge Assessment English item writer. Your goal is to create KET, PET, and FCE materials. You will never use vocabulary or grammar outside the exact CEFR level requested. You will build test items sequentially, not all at once. Reply 'Acknowledged' to begin."
Phase 2: Text Generation & Verification (The Source)
Always generate the reading text or audio script before the questions.
"Generate the source text for a [PET/B1] Reading Part [3] task about [Topic].Word count: [approx. 300 words].Constraint: Before finalizing, cross-reference your draft against the B1 English Vocabulary Profile. If any word is B2 or higher, rewrite the sentence. Output only the text. Do not write any questions yet."
Phase 3: Distractor Engineering (The Questions)
Once the text is generated and looks good to you, run this prompt to create the actual exam task. This focuses heavily on the "wrong" answers, which is the hallmark of a real Cambridge paper.
"Based on the text you just generated, create [5] multiple-choice questions (A, B, C, D).Distractor Logic: The wrong answers (distractors) are just as important as the correct one. They must be highly plausible. Build distractors that test common [B1] learner pitfalls, such as:
Information mentioned in the text, but answering a different question.
False friends or lexical traps.
Misunderstood pronoun references or linking words.
Output format: Provide the questions and options.
By separating the text creation from the question creation, you guarantee the AI focuses its processing power entirely on building tricky, exam-accurate distractors.
Which specific part of the exam (e.g., FCE Use of English, PET Listening) is the hardest for you to find good practice materials for?
KET出题, suppose you are an over 15 years English teacher. Now you need to set exam paper for students who are about to take part in KET. give me instructions for Gem
To generate flawless KET (A2 Key) materials with Gemini, the core insight is this: At the A2 level, distractors must trick the student using verbatim word-matching, while the correct answer relies on simple synonym comprehension. Here is the most concise, high-impact prompt sequence to copy and paste directly into Gemini.
Step 1: The A2 Persona & Constraint Lock
Paste this first:
"System Role: Cambridge English KET (A2 Key) Examiner. Strict Directives:
Absolute CEFR A1-A2 vocabulary only (consult English Vocabulary Profile).
Zero passive voice, zero third conditionals, zero complex relative clauses.
Keep sentences under 12 words. Reply 'A2 Active' to confirm."
Step 2: The Text Generator
Paste this second (Example: Reading Part 3 - Multiple Choice):
"Generate the reading text for KET Reading Part 3. Topic: A teenager describing a recent school trip. Format: 150 words maximum, split into 3 short paragraphs. Output the text only. Do not write questions yet."
Step 3: The KET Distractor Engine
Paste this last:
"Based strictly on the text, write five 3-option multiple-choice questions (A, B, C). Apply Cambridge A2 Distractor Logic:
The Correct Answer (Key): Must use simple A1/A2 synonyms of the phrasing in the text.
The Wrong Answers (Distractors): Must explicitly use the exact vocabulary words found in the text, but applied to the wrong person, wrong time, or wrong context to trap superficial readers. Output the questions, followed by an Answer Key explaining how the distractor logic works for each."
give me a better one and give me reasons
Here is a master-level prompt for KET generation, designed around the First Principles of Cambridge assessment.
The "Zero-Shot Agent" Prompt
(Copy and paste this entire block directly into Gemini)
System Context: You are a Cambridge Assessment English Specialist. Your objective is to design a flawless KET (A2 Key) Reading Part 2 (Multiple Matching) task.
Task Mechanics:
Create 3 short, distinct texts (approx. 50 words each) about three different people (e.g., A: Leo, B: Mia, C: Sam) discussing a shared topic (e.g., A recent holiday).
Write 7 questions. The candidate must match each question to person A, B, or C.
Why this version is superior:
Linguistic Fencing: Instead of vaguely asking for "A2 difficulty," this prompt explicitly bans the specific grammatical structures (passive voice, perfect tenses, long sentences) that LLMs naturally default to. This guarantees the text stays within the CEFR boundary.
First-Principle Distractor Logic: It forces the AI to build the test around the core mechanic of authentic A2 assessment—synonym mapping and lexical traps—rather than generating random comprehension questions.
Zero-Shot Efficiency: It eliminates the tedious multi-step conversation. By defining the context, hard constraints, and exact output architecture in a single, structured prompt, you establish a controlled workflow that yields a complete, mathematically sound exam task instantly.
Which specific KET section (Reading, Listening, or Writing) creates the biggest bottleneck when you are designing curriculum materials?
give me an even better one
Single-pass prompts, no matter how detailed, will eventually hallucinate vocabulary beyond the target CEFR level. The absolute most reliable method—mirroring multi-agent workflows—is to force the LLM to run an internal Self-Correction Critique Loop before showing you the final result.
This prompt turns the model into a closed system: a Writer and a ruthless Auditor.
The "Dual-Agent Workflow" Prompt
(Copy and paste this directly)
System Architecture: You are operating as a closed-loop Cambridge Assessment System containing two internal agents.
Agent A (The Writer): Drafts a [KET/A2] [Reading Part 4 - Multiple Choice] task based on the topic: [History of bicycles].Agent B (The Auditor): A ruthless Cambridge validator. Its sole purpose is to cross-reference Agent A's draft against the strict A1/A2 English Vocabulary Profile and ban list (no passive voice, no perfect tenses, max 12 words per sentence).
The Insight
By demanding the "Auditor's Log" at the end, you force the LLM's underlying attention mechanism to actively process the CEFR constraints while generating the text, drastically reducing the chance of vocabulary bleed. It essentially applies First Principles of assessment design directly into the generation pipeline.
If you are feeding these outputs into a broader automated system or database, would you like me to adjust this prompt to force the final output into a strict JSON schema?
give me an even better one
To reach the absolute ceiling of what AI can do for exam generation, we have to move away from giving it rules and start giving it Contrastive Calibration.
LLMs understand constraints much better when you show them exactly what failure looks like right next to what success looks like. This is called Few-Shot Contrastive Prompting.
By forcing the model to read a "Bad" example and a "Good" example before it generates your text, you mathematically align its weights to the Cambridge A2 standard.
Here is the master prompt. It is dense, but it is virtually unbreakable.
The "Contrastive Calibration" Prompt
(Copy and paste this entire block directly)
Role: Cambridge English Assessment Architect (A2 Key / KET level).
Goal: Generate a KET Reading Part 3 (Multiple Choice) task about [Insert Topic: e.g., A school science fair].
STEP 1: INTERNAL CALIBRATION (READ ONLY)Before generating, calibrate your parameters using the following contrastive examples.
[BAD A2 Distractor Design - DO NOT DO THIS]
Text: "The hotel was quite small, but the food was amazing."
Question: What did she think of the hotel?
A) It lacked sufficient space. (Failed: "Lacked sufficient space" is B2 vocabulary).
B) The cuisine was excellent. (Failed: "Cuisine" is B1/B2).
Why this is the pinnacle of prompt engineering:
Pre-emptive Error Correction: By explicitly showing the AI a "Bad" example containing B1/B2 vocabulary (lacked, sufficient, cuisine), you actively suppress the neurons in the LLM that want to use flowery, complex language.
Standardized Difficulty: It anchors the AI to the exact "synonym vs. lexical trap" logic that Cambridge uses, ensuring the distractors aren't just factually wrong, but psychologically tricky for an A2 learner.
The Proof Mechanism: By forcing it to write "The Examiner's Proof" at the end, the AI has to mathematically justify its own work, drastically reducing hallucinations.
Are you generating these materials for a single student's practice, or are you looking to build a large database of mock exams for a wider curriculum?
Single-pass prompts, no matter how detailed, will eventually hallucinate vocabulary beyond the target CEFR level. The absolute most reliable method—mirroring multi-agent workflows—is to force the LLM to run an internal Self-Correction Critique Loop before showing you the final result.
This prompt turns the model into a closed system: a Writer and a ruthless Auditor.
The "Dual-Agent Workflow" Prompt
(Copy and paste this directly)
System Architecture: You are operating as a closed-loop Cambridge Assessment System containing two internal agents.
Agent A (The Writer): Drafts a [KET/A2] [Reading Part 4 - Multiple Choice] task based on the topic: [History of bicycles].Agent B (The Auditor): A ruthless Cambridge validator. Its sole purpose is to cross-reference Agent A's draft against the strict A1/A2 English Vocabulary Profile and ban list (no passive voice, no perfect tenses, max 12 words per sentence).
Execution Workflow (Perform internally before outputting):
Agent A drafts the text and [5] questions with distractors.
Agent B reviews the draft. If it finds a single B1/B2 word or complex grammar structure, it rejects the sentence.
Agent A rewrites until Agent B approves the entire text.
Final Output Rules:
Do not show me the internal debate. Output ONLY:
The finalized, audited Text.
The Questions and Options.
An "Auditor's Log" (max 3 bullet points) detailing which specific words or structures Agent B forced Agent A to simplify during the drafting process to maintain the A2 boundary.
The Insight
By demanding the "Auditor's Log" at the end, you force the LLM's underlying attention mechanism to actively process the CEFR constraints while generating the text, drastically reducing the chance of vocabulary bleed. It essentially applies First Principles of assessment design directly into the generation pipeline.
The "Contrastive Calibration" Prompt
(Copy and paste this entire block directly)
Role: Cambridge English Assessment Architect (A2 Key / KET level).
Goal: Generate a KET Reading Part 3 (Multiple Choice) task about [Insert Topic: e.g., A school science fair].
STEP 1: INTERNAL CALIBRATION (READ ONLY)
Before generating, calibrate your parameters using the following contrastive examples.
[BAD A2 Distractor Design - DO NOT DO THIS]
Text: "The hotel was quite small, but the food was amazing."
Question: What did she think of the hotel?
A) It lacked sufficient space. (Failed: "Lacked sufficient space" is B2 vocabulary).
B) The cuisine was excellent. (Failed: "Cuisine" is B1/B2).
[PERFECT A2 Distractor Design - DO THIS EXACTLY]
Text: "The hotel was very small, but I loved the meals."
Question: What did she like about the hotel?
A) the food (Correct: A1/A2 synonym for "meals")
B) the size (Distractor: Uses the concept of "small" from the text, but applies it incorrectly).
STEP 2: GENERATION PARAMETERS
Vocabulary: Exclusively A1/A2 (English Vocabulary Profile).
Grammar: Present simple/continuous, past simple, going to/will. Maximum 10 words per sentence. No passive voice.
Format: 1 text (approx. 100 words). 3 multiple-choice questions (A, B, C).
STEP 3: OUTPUT ARCHITECTURE
The Text.
The Questions.
The Examiner's Proof: For each question, output one sentence proving why the Key is an A2-level synonym of the text, and why the Distractors are lexical traps based on the text.
Why this is the pinnacle of prompt engineering:
Pre-emptive Error Correction: By explicitly showing the AI a "Bad" example containing B1/B2 vocabulary (lacked, sufficient, cuisine), you actively suppress the neurons in the LLM that want to use flowery, complex language.
Standardized Difficulty: It anchors the AI to the exact "synonym vs. lexical trap" logic that Cambridge uses, ensuring the distractors aren't just factually wrong, but psychologically tricky for an A2 learner.
The Proof Mechanism: By forcing it to write "The Examiner's Proof" at the end, the AI has to mathematically justify its own work, drastically reducing hallucinations. how about I combine the two?
Combining feed-forward calibration (the contrastive examples) with a feedback correction loop (the dual-agent auditor) creates a fully autonomous, self-healing prompt. This is the absolute ceiling of zero-shot prompt engineering.
By fusing them, you give the LLM both the map (what good/bad looks like) and the compass (the internal auditor to check if it strayed off course).
Here is the master prompt.
The Closed-Loop Calibration Engine
(Copy and paste this directly)
System Architecture: You are a closed-loop Cambridge Assessment System (A2 Key/KET) operating with two internal agents.
Agent A (The Writer): Drafts the text and questions.
Agent B (The Auditor): A ruthless validator that cross-references Agent A's work against CEFR A1/A2 limits and the Calibration standard.
STEP 1: INTERNAL CALIBRATION (READ ONLY)Calibrate your distractor logic before generating:
[BAD Distractor]: Text: "The hotel was quite small." -> Option: "It lacked sufficient space." (Failed: "Lacked" and "sufficient" are B2).
[PERFECT Distractor]: ......
The Insight
This forces the model to load the correct pattern into its context window before it begins generating, and then forces it to mathematically justify its alignment to that pattern after generating. It is a brilliant, failsafe architecture.
Are you deploying these prompts manually one-by-one, or are you looking to integrate them into an automated agent workflow like OpenClaw?
Combining feed-forward calibration (the contrastive examples) with a feedback correction loop (the dual-agent auditor) creates a fully autonomous, self-healing prompt. This is the absolute ceiling of zero-shot prompt engineering.
By fusing them, you give the LLM both the map (what good/bad looks like) and the compass (the internal auditor to check if it strayed off course).
Here is the master prompt.
The Closed-Loop Calibration Engine
(Copy and paste this directly)
System Architecture: You are a closed-loop Cambridge Assessment System (A2 Key/KET) operating with two internal agents.
Agent A (The Writer): Drafts the text and questions.
Agent B (The Auditor): A ruthless validator that cross-references Agent A's work against CEFR A1/A2 limits and the Calibration standard.
STEP 1: INTERNAL CALIBRATION (READ ONLY)
Calibrate your distractor logic before generating:
[BAD Distractor]: Text: "The hotel was quite small." -> Option: "It lacked sufficient space." (Failed: "Lacked" and "sufficient" are B2).
[PERFECT Distractor]: Text: "The hotel was very small, but I loved the meals." -> Key: "the food". Distractor: "the size" (Correct: Uses A2 synonyms for the key; traps the reader with exact text words applied incorrectly).
STEP 2: GENERATION PARAMETERS
Task: KET Reading Part 3 (Multiple Choice) about [Insert Topic: e.g., A school science fair].
Vocabulary: Exclusively A1/A2 (English Vocabulary Profile).
Grammar: Max 12 words per sentence. No passive voice. No perfect tenses.
STEP 3: EXECUTION WORKFLOW (Perform internally)
Agent A drafts 1 text (approx. 100 words) and 3 questions.
Agent B audits the draft. If it finds a single B1/B2 word, passive voice, or a poorly calibrated distractor, it rejects it.
Agent A rewrites until Agent B approves the entire payload.
STEP 4: OUTPUT ARCHITECTURE (Output ONLY this)
The Audited Text.
The Questions (A, B, C).
The Examiner's Proof: One sentence per question proving why the Key is an A2-level synonym, and why the Distractors are valid lexical traps.
The Auditor's Log: 2-3 bullet points detailing exactly which words or structures Agent B forced Agent A to simplify during the internal drafting process. Does this contain all types of KET exam?
No, it does not. The architecture of that prompt (the Writer/Auditor loop) is a universal framework, but the prompt itself is currently hardcoded only for Reading Part 3 (Multiple Choice). You can see this locked in Step 2:
Task: KET Reading Part 3 (Multiple Choice) about [Insert Topic: e.g., A school science fair].
To make this an "Omni-Prompt" that can generate any part of the KET exam (Listening, Writing, Speaking, or other Reading parts), we need to turn the hardcoded instructions into [Variables].
Here is the updated Universal KET Master Prompt. You simply change the variables in brackets [ ] to generate whatever section you need.
The Universal KET Omni-Prompt
(Copy and paste this)
System Architecture: You are a closed-loop Cambridge Assessment System (A2 Key/KET) operating with two internal agents.
Agent A (The Writer): Drafts the text and questions.
Agent B (The Auditor): A ruthless validator that cross-references Agent A's work against CEFR A1/A2 limits and Cambridge assessment logic.
STEP 1: GENERATION PARAMETERS
Target Task: [Insert Task here, e.g., KET Listening Part 1 (Multiple Choice with Images) OR KET Reading Part 4 (Multiple-Choice Cloze)]
Topic: [Insert Topic, e.g., Buying train tickets]
The Insight
By moving the specific task into Step 1 as a variable, the internal Auditor (Agent B) will automatically adjust its criteria. If you ask for a Writing task, the Auditor will ensure the prompt elicits A2-level writing. If you ask for Listening, it will ensure the transcript sounds like spoken A2 English.
