--- title: "The AI Prompt Bank for IELTS" category: metodo --- # Prompt Bank — Speaking and Writing Every prompt used across the preparation sessions, reconstructed and cleaned up. All of them are written to be copied and pasted straight into any conversational model (ChatGPT, Claude, Gemini, etc.). > **One warning before the list:** None of these prompts asks the model to write for you. Every one asks it to assess you, to generate an exercise, or to point at a pattern. That distinction is the only reason the set works at all. A prompt that produces someone else's text trains nothing. **A note on provenance:** several of the speaking prompts were improvised inside live mock sessions and were never recorded word for word. What follows is the functional reconstruction, not the literal transcript. Where the mechanism matters more than the wording, I say so. --- ## PART I — SPEAKING ### 1. The silent mock (the core prompt) This was the scaffold under every oral practice session. The mechanics matter as much as the text: the model does not interrupt, does not correct, does not comment. It stays quiet until you say the codeword. > You are going to act as an IELTS Speaking examiner. Give me the full test: Part 1 (four or five questions across two topics), Part 2 (a cue card with one minute of preparation and two minutes of response), and Part 3 (four or five abstract questions derived from the Part 2 topic). > > Strict rules during the test: > — Ask one question at a time and wait for my answer. > — Correct nothing, comment on nothing, give me no feedback. > — Do not tell me whether an answer was good. Just move to the next question. > — If my answer arrives truncated or with strange words, assume it is a microphone transcription failure, not my error. Ignore it. > > When I type the word **CLOTH**, drop the examiner role and give me the full feedback. **Why the codeword.** The model tends to break character and congratulate you after each answer. That destroys the exam condition, which is precisely what is being simulated. An arbitrary word, foreign to the exam's own vocabulary, is the only reliable switch. Pick your own. Keep it short, and make sure it could never surface inside a real IELTS answer. --- ### 2. Post-mock feedback > Drop the examiner role. Analyse every answer from this session and give me: > > 1. Grammatical errors grouped by **pattern**, not in order of appearance. I want to know which error repeats, not which one came first. > 2. How many times each pattern appeared. > 3. Incorrect collocations and prepositions, with the correct version beside each. > 4. Words I overused, with exact frequency counts. > 5. A band estimate per criterion: Fluency and Coherence, Lexical Resource, Grammatical Range and Accuracy. > > Do not correct pronunciation. Do not encourage me. Keep every correction short: the error, the correct version, nothing else. --- ### 3. Error density count The prompt that turns practice into measurement. Without it, the feeling of improvement cannot be told apart from familiarity with the topic. > Count the morphological errors in this answer and divide by the word count. Count only: subject-verb agreement, wrong tense, dropped plural, missing or intrusive article. Ignore vocabulary, structure and content. Give me the raw error count per sample. Research on spoken error analysis suggests the boundary between Band 7 and Band 8 is more strongly correlated with a reduction in morphological error density than with lexical sophistication, though individual marking will always vary. Five slips per answer is the threshold worth chasing. --- ### 4. Self-transcription and task repetition The best-evidenced protocol in the whole set. You record, you transcribe yourself, you mark your own errors, and you **repeat the same task**. > I am pasting my own transcription of a Part 2 answer that I recorded and transcribed myself. I have already marked the errors I spotted. Tell me: > > 1. Which errors I marked correctly. > 2. Which errors I missed. > 3. What I flagged as an error that was not one. > > Then give me the same cue card again so I can repeat it. Do not give me a model version — I want to produce it again myself. Transcribing yourself is the step that does the work. Receiving a transcript made by someone else does not produce the same effect. --- ### 5. Cue cards with a grammatical target > Generate an IELTS Speaking Part 2 cue card on the topic: [TOPIC]. Give me one minute to prepare. After my answer, review **verb tenses only**: tell me at which point I slipped from past into present without justification. Comment on nothing else. Variants used, changing the review target: — third person singular agreement — second conditional (*if I had*, never *if I will have*) — fixed prepositions — articles with abstract nouns One target per round. Two simultaneous targets dilute attention and you catch neither. --- ### 6. Part 3 with structural critique > Ask me three IELTS Speaking Part 3 questions on [TOPIC]. After each of my answers, assess whether I followed the AEEC structure: Answer, Expand, Example, Conclusion. Tell me which step I cut short. Also check whether I used conditionals correctly and whether I conceded before stating my position. --- ### 7. Agreement drill with abstract subjects > Give me ten incomplete sentences whose subject is an abstract noun (evidence, research, trust, politics, data, the number of, the majority of, staff, team). Ask me to complete the verb. After I answer, correct me and explain the mechanism, not the rule. This drill targets the most expensive error in the profile: the singular abstract subject that the brain reads as plural under load. --- ### 8. Emergency frame for impossible questions > Give me ten Part 3 questions on topics I know nothing about (agricultural policy, sport, fashion, the car industry). After each one, assess whether I managed to answer without falling into silence, using the *it depends* frame: acknowledge two scenarios, pick one, give an example. --- ## PART II — WRITING ### 9. Examiner correction with band estimate The base prompt for any essay. It works for Task 1 and Task 2 by swapping the criteria. > Act as an IELTS Academic examiner. Mark this text against the four official criteria: Task Response, Coherence and Cohesion, Lexical Resource, Grammatical Range and Accuracy. Give me an estimated band per criterion and an overall band. Then rewrite the text at Band 8 and **state explicitly what changed and why**, sentence by sentence. Do not give me generic advice of the "use more varied vocabulary" kind: tell me which specific word you replaced and with what. > > [PASTE YOUR TEXT] The last instruction is what makes the difference. Without it the model returns categories of improvement instead of improvements. --- ### 10. Correction against your closed error set After three or four marked essays you already know which faults you repeat. From that point on, generic correction stops earning its keep. > Mark this essay by checking **exclusively** against this list of my errors. Ignore everything else: > > 1. Subject-verb agreement with singular abstract subjects > 2. Comma splices > 3. Tense collapse within a single paragraph > 4. Prepositions: different from, arrive at/in, based on, depend on, responsible for > 5. this/these, spend/spent, base form after modals, a/an, amount/number > 6. Stray comma between subject and verb > 7. Full stop after a sentence-initial transition word > > Give me a count of how many times each one appeared. I want the frequency table, not the explanation. > > [PASTE YOUR TEXT] The frequency table is the instrument. An error that drops from five to one over three weeks is evidence; "I think I'm writing better" is not. --- ### 11. Separating production from revision The single habit change that paid better than all the grammatical content put together. > I am going to write a timed Task 2 in 40 minutes, in one pass, without correcting as I write. When I finish, I will paste it to you **unrevised**. Mark it. Then I will paste my own revision pass, done before seeing your corrections. Compare the two: tell me which errors I caught on my own and which ones escaped me. Correcting while producing splits attention between generating the idea and policing the form. Both are lost. --- ### 12. Reverse-order proofreading > Give me instructions for revising my essay by reading the sentences from last to first. Explain why that order catches mechanical errors that normal reading does not. The mechanism: reading in natural order engages semantic processing, which automatically fills in whatever is missing. Reverse reading breaks the narrative thread and leaves the form exposed. --- ### 13. Collocation extraction from your own text > Here is my Task 2 essay. Extract five collocations I used (verb + noun or adjective + noun) and give me three alternatives for each. Flag which of my original collocations are wrong or read as a literal translation from Spanish. The second instruction is the valuable one. Cross-linguistic interference traps do not announce themselves. --- ### 14. Generating Task 1 tasks > Describe two maps of a coastal town in 1990 and 2020, with at least five changes between them. After I write my report, review **only**: passive voice, comparatives, and prepositions of place. Variants: a line graph across two periods, a table with four categories, a process diagram with six stages, a grouped bar chart. > Describe a six-stage process diagram about [TOPIC]. After my report, review only temporal sequencing and passive voice. --- ### 15. Planning with 5W + H > I am going to write a Task 2 essay on [PROMPT]. Before I write, ask me the 5W + H questions I need to answer to build the plan: who, what, when, where, why, how. Help me identify the prompt type (opinion, discussion, advantage/disadvantage, problem/solution, two-part) and tell me which paragraph structure that specific type demands. One caution: the *discussion* type requires symmetric presentation of both views, with no verdict until the conclusion. The concession-then-position structure belongs only to the balanced *opinion* type. Confusing the two costs Task Response. --- ### 16. Dangling participle check > Find every sentence in my essay that opens with a participle, gerund or infinitive. For each one, run this test: who performs the action of the participle? Does that subject match the grammatical subject of the main clause? If they do not match, rewrite it. This error costs more than a dropped *-s*, because it changes the literal meaning of the sentence. It gets checked at the end, never during writing. --- ### 17. Band 7 against Band 8 > Take my essay. Generate two versions: one at Band 7 and one at Band 8. Put them in parallel columns and bold **only** what changes. Do not rewrite the content, do not change the examples, do not touch the structure. I only want to see the formal difference between the two bands. Seeing the delta in isolation teaches more than reading a complete flawless model. --- ### 18. Cumulative error log > Here are my last five marked essays. Build a table with error type in the rows, essay number in the columns, and frequency in the cells. Tell me which error is falling, which is holding steady, and which got worse. The prompt that closes the loop. Without it, each correction is judged in isolation and you never learn whether the whole system is working. --- ## PART III — CROSS-CUTTING ### 19. Initial diagnosis > I am going to paste three samples: one transcribed Speaking answer and two essays. Without correcting them yet, tell me my **root error**: the process habit the others derive from. Distinguish between knowledge errors (I do not know the rule) and execution errors (I know the rule and it fails under pressure). Rank them by how much band they cost me, not by frequency. The line between not knowing and not executing decides the entire study strategy. Knowledge errors are studied; execution errors only yield to volume under pressure. --- ### 20. Detecting penalised templates > Review my text and flag any phrase that reads as a memorised template. Prefabricated phrases are penalised under Lexical Resource. Flag in particular generic openers of the "In recent years, X has become a matter of significant concern" kind. Working rule: if your first sentence could appear in an essay on any topic by swapping a single noun, it describes context instead of arguing. Delete it. --- ## Comparing the two banks | Criterion | Speaking | Writing | |:---|:---|:---| | **Core prompt** | Silent mock with codeword | Examiner marking with band per criterion | | **Model's role** | Examiner who never interrupts | Marker who justifies every change | | **Dominant error targeted** | Tense collapse under cognitive load | Agreement with abstract subjects; comma splices | | **Unit of measurement** | Morphological slips per answer | Frequency by error type per essay | | **Supporting evidence** | Task repetition and self-transcription | Corrective feedback on a closed error set | | **Main risk** | Model breaks character and congratulates | Model returns improvement categories, not edits | | **Instruction that saves it** | The arbitrary codeword | "Tell me which word you replaced and with what" | | **Suggested frequency** | Daily, short sessions | One full essay every two days | ### Verdict The Speaking bank serves the candidate who loses band under pressure despite knowing the rule; the Writing bank serves the one who repeats a narrow set of mechanical faults that survive revision. Both converge on the same design instruction: the model measures and points, and never produces the text you will then try to imitate. Anyone starting out gains more from the initial diagnosis (prompt 19) than from any mock, because without a closed error set the correction scatters and neither track pays off.