O Holy Spirit, Love of the Father and the Son, inspire me always with what I should think, what I should say, how I should say it, what I should keep silent, how I should act, and what I should do, for the glory of God, the good of souls and my own sanctification.
Holy Spirit, give me sharpness to understand, capacity to retain, method and ability to learn, subtlety to interpret, and grace and effectiveness to speak.
Grant me clarity at the beginning, direction as I progress and completeness at the end.Amen.
Learning to speak to AI.
"Write me something about how to bring up my child."
What you get: a generic text, with no focus, that could apply to any family or age. You spend 20 minutes rewriting it. The AI had no idea what you actually needed.
"I am the mother of a 10-year-old child preparing for his First Communion. Write me a 200-word explanation of what the Eucharist is, in simple language, with one everyday example and one question I can ask my child after reading it."
The more specific your question, the fewer corrections you will make afterwards.
As with any good conversation, technique matters. Six building blocks for a strong prompt:
Sets the scene. Gives the AI the background, the situation and the audience so it can tailor its response.
Who is this for? What is the situation? What does the reader already know?
Tell the AI who it should be. A specific persona shapes its expertise, vocabulary and perspective.
Assign profession and personality. Define communication style and level of expertise.
Define what to do and how. Give numbered steps so the AI follows a process rather than a vague instruction.
List specific tasks in order. Include decisions and conditions.
Specify the shape of the result. Structure, length and visual layout so you get exactly what you need.
State the document type, word count and sections. Use of bullet points, tables or headings.
Set the emotional register. The same content sounds different when the tone changes.
Describe the feeling and the level of formality. Define words or phrases to use or to avoid.
Build in quality control. The AI reviews its own output against your criteria before delivering it.
Give a checklist the AI must verify internally. If anything fails, it corrects itself before responding.
How to get the AI to think better.
What it does: ask the AI to reason step by step before giving you the final answer. Forces a logical breakdown.
When: calculations, logic problems, multi-step analysis, verification against criteria.
How:
"Think this through step by step before giving me your answer."
"Walk me through your reasoning, then give me your conclusion."
Tip: works best with models that do not reason internally (classic ChatGPT, basic Copilot). On models that already reason on their own (o1, Claude with thinking) the gain is smaller. Start here.
What it does: explores several reasoning paths in parallel, evaluates each one and chooses the strongest.
When: strategic decisions, comparing approaches, complex planning, choosing between alternatives.
How:
"Consider 3 different approaches. Evaluate the pros and cons of each, and recommend the best."
"Explore 3 solutions. Rank them by feasibility and risk."
Tip: produces better analysis than a single path. Essential for recommendations you will need to defend.
What it does: generates a response, then systematically verifies each claim. Two passes: create, then verify.
When: research, leadership reports, regulatory content, anything where accuracy is critical.
"Now review every claim in your answer. Flag anything you are not sure about and verify the key facts."
Tip: essential for catching hallucinations. Use it on any document you intend to share or present.
What it does: ask the AI to argue AGAINST its own answer to find weaknesses and counter-arguments.
When: risk assessment, proposals, important decisions, formal letters.
"Now play devil's advocate. What are the 3 strongest arguments against this recommendation?"
Tip: avoids biased analysis. Excellent for preparing difficult conversations: anticipate objections before they come up.
What it does: runs the same question several times and compares the answers. Reveals the AI's uncertainty.
When: high-impact decisions, regulatory topics, numerical estimates.
"Answer this question 3 times independently, then compare them and explain the differences."
What it does: ask the AI to write its own optimal prompt, then use that prompt for the final output.
When: new or ambiguous tasks, when you do not know how to structure the request, prompt optimisation.
"Create a master prompt for me to get the ideal result on this task. DO NOT INFER — ask me up to 3 critical questions so I can give you the right input."
Tip: surprisingly effective. The AI often instructs itself better than we do. Ideal for getting started.
When AI does not answer: it acts.
Six building blocks for an agent that acts on your behalf.
Who is this agent, always? Defines expertise, perspective and tone for ALL conversations, not just one.
Example: "You are a bilingual assistant who helps Latin American families in London with school letters, NHS correspondence and paperwork in English."
Why does it exist? A clear purpose stops it drifting from the job you asked it to do.
Example: "Your sole purpose is to translate letters from English to Spanish, summarise what matters, and suggest the next step to take."
Give it what it needs to know. Upload files, examples or data. The agent references them in every reply.
Upload: previous school letters, an example of a good summary, a list of English terms you do not understand.
Hard rules: ALWAYS / NEVER. Prevents common errors. If you do not set them, it will invent them.
Example: "Always reply in Spanish. Never make up medical facts. If you are not sure, say 'I do not know, please check with the person concerned'."
Numbered steps it follows every time. Forces a logical and consistent output rather than an improvised one.
Example: "When I give you a letter: 1. Identify the sender. 2. Summarise in 3 bullet points. 3. Flag dates or actions. 4. Suggest a reply if relevant."
Quality control before delivery. The agent verifies its own output against your criteria.
Example: "Before delivering, check: is it in Spanish? Are dates clear? Is there a next step? If anything fails, correct it and tell me."
Markdown, XML, JSON, Python.
Talking to AI is like sending a long WhatsApp message: if it all runs together, the meaning is lost. If you separate the question, the context and the format you want, the AI does not get confused.
Markdown, XML and JSON are three ways of marking the parts — like opening quotation marks, making a list, or setting a title.
As with any good conversation, you also need a bit of grammar.
| Symbol | Syntax | What it is for | How to use it |
|---|---|---|---|
| # | # Title | Main heading | # Risk assessment report |
| ## | ## Section | Section heading | ## Executive summary |
| ### | ### Subsection | Sub-heading | ### Key findings |
| ** | **text** | Bold / emphasis | Highlights critical risks |
| * | *text* | Italic / soft emphasis | Adds context notes |
| - | - item | Bulleted list | - Risk 1 - Risk 2 |
| 1. | 1. first | Numbered list | 1. Identify 2. Assess 3. Mitigate |
| ` | `code` | Inline code / term | Use the CRAFTS framework |
| > | > quote | Quotation or callout | > Key insight: the quality of the data… |
| | | | col | col | | Table structure | | Risk | Impact | Likelihood | |
| --- | --- | Horizontal divider | Separate sections with --- |
| ``` | ```language | Code / data block | ```json {...} ``` |
Use [BRACKETS] as placeholders to build templates. Define variables once, swap values easily.
[ROLE] Mother supporting catechesis
[AUDIENCE] My 9-year-old daughter
[TASK] Explain in simple language
[TOPIC] What the Eucharist is
[FORMAT] 200 words, 1 everyday example
[TONE] Warm, motherly
[LENGTH] 200 words, 1 closing question
[LIMITS] Basic Catholic doctrine onlyYou are a [ROLE] at [INSTITUTION].
## Task
[TASK] about [TOPIC] for [AUDIENCE].
## Constraints
- **Tone:** [TONE]
- **Length:** [LENGTH]
- **Scope:** [LIMITS]
- **Format:** [FORMAT]<role> · Defines who it should be<context> · Background information<task> · What it must do<constraints> · Limits<example> · What a good output looks like<input> · Data to process<output_format> · Desired structure<instructions> · Step-by-step guidance<role>Bilingual catechesis tutor</role>
<context>
My 9-year-old daughter is going
to make her First Communion.
</context>
<task>
Explain it to her in 200 words,
simple language
</task>
<constraints>
<tone>warm, motherly</tone>
<include>1 everyday example</include>
</constraints>{ } · Object: groups related data[ ] · Array: list of items"key": "value" · Text"key": 42 · Number"key": true · Boolean"key": null · Empty"key": [...] · Array"key": {...} · Nested object{
"task": "weekly_family_agenda",
"schema": {
"activity": "text",
"priority": "low|medium|high",
"estimated_time": "minutes",
"owner": "text",
"help_needed": "text"
},
"max_items": 10
}for x in list: · Iterates itemsif condition: · Conditional logicelif / else: · Alternative branchesdef func(): · Defines a stepreturn value · Function outputvariable = x · Stores a valuelist = [a, b] · Collectiondict = {k: v} · Key-value mappingfor task in family_agenda:
summary = explain(task,
max_words=100)
questions = extract_questions(summary)
for question in questions:
if question.difficulty == "high":
check_with_friar(question)
notify(partner)
elif question.difficulty == "medium":
explain_directly(question)20 minutes. Choose ONE option and work with your own real case.
We give you thanks, Holy Spirit, for what we have learnt today.
Grant us discernment to use these tools wisely, vocation to put them at the service of our families and our work, and mission to walk alongside our children in the world that is coming.Amen.