TL;DR: Church AI prompts earn their place in the admin pile: newsletters, rotas, event copy, grant drafts, translated notices, minutes. They should not write sermons, devotionals or anything pastoral. Models invent scripture references with total confidence, so verify every citation against the text itself before it is printed or read aloud.
Most faith organisations run their communications on donated evenings. Somebody's mother-in-law does the newsletter. A retired teacher keeps the rota in a spreadsheet nobody else can open. The grant application is due Friday and nobody has written one before. There is no comms budget, there is rarely a comms person, and the work still has to go out every single week.
That is the gap these prompts are for. Everything below is administrative and communications work, because that is where a language model is genuinely useful to a congregation and where the risk of getting it wrong is a typo rather than a crisis of trust. The parts of the work that are actually ministry stay with the people doing the ministry, and the section on that line comes before the templates rather than after, because it matters more than the templates do.
What can AI actually do for a faith organisation?
It can produce first drafts of things that already have a shape. Notices, event announcements, volunteer role descriptions, thank-you notes, grant answers, plain-language rewrites, translations, meeting summaries. Work where the facts come from you and the model supplies the arrangement.
It cannot do the parts that depend on knowing your community. It does not know that the Tuesday group moved rooms, that the family in the third row lost someone in March, or that the phrase you used last year landed badly. Every draft below assumes a person who knows those things reads it before it goes anywhere.
The practical test is simple. If the work is "turn these facts I already have into readable prose," a model helps. If the work is "decide what we should say," it does not.
| The work | Give it to a model? | What still has to happen |
|---|---|---|
| Weekly notices, event copy, social posts | Yes, as a first draft | Somebody who knows the diary checks every date and name |
| Volunteer role descriptions, handover notes | Yes | The person doing the role reads it and corrects it |
| Rota conflict checking | Yes, with initials instead of names | You build the actual rota; the model only flags clashes |
| Grant drafts, funding reports | Yes, section by section | Every number comes from your accounts, not the model |
| Plain-language rewrites, translation | Yes, as a draft | A fluent speaker or the affected reader approves it |
| Sermons, homilies, devotionals, teaching | No, not as finished work | The delivered word is the speaker's, start to finish |
| Pastoral care, safeguarding, anything about a named vulnerable person | No | Your policy, your people, your regulator |
Where does AI stop and ministry begin?
Do not generate sermons, homilies, devotionals or teaching material as finished work. Not because a model cannot produce fluent, plausible, even moving text on a religious theme. It can, easily, in seconds. The reason is that the person standing up to deliver it is accountable for every word, and a congregation can tell.
They can tell in a way that is hard to describe and impossible to argue with. Something in the rhythm is wrong. The examples are slightly too neat. The illustration does not quite belong to the person telling it. People who listen to the same voice every week have a very fine ear for when that voice changes, and the credibility you spend there is not cheap to earn back.
If you want a model in the preparation, that is a different question, and a reasonable one. Structuring an outline you then write yourself is preparation. Listing the questions a passage raises is preparation. Producing candidate illustrations that you go and verify is preparation. Handing over the outline, the argument and the words is not preparation, it is substitution. Two bounded prompts for that work are at the end of the template list, written so the model returns questions and candidates rather than prose.
Why do models invent scripture references?
Because a language model predicts plausible next text rather than looking anything up. A reference is a short, highly patterned string, which makes it exactly the kind of thing a model produces confidently and wrongly: the right-sounding book, a chapter that is close, a verse that says something else entirely, or a quotation assembled from the shape of the thing rather than the thing. This is the ordinary hallucination failure mode, and why models make things up covers the mechanism in full.
In most domains a wrong citation is embarrassing. Here it is worse. A misquoted verse read from the front, or a fabricated historical detail in a printed history of the building, is the kind of error that people remember and repeat. The same applies to attributed quotations, dates, and anything about a named historical figure.
The fix is procedural, not clever. Every reference and every quotation gets checked against the actual text before it goes out. Not skimmed, opened. If a model gives you a reference you cannot immediately verify, delete it rather than shipping it with a mental note. And instruct the model up front to tell you when it is unsure, which does not make it reliable but does surface some of the weakest claims for you.
Before you finish, list every factual claim in your answer that I should verify independently. Include every scripture reference, every quotation, every date, every name, and every historical detail. For each one, tell me exactly where I would go to check it. Mark anything you are reconstructing from memory rather than from something you can point me to.
What should never go into a consumer AI tool?
Congregation lists. Pastoral notes. Safeguarding records. Giving records tied to names. Anything about a named person's health, family situation, immigration status or finances. Contact details in bulk.
The reason is not that vendors are careless. It is that consumer surfaces and business surfaces are governed by different documents with different defaults, those defaults change, and the person pasting the spreadsheet at eleven at night is not the person who read the terms. Rather than characterise any vendor's posture second-hand, read their own pages. Three worth having open:
- Google, Gemini Apps Privacy Hub (support.google.com/gemini/answer/13594961, accessed 27 August 2026). It states directly: "Please don't enter confidential information that you wouldn't want a reviewer to see or Google to use to improve our services, including machine-learning technologies." The same page notes that deletion is not always total: "Chats reviewed by human reviewers (and related data like your language, device type, location info, or feedback) are not deleted when you delete your activity. Instead, they are retained for up to three years."
- OpenAI, privacy policy (openai.com/policies/privacy-policy, accessed 27 August 2026; the page is dated 28 April 2026). Deletion has a similar caveat: content is removed within 30 days "or it has already been de-identified and disassociated from your account when you allow us to use your Content to improve our models".
- Anthropic, privacy policy (anthropic.com/legal/privacy, accessed 27 August 2026), whose retention section states that "Anthropic retains your personal data for as long as reasonably necessary for the purposes and criteria outlined in this Privacy Policy and explained further in our privacy center".
Business and enterprise tiers are a different contract. OpenAI's enterprise privacy page says "By default, we do not use your business data for training our models" (openai.com/enterprise-privacy, accessed 27 August 2026), which is worth noting precisely because it is stated about business data specifically. Read the page that covers the tier you are actually on, on the day you decide.
The practical workaround costs nothing. Strip identifying detail before you paste. Initials instead of names, "a family in the congregation" instead of a family, invented names in any example. A model checking a rota for clashes does not need to know that J.M. is Joan.
Which prompts are worth saving?
Twenty-six templates, grouped by the job. Square brackets mark what you replace. They are written to be boring on purpose: constrained, specific about what not to do, and biased towards asking you for missing information rather than inventing it.
Newsletters and weekly notices
1. Raw items into ordered notices
You are drafting the weekly notices for [ORGANISATION NAME], a [TRADITION, e.g. "Methodist church", "Sikh gurdwara", "Reform synagogue", "Buddhist meditation centre"] with about [NUMBER] regular attenders in [TOWN OR CITY].
Here are this week's raw items, in no particular order:
[PASTE YOUR BULLET LIST: items, dates, times, locations, contact names]
Turn them into notices for our weekly email.
Rules:
- Order by date, soonest first. Anything with a deadline this week goes at the top.
- One short paragraph per item: what it is, when, where, who to contact.
- Plain, warm, unhurried. No marketing language, no exclamation marks, no urgency the item does not actually have.
- Do not add any item I did not give you. Do not invent times, room names, or contact details.
- If an item is missing a date, time or location, list it at the end under "Needs details from me" rather than guessing.
2. Last year's notice, updated
Here is a notice we sent last year:
[PASTE LAST YEAR'S TEXT]
Here is what has changed this year:
[PASTE YOUR LIST OF CHANGES]
Rewrite it for this year. Keep the structure and the tone exactly. Change only what my list says has changed.
Then give me a bullet list of every change you made, so I can check it against my list line by line before I send it.
3. Subject lines that do not oversell
Write 8 subject lines for this week's email from [ORGANISATION NAME].
The main item is [ITEM]. The audience is [DESCRIBE: e.g. "mostly over 60, reads on a phone, about half open it in any given week"].
Rules: under 55 characters. No emoji. No false urgency. No curiosity-gap phrasing. State the thing plainly.
Two of the eight should be nothing but the date and the main event.
Events and promotion
4. One event, three lengths
We are running [EVENT] on [DATE] at [TIME], at [LOCATION]. It is for [WHO IT IS FOR]. It costs [PRICE OR "free"]. To attend, people need to [BOOK VIA X / JUST TURN UP].
Practical details people always ask about: [PARKING, STEP-FREE ACCESS, CHILDCARE, FOOD, WHAT TO BRING, HOW LONG IT LASTS].
Write three versions of the announcement:
1. 90 words, for the weekly email.
2. 40 words, for a printed slip.
3. 25 words, for the noticeboard and the website banner.
All three must contain the date, the time, the place and how to attend. Do not add benefits, promises or descriptions I did not give you.
5. The awkward ask
Write a short message asking people to help with [TASK] at [EVENT]. We need [NUMBER] people for [TIME WINDOW]. In previous years the same handful of people did all of it.
Tone: direct and unembarrassed. Say exactly what is being asked, how long it takes, and what someone needs to be able to do (for example lifting, driving, a DBS check, none of the above).
No guilt. No "we desperately need". No suggestion that declining is a failure. End with one clear way to say yes: [HOW].
6. Post-event note
Write a thank-you note to everyone who came to [EVENT] on [DATE].
Here is what actually happened: [ATTENDANCE, WHAT WENT WELL, ANYTHING RAISED OR COLLECTED, ANYTHING THAT DID NOT GO TO PLAN].
Rules: 120 words maximum. Name no individuals except these: [NAMES OR "none"]. No superlatives. One closing sentence pointing to the next thing, which is [NEXT EVENT AND DATE].
Volunteers, rotas and coordination
7. Rota conflict check (initials only)
Here is a draft rota using initials only, with availability:
[PASTE THE GRID]
The rules for [MONTH] are:
[e.g. "nobody two weeks running", "at least one first-aider each week", "J.M. cannot do mornings", "two people minimum on the door"]
Check the draft against the rules. Output only:
1. Every rule the draft breaks, and the date it breaks on.
2. Every date with nobody assigned.
3. Anyone rostered more than [NUMBER] times this month.
Do not produce a new rota. Do not fill the gaps yourself. Do not guess at anyone's availability.
8. Volunteer role description
Write a role description for [ROLE] at [ORGANISATION NAME].
What the role actually involves: [TASKS]
Time commitment: [HOURS, HOW OFTEN, FOR HOW LONG]
Requirements: [CHECKS, TRAINING, PHYSICAL REQUIREMENTS, OR "none"]
Who they work with: [NAMES OF ROLES, NOT PEOPLE]
Who to contact: [ROLE, NOT NAME]
Rules: under 250 words. Be honest about the boring parts. Do not describe it as rewarding, fulfilling or a chance to make a difference. Someone reading this should be able to work out in thirty seconds whether it is for them.
9. Handover note before someone stands down
[NAME OF ROLE, not the person] is standing down after [TIME]. Here is a brain-dump of everything they do, in the order they thought of it:
[PASTE THE DUMP]
Turn it into a handover document with:
- A month-by-month calendar of what happens when.
- A list of recurring tasks with rough time estimates.
- Every login, key, supplier or contact mentioned, listed as "[DESCRIPTION] - details held by [ROLE]". Do not include any actual credentials.
- A section headed "Questions for the outgoing volunteer" with everything the dump does not explain.
Social and short-form
10. One notice, three platforms
Here is this week's notice:
[PASTE]
Turn it into:
1. A Facebook post, 60 to 80 words, with the practical details intact.
2. An Instagram caption, under 40 words, plus a suggested description of what the photo should show.
3. A WhatsApp message for the group, under 30 words, written as a person would type it.
No hashtag stuffing: two hashtags maximum, and only if they are ones people locally actually use. No emoji unless I have used them in the source text.
11. Alt text and image description
Here is a description of a photo we want to post: [DESCRIBE THE PHOTO IN PLAIN WORDS]
Write:
1. Alt text under 125 characters, describing what is visibly in the image, for someone using a screen reader.
2. A one-sentence caption.
Do not describe anyone's identity, age, or condition unless I have stated it. Do not name individuals. Do not editorialise about the mood of the scene.
Grants, funding and reporting
12. Grant question, drafted from your own facts
We are applying to [FUNDER] for [AMOUNT] towards [PROJECT].
The question is: "[PASTE THE EXACT QUESTION]"
The word limit is [NUMBER].
Here are our facts, and the only facts you may use:
[PASTE: what the project is, who it serves, numbers you can evidence, what you have done before, who else is involved, what the money buys]
Draft an answer within the word limit. Every claim must trace back to something in my list. If the question asks for something I have not given you, do not invent it: write "[MISSING: what you need from me]" in place of that sentence.
13. Plain-English budget narrative
Here is our budget for [PROJECT]:
[PASTE THE LINE ITEMS AND AMOUNTS]
Write a plain-English narrative explaining what each line buys and why it is needed, for a funder who does not know our organisation.
Rules: do not restate the numbers I gave you incorrectly, and do not calculate any total, subtotal or percentage. If a figure would help the narrative, write "[CALCULATE: what is needed]" and I will fill it in from the spreadsheet.
14. Report to the trustees or committee
Turn these notes into a report for [TRUSTEES / COMMITTEE / COUNCIL] covering [PERIOD]:
[PASTE ROUGH NOTES]
Structure: what we said we would do, what actually happened, what changed and why, what we are asking the meeting to decide.
Rules: under [NUMBER] words. Neutral tone. Do not soften bad news. Anything I have flagged as a problem stays flagged as a problem. End with a list of decisions needed, each phrased as a question the meeting can answer yes or no to.
Plain language and accessibility
15. Reading-age rewrite
Rewrite this so a confident twelve-year-old could read it aloud and understand it:
[PASTE]
Rules: short sentences. Common words. Explain any term specific to our tradition the first time it appears, in a short clause rather than a footnote. Keep every fact, date and name exactly as written. Do not add warmth, encouragement or explanation that was not in the original.
Then list every term you had to explain, so I can decide whether to keep or cut each one.
16. Jargon audit
Read this document as someone who has never been to a place of worship of any kind:
[PASTE]
List every word, phrase, abbreviation, name or assumed reference that would stop them. For each one, say what a newcomer would probably assume it means, and suggest a plain alternative or a short in-line gloss.
Do not rewrite the document. I want the list.
17. Large-print and printed-notice version
Reformat this for a large-print handout and a printed noticeboard:
[PASTE]
Rules: remove anything that only works on screen, such as "click here" or "see the link below". Spell out any URL as words a person could type. Move the essential information (what, when, where, who to contact) into the first three lines. Keep the whole thing under [NUMBER] words so it fits at 18-point type.
Translation for a multilingual congregation
18. Translated notice with a do-not-translate list
Translate this notice into [LANGUAGE], for [WHO WILL READ IT: e.g. "first-generation speakers who read the language fluently", "younger family members who speak it but read it less confidently"]:
[PASTE THE NOTICE]
Do not translate these terms; keep them exactly as written:
[LIST: names of people, the name of the building, festival names you want kept, anything with an official English form]
Keep the register plain and respectful. Where a phrase has no direct equivalent, translate the meaning rather than the words, and note what you did at the end.
19. Back-translation check
Here is a translation of our notice into [LANGUAGE]:
[PASTE THE TRANSLATION]
Translate it back into English, literally, without smoothing anything out. Then list anything that has shifted in meaning, tone or formality compared to this original:
[PASTE THE ENGLISH ORIGINAL]
I am checking for errors before a fluent speaker reviews it, not replacing that review.
Meetings, minutes and internal admin
20. Minutes into actions
Here are the notes from [MEETING] on [DATE]:
[PASTE]
Produce:
1. A list of decisions made, one line each, in the words the meeting used.
2. A list of actions: what, who by role, by when. If the notes do not say who or when, write "[UNASSIGNED]" rather than guessing.
3. A list of items raised but not resolved, for the next agenda.
Do not summarise the discussion. Do not attribute opinions to named individuals.
21. Agenda from a backlog
Here is everything waiting for a decision:
[PASTE THE LIST]
Build an agenda for a [LENGTH] meeting on [DATE]. Put anything time-critical first, group related items, and give each item a realistic time box that adds up to less than the meeting length.
Mark each item as "decision", "discussion" or "information". Anything that does not fit, list at the end under "Deferred", with a one-line reason.
22. Policy document into a one-page summary
Here is our [POLICY NAME]:
[PASTE THE POLICY]
Write a one-page summary for volunteers, covering only: what the policy is for, what it asks them to do differently, who to go to with a question, and where the full document lives.
Rules: do not paraphrase any legal or regulatory obligation loosely. Where the policy states a requirement, quote it exactly and mark it as a quotation. If you are unsure whether something is a requirement or a suggestion, list it separately under "Check with [ROLE]".
Welcome and newcomer materials
23. What happens when you arrive
Write a "what to expect on your first visit" page for [ORGANISATION NAME].
Here is what actually happens: [DESCRIBE THE ARRIVAL, THE TIMINGS, WHAT PEOPLE WEAR, WHETHER ANYONE WILL SPEAK TO THEM, WHAT HAPPENS WITH CHILDREN, WHETHER THERE IS FOOD, HOW LONG IT LASTS, WHERE TO PARK, ACCESS ARRANGEMENTS]
Write it for someone who has never been to a place of worship of any kind and is slightly nervous about doing the wrong thing.
Rules: answer the practical questions first. Say plainly what is expected and what is not. Do not use any term specific to our tradition without explaining it in the same sentence. Do not promise anyone an experience.
24. Enquiry reply templates
Write short reply templates for the five enquiries we get most often:
[LIST YOUR FIVE, e.g. hall hire, weddings or naming ceremonies, funerals, volunteering, "can I just come along?"]
For each: under 100 words, warm and specific, answering the practical question first and giving one clear next step.
Leave [SQUARE BRACKETS] wherever a real person must add a date, a name, a price or an availability, rather than filling them with placeholder text that could be sent by accident.
Preparation for teaching, bounded on purpose
These two return questions and candidates, not prose. That is deliberate. The writing stays with the person who has to stand up and say it.
25. Questions a text raises
I am preparing to teach on [PASSAGE OR TEXT OR TOPIC], for [AUDIENCE], for [LENGTH].
Do not write the talk. Do not summarise the text for me. Do not suggest an outline.
Instead, give me:
1. Ten questions this text raises that a thoughtful listener might ask.
2. Three places where the wording is commonly misread, and what the misreading is.
3. Anything above where you are not confident, marked clearly as uncertain.
Do not quote the text back to me; I will read it myself. Do not cite a commentary, a translation or a historical claim unless you name it precisely enough for me to look it up.
26. Illustration candidates to verify
I am looking for illustrations for a talk about [THEME], for an audience of [WHO].
Suggest 8 candidates: everyday situations, historical episodes, or well-documented events.
For each one give me:
- a one-line summary
- the specific factual claim I would need to verify
- where I would go to verify it
Mark clearly any that you are reconstructing from memory rather than from a source you can name. I will check every one before I use it, so flag the risky ones rather than smoothing them over.
How do you keep it sounding like your community?
Not with adjectives. Telling a model to be "warm but not saccharine" produces the average of every organisation that has ever described itself that way, which is exactly the sound people recognise as machine-written.
Give it examples instead. Paste three or four things you have actually sent, say what to keep and what to drop, and ban your specific dislikes by name. Then save that as a prompt template you reuse rather than retyping it, because the version you retype from memory drifts within a fortnight. Building a brand-voice context walks through the fuller version of this, and it transfers directly.
Here are four things we have actually sent out:
[PASTE FOUR SAMPLES]
Describe our writing in specific, observable terms: sentence length, how we open, how we close, what we call things, what we never do.
Then write me a set of instructions I can paste at the top of every future request, so a model writes in that voice. Include a "never do this" list based on what is absent from my samples.
Two habits matter more than any prompt. First, one person owns the final read. Rotating that job across six volunteers produces six voices, and the inconsistency reads as carelessness even when every individual draft is fine. Second, the person who writes the prompt is not the person who approves the output. That separation catches the wrong date before it reaches four hundred inboxes.
How do you set this up in an afternoon?
- Pick the three tasks that eat the most volunteer time. For most organisations that is the weekly notices, the rota, and whatever is due to a funder.
- Write one prompt per task, using the templates above as a starting point. Fill in the real names, the real constraints, the real audience.
- Run each one against last week's real work and compare the output to what you actually sent. Where it is wrong, fix the prompt rather than the output.
- Save the working versions somewhere the next volunteer can find them. A shared document is fine to start with. What is not fine is one person's chat history.
- Write down the verification step for each: who checks the dates, who checks the references, who reads it before it sends.
- Agree, once, in writing, what does not go into a chat window. See the list further up.
If step four sounds like the part that quietly fails, it usually is. Prompts that live in one volunteer's chat history leave when that volunteer leaves, and the next person starts from a blank box. A shared library with fill-in-the-blank variables fixes that, and how to build a prompt library covers the mechanics for a small team. Prompt Architects is built for exactly that job, with a free plan that includes 5 prompt enhancements per day, forever, according to our FAQ page as of 27 August 2026.
It is worth being plain about what it is not. Prompt Architects is a prompt library and enhancer. It is not church management software: it does not hold your membership list, run your rota, take your donations or track attendance, and it should not. If your problem is that nobody knows who is on the door in three weeks, a better prompt does not solve that. Your rota system does.
For the other side of this work, two adjacent guides do more than this post has room for: grant writing and abstract templates if funding applications are the bottleneck, and prompting in languages other than English if your congregation reads in more than one.
What does this look like after a month?
The newsletter goes out on time because the first draft takes ten minutes instead of ninety. The grant application gets written because the blank page is no longer the obstacle. The Punjabi and Polish versions of the notices exist, imperfectly, where before they did not exist at all. Somebody who is not the retired teacher can now produce the rota.
What has not changed is anything that matters most. The talk is still written by the person giving it. The visit to the hospital is still made by a person. The decision about what to say when something difficult happens is still made by people who know the family. That is the correct division of labour, and it does not need defending to anyone who has done the work.
Start with one task. Write one prompt. Save it where the next volunteer will find it.
Stop rewriting prompts. Start shipping.
Works with ChatGPT, Claude, Gemini, Grok, Midjourney, Ideogram, Veo3 & Kling. 5.0★ on the Chrome Web Store.
Create An Account