1. Test whether the post is real

Here is the role description for a community link worker post: [paste].
Give me: (a) the three decisions this person can actually make,
(b) three things they are only told about, (c) the questions I should
ask before accepting. Quote the words in the description that support each answer.

Check: if the model claims a power the description does not grant, reject the answer and ask it to quote again.

2. Turn the protocol into a plain explanation

Here is a participant information sheet: [paste, anonymised].
Write a two-minute spoken explanation for a community member with no
research background. Reading age 11. No jargon. Cover: purpose, what
taking part involves, time, payment, data, and how to say no.
Then list anything in the sheet that the explanation cannot honestly simplify.

Check: read the "cannot honestly simplify" list first. That list is where consent goes wrong.

3. Find who the recruitment plan misses

Here is our recruitment plan: [paste].
For each channel, tell me: who it reaches, who it cannot reach,
and who it systematically excludes. Then name three channels that would
reach the people missing, and what each one would cost in time and money.

Check: a channel that reaches "everyone" is a channel that was not analysed. Ask for the exclusion again.

4. Rehearse the invitation conversation

Role-play a community member who is interested but says "my English
is not good enough for this". Ask me the questions a real person would
ask. Do not be easy to persuade. After ten turns, stop and tell me the
three moments where I pressured the person.

Check: if the role-play gives in immediately, it is not testing anything. Ask for a harder version.

5. Write the invitation in the community's words

Draft three versions of a one-paragraph invitation to this study: one
for a WhatsApp group, one for a faith setting notice, one to be read
aloud. Audience: [describe]. Reading age 11. Then state which version
you would not use, and why.

Check: a version that sounds like a leaflet from a hospital is not usable. Say so and ask again.

6. Pressure-test the topic guide

Here is my topic guide: [paste].
For each question, tell me: is it open, is it single-barrelled, does it
suggest its own answer, and what might it shut down? Rewrite the two
weakest. Keep the meaning of every question.

Check: rewritten questions must ask the same thing. If the meaning changed, the guide is no longer the approved one.

7. Plan for the difficult moment

A participant in a focus group discloses that a relative is being
neglected by a paid carer, and asks me not to tell anyone.
Give me: what I say at that moment, what I record, who I tell and by
when, and what I do not do. Then tell me which part of this I must
confirm with the study lead before the session.

Check: any answer that promises secrecy is wrong. Any answer that investigates is wrong. Rerun with that instruction if needed.

8. Write the debrief from your notes

Here are my notes from today: [paste, anonymised].
Produce: what went well, what I would change, what I noticed that is not
in the recording, anything for the team, anything for supervision.
Use only what is in my notes. List separately anything you think I have
missed, and ask me rather than filling the gap.

Check: the "you have missed" list is the useful part. Anything the model filled in without asking is invented.

9. Interrogate a draft theme

Draft theme: "[paste]". Supporting quotations: [paste, anonymised].
Tell me: what this theme claims, what the evidence actually supports,
whose voice is missing, and one alternative reading. Do not write the
final wording.

Check: if it agrees with the theme too quickly, give it the counter-cases and ask again.

10. Write the member-checking question set

Here are three draft findings: [paste].
Write four questions I can ask a community group to test whether these
are recognisable, in plain language, that do not lead the answer, and
that invite disagreement. Then write how I record a disagreement.

Check: a question that contains the finding is not a test. Ask for the neutral version.

11. Phrase the refusal

I have been asked to do this task: [describe]. It is outside my training
or my written agreement because: [reason]. Write three versions of a
refusal: to a supervisor, to a researcher in the room, and to a community
partner. Each must name the limit, offer what I can do, and put the
decision where it belongs.

Check: a refusal that apologises more than once is a refusal that will be argued with.

12. Prepare for supervision

Here is what happened this week: [notes].
Give me: the one thing I most need to raise, the evidence I should bring,
the two options I can propose, and the question I should ask my supervisor
to answer directly.

Check: supervision is not a performance review. If the output is a confession, rewrite the request.

The standing rule for all twelve

An AI assistant will invent a policy, a rate, a form name or a legal duty to make an answer look complete. Every time you use one of these prompts, ask one more question: "Which of these statements are from the material I gave you, and which did you infer?" Use only the first group.