AI prompts for trainers: 10 examples to adapt

Learning pathway design

• “Break this learn­ing ob­ject­ive down into 5 pro­gress­ively se­quenced activ­it­ies, each with an ob­serv­able de­liv­er­able.”
• “Sug­gest three dif­fer­ent ap­proaches to teach­ing [concept] for be­gin­ner, in­ter­me­di­ate and expert learners.”
• “Using this course out­line, list the im­pli­cit pre­requis­ites that I should make ex­pli­cit in the in­tro­duc­tion.”

Content production

• “Re­write this para­graph, keep­ing its mean­ing but chan­ging the aca­dem­ic re­gister to prac­tic­al lan­guage for front­line man­agers.”
• “Gen­er­ate three con­crete ex­amples from dif­fer­ent sec­tors that il­lus­trate [concept].”
• “Turn this list of bullet points into a 90-second voice-over script in a con­ver­sa­tion­al tone.”

Assessment

• “Create 5 quiz ques­tions of in­creas­ing dif­fi­culty on this con­tent, in­clud­ing 2 com­pre­hen­sion ques­tions and 3 ap­plic­a­tion ques­tions.”
• “Sug­gest a work­place scen­ario that draws on the con­cepts in [chapter 3].”
• “Ana­lyse these learners’ open-ended re­sponses and identi­fy the 3 most common reas­on­ing errors.”
• “Gen­er­ate an as­sess­ment rubric based on this course’s ob­ject­ives.”

To decide what can be an auto­mated sug­ges­tion and what re­mains a human re­spons­ib­il­ity, read AI and trainers: planning thoughtful collaboration. Do not submit per­son­al data or con­fid­en­tial ma­ter­i­als to a tool without an au­thor­ised frame­work; an­onymise ex­amples where pos­sible.

Going further: AI within the learning process

These ex­amples con­cern train­ing pre­par­a­tion. Tools may also offer in­ter­ac­tion with learners during a course; avail­able fea­tures depend on the solu­tion and must be checked before you de­scribe them. To ex­plore this topic, see voice interaction in learning and the solu­tion for independent trainers.

Specify the output rather than making a vague request

The ten prompts above be­ne­fit from an output spe­cific­a­tion. De­scribe the audi­ence in terms of ob­serv­able pre­requis­ites, the work­place task to be learned, the au­thor­ised source ma­ter­i­al and the ex­pec­ted de­liv­er­able. Also spe­cify what the AI must not do: fill in a miss­ing pro­ced­ure, create a ref­er­ence or turn a re­com­mend­a­tion into a re­quire­ment. A short re­quest may be suf­fi­cient if these con­straints are ex­pli­cit; piling up con­tra­dict­ory in­struc­tions makes val­id­a­tion harder.

Fic­tion­al ex­ample of an ex­pan­ded prompt: “Using only the ex­tract below, pre­pare two scen­ari­os for novice case ad­min­is­trat­ors who need to identi­fy a miss­ing doc­u­ment. For each scen­ario, provide the in­struc­tions, avail­able in­form­a­tion, ex­pec­ted answer, its jus­ti­fic­a­tion and the exact pas­sage sup­port­ing it. If the ex­tract does not allow a de­cision, state in­suf­fi­cient in­form­a­tion. Do not add any rules.” The train­er’s de­liv­er­able is not the entire con­ver­sa­tion, but a re­viewed activ­ity sheet ac­com­pan­ied by its doc­u­ment­ary ref­er­ence. Keep the source ex­tract in an au­thor­ised loc­a­tion, sep­ar­ate from the learner-facing ver­sion.

Test the prompt on a small set of cases

Before re­using a prompt, create three fic­tion­al inputs: an or­din­ary case, an am­bigu­ous case and a case that cannot be re­solved using the source ma­ter­i­al. This last input checks wheth­er the in­struc­tion not to invent in­form­a­tion is fol­lowed. Also repeat the same re­quest: one cor­rect re­sponse is not enough to es­tab­lish re­li­ab­il­ity. Record the tool’s dis­played name and ver­sion, the date, the prompt text and the cor­rec­tions re­quired. This will help you dis­tin­guish a change in source ma­ter­i­al from a change in the as­sist­ant’s be­ha­viour.

Use a simple de­cision rubric: every sug­ges­tion must be ac­cur­ate, trace­able, ap­pro­pri­ate to the level and usable without hidden in­form­a­tion. An error con­cern­ing an es­sen­tial rule means re­jec­tion, even if the writ­ing is ex­cel­lent. Ex­cess­ive length calls for re­vi­sion; it never com­pensates for a sub­stant­ive error. The de­liv­er­able is a table with input, output, iden­ti­fied defect, human cor­rec­tion and de­cision. This pro­tocol is a local pro­duc­tion check, not model cer­ti­fic­a­tion or a meas­ure of in­struc­tion­al ef­fect­ive­ness.

Check questions before releasing them

For an AI-gen­er­ated quiz, have a col­league answer each ques­tion without show­ing them the answer key. Ask them to identi­fy the clue that sup­ports the answer and ex­plain why the other op­tions are less ap­pro­pri­ate. If two an­swers can be jus­ti­fied using the ma­ter­i­als, re­write the ques­tion or ex­pli­citly accept mul­tiple solu­tions. Remove absurd dis­tract­ors, nested neg­at­ives and de­tails that test tricky read­ing rather than the target skill. In­creas­ing dif­fi­culty should come from the reas­on­ing re­quired, not un­ne­ces­sar­ily com­plic­ated vocab­u­lary.

Then pre­pare feed­back for each plaus­ible error. In a fic­tion­al case-file clas­si­fic­a­tion ex­ample, con­fus­ing a miss­ing doc­u­ment with an ex­pired doc­u­ment calls for two dif­fer­ent ex­plan­a­tions. The answer key should re­state the cri­terion, show how it ap­plies and sug­gest a next action. Do not use auto­mated ana­lys­is of open-ended re­sponses as the sole basis for high-stakes grad­ing. Start by com­par­ing it with human mark­ing on fic­tion­al re­sponses, then ex­am­ine dis­agree­ments before using it with learners.

Maintain a safe, revisable prompt library

The UNESCO guidance published on 7 September 2023 re­com­mends a human-centred ap­proach, with eth­ic­al and ped­ago­gic­al val­id­a­tion of edu­ca­tion­al uses of gen­er­at­ive AI. It does not provide evid­ence that a par­tic­u­lar prompt im­proves a course. To put this guid­ance into prac­tice, assign each prompt sheet an owner, an au­thor­ised scope, a review date and situ­ations in which the prompt must not be used.

Do not in­clude iden­ti­fi­able learner sub­mis­sions or con­fid­en­tial ex­changes in the lib­rary. Re­pla­cing a name alone may still leave someone iden­ti­fi­able through their role or cir­cum­stances; prefer en­tirely fic­tion­al data for test­ing. Keep a val­id­ated ver­sion of the de­liv­er­able sep­ar­ate from drafts and record the checks per­formed. When source ma­ter­i­al changes, rerun the test inputs before re­pub­lish­ing. A useful lib­rary there­fore con­tains fewer magic for­mu­las and more tested prompt sheets, com­plete with lim­it­a­tions and cor­rec­tion his­tor­ies.

Key takeaways

These ten prompts are start­ing points to adapt and review. AI can con­trib­ute to design and to some learn­ing activ­it­ies, provided its re­sponses are checked and ap­pro­pri­ate in­struc­tion­al over­sight is main­tained.