Course resource
HR and Recruitment Prompt Library
Prompts for people work, and the boundaries that are not negotiable.
The boundaries
Never let AI decide about a person. Not hiring, not firing, not promotion, not discipline, not performance rating. In many jurisdictions an automated decision materially affecting someone carries specific legal obligations including a right to human review and an explanation. Beyond the law, it is the wrong thing to do.
Never put identifiable employee or candidate data into a general tool. This is among the most sensitive data any organisation holds. Anonymise, or use an approved tier with a processor agreement.
Test for bias before you rely on anything. Matched-pair testing is described at the end of this document and takes twenty minutes.
AI drafts. A person decides, and their reasoning is recorded.
Job descriptions
Writing one
Write a job description for [ROLE].
Actually does: [THE REAL DAY-TO-DAY, NOT ASPIRATIONS]
Must be able to: [GENUINE REQUIREMENTS]
Nice to have: [SEPARATE THESE PROPERLY]
Team: [SIZE, STRUCTURE]
Salary range: [STATE IT]
Rules:
- No "rockstar", "ninja", "wear many hats", "fast-paced environment"
- No degree requirement unless genuinely necessary — say why if it is
- Separate must-have from nice-to-have rigorously. Long must-have lists
disproportionately deter women and under-represented candidates from applying.
- Describe the work, not the culture
- Include the salary range
The bias pass
Review this job description for language that could deter qualified
candidates:
[PASTE]
Check for: gendered language, unnecessary requirements, jargon that
assumes a particular background, cultural references, physical
requirements not essential to the role, and an inflated years-of-experience
figure.
For each issue, suggest a replacement.
Screening — with care
AI may help you read more carefully. It must not filter.
Here is an anonymised CV and the role requirements: [PASTE]
For each MUST-HAVE requirement, state:
- Evidence present, with the specific line
- Evidence absent
- Unclear
Do not score. Do not rank. Do not recommend.
Do not comment on anything not related to a stated requirement.
Anonymise before pasting: remove name, age, address, photograph, university, and anything indicating nationality, gender or family status. If you are not removing these, do not use AI for screening at all.
A human reads every rejection. If nobody looks at the people being screened out, you have built an automated decision system regardless of what you call it.
Interviewing
Design interview questions for [ROLE].
Key requirements: [LIST]
For each requirement, give one behavioural question ("tell me about a
time...") and what a strong answer would contain.
Rules:
- Questions must relate only to the requirements
- No questions about personal circumstances, family, health, age,
or anything protected
- Same questions for every candidate — this is a structured interview
Structured interviews with consistent questions are both fairer and more predictive than unstructured ones. This is one of the more robust findings in the field.
Notes, afterwards
Here are my interview notes: [PASTE]
Requirements: [LIST]
Organise the evidence against each requirement.
Flag any requirement where I have no evidence either way.
Do not score. Do not compare to other candidates.
The "no evidence" flag is useful — it catches the requirement you forgot to probe.
Onboarding and internal comms
Onboarding assistant
# ROLE
You answer new-starter questions for [COMPANY] from the documents provided.
# RULES
- Answer only from the knowledge files
- If not covered, say "ask [PERSON/CHANNEL]"
- Never give advice on pay, contracts, immigration, health or legal matters
- Quote the policy and its date
- If someone seems distressed, hand off to a human immediately
Policy communication
Announce [POLICY CHANGE] to [AUDIENCE].
What changes: [WHAT]
Effective: [DATE]
What they must do: [ACTION, OR NOTHING]
Why: [REASON]
Questions to: [WHO]
Lead with what changes and what they must do. Reasoning after.
No corporate euphemism — if it is a reduction in something, say so.
Performance and difficult conversations
I need to give [ROLE — no name] feedback about [BEHAVIOUR].
Specific examples with dates: [TWO OR THREE]
Effect on the work: [IMPACT]
What I want instead: [OBSERVABLE]
Rules:
- Behaviour, never character
- No compliment sandwich
- The requested change must be specific enough to verify
- End by asking for their view
Never draft a dismissal, a disciplinary outcome, or a performance rating with AI. Structure your own thinking with it if you like; the document and the decision are yours, and they may be scrutinised.
The bias test — run this before relying on anything
Twenty minutes, and it is the difference between hoping and knowing.
- Take one real (anonymised) CV or piece of work.
- Create matched versions varying only one attribute: a name suggesting a different gender or ethnicity, a different university, a gap in employment, a non-native phrasing pattern.
- Run each through your prompt in a fresh conversation.
- Compare the outputs.
| Version varied | Output differed? | How |
|---|---|---|
| Name (gender-coded) | ||
| Name (ethnicity-coded) | ||
| University | ||
| Employment gap | ||
| Phrasing (non-native) |
Any systematic difference means stop using it for that purpose. Not "adjust the prompt and hope" — stop, until you can show it does not.
Compliance checklist
- Approved tool with a processor agreement
- Candidate and employee data anonymised before any AI processing
- No automated decision affecting an individual
- A human reviews every rejection
- Bias testing done and documented
- Candidates informed that AI is used in the process, where required
- Retention of AI-processed data covered by your policy
- Decision reasoning recorded separately from any AI output