Most recruiters have the same first experience with AI. They ask for a candidate outreach message, get back something that reads like it was written for anyone in any industry, and quietly decide the tool is not for them. The model is not the problem. The prompt asked for a message without saying who it is for, what the role actually needs, what the recruiter already knows and what the message has to achieve.
Prompting is not a mystical skill. It is closer to briefing a very fast, very literal contractor who has never met your client. If you would not send that brief to a new resourcer on their first morning, an AI model will not do much better with it.
What makes a recruitment prompt fail
Six failure modes cover almost everything we see. They are all fixable, and none of them require special software.
1. The prompt carries no real role context
"Write an outreach message for a backend engineer role" leaves the model to invent the interesting parts: the product, the stack, the reason a good engineer would move. It will invent them, confidently, and you will spend longer editing than writing.
2. The audience is missing
A message to a passive senior engineer who is not looking, a candidate who applied last week and a hiring manager chasing an update are three different pieces of writing. If the prompt does not name the reader and their likely state of mind, you get the average of all three.
3. There is no recruitment objective
Output improves sharply when the prompt states what should happen next: a reply, a booked call, a decision on two shortlisted candidates, an answer to one specific question. Without that, the model optimizes for sounding complete rather than moving the search forward.
4. No constraints
Length, tone, format, what to avoid, what must not be claimed. Recruiters usually have strong opinions here and rarely write them down. Two lines of constraint remove most of the editing.
5. Too much irrelevant information
The opposite error is just as common. Pasting an entire job description, the client's boilerplate and three pages of notes into a prompt for a two paragraph message buries the signal. The model weights what it is given. Give it the parts that matter for this task.
6. One prompt trying to do everything
Screening criteria, an outreach sequence, interview questions and a client update in one request produces four mediocre outputs. Split the work. Each task gets its own prompt with its own audience and objective.
A framework that works without any particular tool
Answer-first version: a recruitment prompt should carry the role, the audience, the objective, the material, the constraints and the boundaries. Anything else is optional.
The six-part recruitment prompt
- Role: the job, the level, the market, and the two or three things that genuinely define a strong candidate for it.
- Audience: who reads this, what they already know, and how they are likely to feel about being contacted.
- Objective: the specific next step this output should make possible.
- Material: only the source text this task needs, such as the CV section, the call notes or the requirement being tested.
- Constraints: length, tone, format, and anything the output must not do or claim.
- Boundaries: what the model should flag as unknown rather than guess.
The last one earns its place. Recruitment prompts fail badly when a model fills gaps to be helpful. Adding a line such as "if the material does not establish something, say so rather than assuming it" turns invention into a visible question you can go and answer.
A weak prompt and a workable one
| Element | Weak prompt | Workable prompt |
|---|---|---|
| Request | Write a LinkedIn message for a backend engineer role. | Write a first approach to a senior backend engineer who is not actively looking. |
| Role context | None. | Payments platform, Python and Go, moving off a monolith, London hybrid three days. |
| Objective | Implied. | Get a 15 minute exploratory call, or a clear no. |
| Constraints | None. | Under 120 words, no superlatives, no salary claim, one question at the end. |
| Boundaries | None. | Do not invent benefits or team size. Flag anything you would need from me. |
The second prompt takes about forty seconds longer to write and removes most of the rewriting afterwards. That is the entire trade.
How can recruiters use ChatGPT or Claude effectively?
Use it where the work is structuring, drafting and comparing, and keep judgment with yourself. Practically that means: summarizing a long call into notes you check, drafting a message you edit, listing what a CV does and does not evidence against your requirements, or turning a rambling brief into a structured set of questions for the client. It does not mean asking a model who to shortlist.
The other habit that helps is reusing what worked. If a prompt produced a good screening summary once, it will produce a good one next week, provided the role context is swapped correctly. Rewriting it from scratch every time is where most of the wasted effort lives.
Where prompt quality and role context meet
The single biggest lift in recruitment prompt quality is not clever phrasing. It is having the role written down once in a form you can reuse: the outcomes, the must-haves, the preferences, the client context and the questions still open. Once that exists, every prompt for that search starts from the same understanding instead of whatever you can remember at the time.
Where prompt quality comes from
- Role contextOutcomes, must-haves, preferences, open questions
- Task framingAudience, objective, constraints
- Relevant material onlyThe CV section or notes this task needs
How Creo Access approaches this
Creo Access is an AI recruitment workspace built around the Role. Two parts of it deal directly with prompting.
The Prompt Generator builds a prompt for the task you describe, and it is Role aware: if you attach a Role, it pulls in the parts of that Role which matter for the task at hand rather than pasting everything. Outreach draws on the pitch and the market. Screening draws on the must-haves and the open questions. It also distinguishes between what is already known about the Role and what is unresolved, so the generated prompt does not ask you for information you have already recorded.
Building a recruitment prompt from a Role
Open each numbered part to see how it enters the finished prompt.
The Prompt Generator composes the prompt from parts you can inspect. Attaching a Role adds only the parts of that Role the task needs.
Read this figure as text
- 1. Role: Who is doing the work, and for which search. You are helping a recruiter working a Finance Systems Lead search for a mid-market group.
- 2. Context: The parts of the Role that matter for this task only. The outcome is cutting month-end close from nine days to three. Owning a close is an established requirement; the group's new ERP is a preference.
- 3. Task: One instruction, not a wish list. Draft five screening questions that would establish whether a candidate has personally owned a close.
- 4. Format: The shape of the answer you can actually use. Return a numbered list. One question per line, no preamble, no commentary.
- 5. Exclusions: What the model should not do. Do not score the candidate, do not invent detail about the company, and do not ask anything answerable from a CV.
The Prompt Library is the other half: a curated set of recruitment prompts for outreach, screening, client work and research, which you can refine into your own working version rather than rewriting from memory each week.
Neither replaces the framework above. They make it cheaper to apply consistently, which is usually the difference between a good prompt habit and one that lasts a fortnight.
Try it on a live search
Want to see prompt output that already knows the Role?
Build a Role once, then generate prompts that carry it. Start free for 14 days. Then choose monthly or annual billing. Cancel anytime.
Start 14-day free trial