ChatGPT Prompts for Recruiters: Intake, Outreach and Shortlists

A recruiter can use ChatGPT for job intake notes, candidate outreach and client shortlists by feeding it redacted facts, never raw CVs. Eight prompts cover the cycle: intake call write-up, role brief, outreach, follow-up, screening questions, shortlist summary, client update and rejection note. Use a business plan and strip identifiers before pasting anything about people.

Recruiting is a data-protection business first and a writing business second, so the rules come before the prompts. The sample outputs below use invented roles and placeholders, and each prompt says what to redact. Treat every draft as a draft: you remain responsible for accuracy and for how candidate information is handled.

Follow me on Instagram@sagnikteaches

Candidate data rules to settle before the first prompt

Start with the default that catches agencies out. A consultant using a personal Free, Plus or Pro ChatGPT account is sharing chats for model training until they switch off "Improve the model for everyone" in Data controls, and each consultant has to do that on their own account. OpenAI's help page on data controls says the content of Business, Enterprise and Edu workspaces is not used for training by default. A CV is personal data by any reading, so personal plans are off-limits for CVs, interview notes and anything with a name, email, phone number or full employment history attached.

Connect on LinkedInSagnik Bhattacharya

Even on a business plan, adopt redaction as your default. Replace the candidate's name with "Candidate A", employers with "a regional retailer" and "a logistics firm", and delete contact details, dates of birth, photographs, addresses and anything that reveals protected characteristics. Keep the skills, scope, results and tenure, which are what the prompts need. Put the real names back only in your own system, not in the chat.

Subscribe on YouTube@sagnikteaches

Write these rules on one page and give them to every consultant:

1. No CVs in any chat. Paste a redacted summary you wrote, not the document.
2. No contact details, ever. Not in prompts, not in outputs.
3. No special-category detail. Health, family status, religion and similar stay out of every prompt.
4. Client confidentiality. Do not name the client or reveal a confidential vacancy; use "a confidential search for a finance director".
5. Human decision. The model drafts; a person reads the CV and decides.

Your privacy notice and client contracts may set stricter limits on where data goes, so check them before introducing any tool. For tools that record calls, see AI note takers for recruiters and candidate consent, because consent is a separate question from redaction.

Intake: from a rambling call to an approved brief

Two prompts cover intake: one turns call notes into a brief and lists the gaps, the other turns the approved brief into a role summary candidates will read. Neither needs any candidate data, which makes intake the safest place to start using ChatGPT in an agency.

Prompt 1: turn a messy intake call into a structured role brief

The intake call decides whether a search succeeds. Your notes are fragments: "needs someone senior, not too expensive, must handle suppliers, maybe remote two days". The first prompt turns fragments into a brief and, more usefully, lists what you failed to ask.

I am a recruiter. Here are my notes from a client intake call (client and people redacted):
[paste notes]

Write a role brief with these headings: purpose of the role, must-have skills, nice-to-have skills, team and reporting line, salary range, location and working pattern, interview process and timeline, deal-breakers.
Use only what is in my notes. Where information is missing, write "Not covered:" and list it.
End with five questions I should ask the client to close the gaps.

Illustration: a client hiring a purchasing manager for a bakery chain. The notes say "multi-site, ten staff, needs supplier negotiation, salary around 55 to 60, start asap". The sample output fills in the brief and marks "Not covered: reporting line, remote working, interview stages, whether the post is new or a replacement". The five closing questions include "Why is the role open?" and "What would make this hire a failure after six months?".

What I would fix: models tidy ranges into single numbers or add benefits that the client never mentioned. Compare the salary line to your notes. For a purpose-built version of this step, filling a job intake form from the call with AI goes deeper on the form itself.

Prompt 2: a role summary that sells the job without hype

Candidates respond to specifics: the team size, the problem to solve, the real salary band. The second prompt writes a short advert or role summary from the brief you just approved, with a rule against filler.

Write a 120-word role summary for candidates from this approved brief: [paste].
Include: what the person will do in the first six months, the team, the salary range, working pattern and one honest challenge of the role.
Do not use: dynamic, passionate, rockstar, fast-paced, world-class. Do not claim benefits not listed. Plain English.

A sample output opens: "You will take over supplier contracts for six bakeries, with a budget of about $2m a year and a team of two buyers." The one honest challenge ("some suppliers have not been renegotiated in years") is what makes it credible. Check every figure against the approved brief, and confirm with the client before publishing.

Outreach: a first message and follow-ups that earn replies

Outreach is where most recruiters already use ChatGPT, and where generic drafts are easiest for candidates to spot. Both prompts below depend on one thing the model cannot supply: a reason this person fits, which you found and checked yourself.

Prompt 3: outreach messages that use a verified reason

Outreach fails when it could have been sent to anyone. The fix is one real, public reason this person fits the role, which you found by looking, not by asking the model. Do not paste a profile; summarise it yourself with the name removed.

Write a first outreach message of under 80 words to a passive candidate.
Role: [one-line summary, client confidential]. Reason this person fits (from their public profile, I verified it): [one specific point].
Include: why now, salary range [range], and one low-effort next step (a ten-minute call).
No flattery, no "I came across your profile", no exclamation marks. Sign off with [your name]. Address them as [first name].

Sample output, illustrative: "Hi [first name], I am recruiting for a purchasing lead at a multi-site bakery group, and your move from category buying into a supplier-negotiation role stood out. The role pays $55,000 to $60,000 and reports to the operations director. If it is worth ten minutes this week, I can send a one-page brief first. [your name]". The model should not know the candidate's history, so the sentence about the move came from your note. If it invents detail, delete it.

Add one more check: the message must not reveal anything the candidate has not made public. "I saw you are job-hunting" is a mistake when the person has not said so. Keep the reason about the role fit, not the person's circumstances. Boolean search prompts and checks cover finding the people in the first place.

Prompt 4: three follow-ups that add something each time

Most replies come from the second or third touch. A follow-up that says "just checking in" adds nothing. Ask for a sequence where each message gives a new piece of information.

Write three follow-up messages after an unanswered first message, each under 60 words, spaced 3 days, 7 days and 14 days later.
Each must add one new fact from this list, in this order: [team size and reporting line], [one concrete project in the first six months], [a clear opt-out line].
Keep the tone polite, no guilt, no urgency tricks. Use [first name] and [your name].

The third message should end with a clear way out: "If this is not for you, a quick no is helpful and I will not follow up." That line tends to draw answers for a plain reason: it lowers the cost of replying. It also respects people who are not interested, which matters when you want them in your network next year.

Shortlists: screening questions and the client summary

This stage carries the most candidate data, so both prompts work from the brief and from your own redacted notes, never from a CV.

Prompt 5: screening questions tied to the brief

The prompt below produces a question set from the must-haves, with a note on what a good answer sounds like. It does not score people; it prepares you.

From this role brief, write eight screening questions for a 20-minute call.
For each: the skill it tests, what a strong answer includes, and a follow-up probe.
Ask only about job-relevant skills and experience. Do not include questions about age, family plans, health, nationality or similar.
Brief: [paste]

A sample question: "Tell me about a supplier price rise you negotiated back. What did you offer in return?" with a strong-answer note of "names a figure, describes a trade-off, shows the relationship survived". The explicit ban on protected topics is not decoration. Language models can suggest questions that are unwise or unlawful in some places, so check the list yourself, and see bias checks every recruiter should run before using AI anywhere near screening.

Prompt 6: a shortlist summary the client will read

This is the step where redaction matters most, because it is tempting to paste the CV. Do not. Instead, write your own five-line notes per candidate (the same notes you would keep anyway) and ask the model to turn them into a client-ready summary. For another view of this task, writing candidate summaries clients read covers it from the client side.

Write a shortlist summary for a client from my notes. Candidate A, Candidate B, Candidate C (no names).
For each, my notes: [current role scope, key results, tenure, salary expectation, notice period, my view on fit against the must-haves, one concern]
Format: one paragraph of 70 words each, then a comparison table against the four must-haves (met / partly / not evidenced).
Use only my notes; do not infer skills. Do not rank unless I ask.

That is the sample table, illustrative. The phrase "not evidenced" is deliberate: it says the notes contain no proof, which is different from "does not have the skill". The model tends to turn "not evidenced" into "lacks", which is unfair to the candidate and can mislead the client, so correct it. Swap Candidate A, B and C for real names only in the document you send, and check each paragraph against the CV you hold.

Keeping the client and unsuccessful candidates informed

The last two prompts handle the messages that slip when a search gets busy: the weekly update to the client and the note to people who did not make the list.

Prompt 7: a weekly client update that shows progress

Clients get nervous when a search goes quiet. A short weekly update keeps trust and saves you phone calls. Give the model your counts and let it write.

Write a 100-word weekly search update for a client.
Numbers this week: approached [n], replied [n], screened [n], shortlisted [n], interviews booked [n].
Market feedback: [one honest line, for example "candidates ask about remote working"]. Next week: [plan].
Tone: plain and factual. Do not predict hiring dates or promise outcomes.

A sample: "This week we approached 34 people, 11 replied and 6 were screened. Two are shortlisted and one interview is booked for Thursday. Several candidates asked whether the role allows two remote days; an answer would help us. Next week we are interviewing and extending the search to adjacent sectors." The numbers are yours, so keep them accurate; do not let the model round 11 replies to "a third". A sum worth adding: 11 out of 34 is 32%, which a client may read as strong. Say the number, not the adjective.

Prompt 8: a rejection note that is brief and kind

Rejections are the messages recruiters avoid and candidates remember. A prepared note takes seconds and protects your reputation. Never include reasons you cannot back up.

Write a short, kind message telling a candidate they were not shortlisted.
Facts: role [title], stage reached [stage]. Reason to give: [one neutral, true reason, for example "the client chose candidates with more multi-site experience"].
Do not give feedback on personal qualities. Offer to keep in touch only if I say so: [yes / no]. Under 80 words. Use [first name] and [your name].

The sample output is two warm sentences, a neutral reason and one line of thanks. Read it for anything that implies a judgement of the person; keep to the facts of the search. If feedback is requested, give it in a call, not in an AI-drafted paragraph.

One search from intake to shortlist, with the numbers

To see the prompts working together, follow an illustrative search for a purchasing manager at a bakery group. Monday: a 40-minute intake call. Prompt 1 turns the notes into a brief in about five minutes and returns six "Not covered" items. A follow-up email to the client closes four of them, and the brief is approved by Tuesday afternoon. Writing it by hand typically takes 25 to 30 minutes, so the saving here is modest; the real gain is the list of gaps.

Wednesday: you search and note 45 people who look right. Prompt 3 produces a draft for each, but you spend two minutes per person adding the verified reason, so 45 messages take about 90 minutes instead of the three hours a fully hand-written set might take. Day 3 and day 7 follow-ups from Prompt 4 go to those who have not replied. By the following Wednesday 14 have replied, nine are interested and six pass the 20-minute screen using questions from Prompt 5.

Friday: three candidates are shortlisted. You write five-line notes for each (ten minutes each), redact them, and Prompt 6 gives you a client summary and comparison table in a couple of minutes. After your edit, the pack goes out the same afternoon. The sums are illustrative, but the pattern is typical: the model saves time on drafting and formatting, not on judgement. The judgement stays with you and eats most of the clock.

Consumer plan or business workspace: what changes for candidate data

Teams often ask whether a paid personal plan is "good enough". The distinction is about the plan type, not the price. The points below come from OpenAI's data controls page and pricing page, fetched in November 2026; check both before you decide because plan terms change.

The last row is the practical one for an agency. On personal accounts every consultant decides alone, and one forgotten switch undoes the policy. A workspace gives one place to set rules. On cost, the Business Standard seat appears on OpenAI's business pricing page at $25 per person per month, or $20 when paid annually, for teams from two people (as of November 2026). A three-consultant agency on annual billing would pay $60 a month. The main ChatGPT pricing page no longer shows the seat price, so check the business page before you budget; prices change.

The generic opener against the verified-reason version

Here is the same opening written two ways. The generic version: "Hi [first name], I came across your impressive profile and think you could be a great fit for an exciting opportunity with a leading company. Do you have time for a quick chat?" It names no role, no salary and no reason, and nothing in it could not have gone to a thousand people.

The improved version, after the Prompt 3 rules: "Hi [first name], a bakery group is hiring a purchasing lead to renegotiate contracts across six sites, and your work moving a supplier base onto fixed-price terms is the closest match I have seen. The salary is $55,000 to $60,000. Worth a ten-minute call? I can send a one-page brief first." It is 52 words and every claim is checkable. The edit instruction that produced it was one line: "Replace every adjective with a fact or delete it."

Ask the model to compare the two versions and list what changed. It will name specificity, salary and a low-effort next step, which is a useful way to train your own writing habits as well as the prompt.

Subject lines and the first line you must rewrite by hand

For email outreach, ask for ten subject lines under 45 characters and pick the plainest: "Purchasing lead role, six sites" beats "Exciting career move!". Then rewrite the first sentence of every message yourself. The first line is what appears in the preview, and it is where templated messages give themselves away. It takes ten seconds per message, and it is often the edit that decides whether the message gets opened at all.

Where recruiter prompts go wrong

• Invented achievements. If you paste a vague note, the model may fill it with plausible results ("reduced costs by 15%"). Add "do not add numbers" to every summary prompt and compare to your notes.
• Salary drift. A range of 55 to 60 becomes "up to 65" in an advert. Copy the figure from the brief, not from the output.
• Revealing the client. Role summaries can leak a confidential client through a specific detail such as "the largest bakery group in the region". Read every advert for identifiable details.
• Bias creeping into wording. Phrases such as "young, energetic team" or "native speaker" can appear in drafts. Ask the model to flag any wording that could exclude a group, then judge each yourself.
• One chat for many candidates. Context from one person leaks into another. Start a new chat per task.

Most of these show up in the first month and disappear once the guardrails are in the prompts. Keep a short log of errors found in sent or near-sent drafts; the log shows which rule to add.

A redaction routine that takes three minutes per candidate

For the shortlist prompt you need clean notes. Make this a fixed routine. First, open your own notes, not the CV. Second, write five fields: scope of current role, two measurable results, tenure pattern, salary and notice, fit and concern. Third, delete any name, place-specific detail or employer that could identify the person. Fourth, label them A, B and C. Fifth, paste.

A quick test before pasting: could a stranger who read this note identify the person? If yes, remove more. "Led procurement at a 40-store grocery chain" may identify someone in a small market; "led procurement at a large retailer" does not. After the model returns the summary, replace the labels in your own document and run a last check that nothing in the output goes beyond your notes.

Here is a realistic mistake. A consultant pasted a candidate's full CV into a personal-plan chat "just to summarise it", forgetting that the same chat held two other candidates from earlier. The fix was a rule, not a reprimand: one chat per task, redacted notes only, and a business workspace for the whole team. If you would like to set up that kind of routine with someone, it is the sort of work my AI implementation consultation covers.

How to tell the prompts are working

Track four numbers for a month: minutes spent per intake write-up, outreach reply rate, minutes per shortlist summary, and client questions after a shortlist. If the first and third fall and the second does not drop, the prompts are saving time without hurting quality. If the reply rate falls, your outreach has drifted generic; go back to one verified reason per message.

Review samples every fortnight: pick three sent messages and ask whether they sound like you and whether every fact was true. Retire any prompt whose output you rewrite by hand more than half the time. A short library of eight working prompts beats a folder of fifty untested ones.

Questions recruiters ask before using ChatGPT on candidate work

Can I paste a CV into ChatGPT?

Not on a personal plan. A CV holds a name, contact details, employment history and often more, and OpenAI says personal accounts share data for training by default unless the setting is switched off. Use a business workspace with your client's and candidates' terms in mind, and redact identifiers before pasting.

Does a business plan make candidate data safe to paste?

It improves the position, since OpenAI says Business workspaces are not used for training by default, but it does not remove your duties. Check your privacy notice, your retention rules and any client contract that limits where candidate data may go. Redaction is still sensible.

Should ChatGPT rank or reject candidates?

No. Use it to organise information you already hold and to draft messages. Decisions about who progresses should stay with a person who has read the CV. Automated ranking raises bias and legal questions that vary by place, so take advice before going near it.

How do I stop outreach messages sounding automated?

Give the model one specific reason the person suits the role, taken from public information you verified, and ban generic openers. Keep each message under 90 words and change the first line by hand. Messages that name something real get replies; templated ones do not.

Further reads

• Resume Formatting AI Tools for Recruiters: Check Nothing Was Altered — Format CVs for clients without altering content.
• AI Sourcing Tools for Recruiters: Seat Costs and Consent Rules Compared — Compare sourcing tools on seat cost and consent rules.
• ChatGPT for Recruiters: Can It Prepare a Counter-Offer Call? — Prepare the call when a candidate gets a counter-offer.
• How to Test Job Candidates' AI Skills in an Interview — Test a candidate's AI skills in an interview.
• How to Build Interview Questions and Scorecards With AI — Build interview questions and scorecards with AI.
• Free AI Tool for Recruitment: What a Small Agency Gets for Nothing — What free AI plans and notetakers cover for a small recruitment agency, where free limits and candidate-data risks bite, and when paying makes sense.
• AI for Recruitment Agencies: Re-Engaging Your Candidate Database — A practical sequence for cleaning a candidate database, drafting relevant outreach and recording replies without letting AI make hiring decisions.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI Help Center page on data controls in ChatGPT; the ChatGPT pricing page and OpenAI's ChatGPT Business pricing page (all checked 4 November 2026).