How to Use Claude for Recruitment at a Small Agency: Specs and Outreach

A recruiter can use Claude to draft job specs, adverts, outreach messages and follow-ups, and to tidy interview notes, while a named person alone shortlists and rejects. Do not ask Claude to score, rank or filter candidates. Keep it to admin and communication, on a business plan, with candidate data minimised.

That line is what keeps the workflow useful and defensible. Writing is where a chatbot is strong and a mistake is cheap to catch. Judging people is where errors are costly, hard to see and potentially unlawful. Below are the prompts, the sample outputs, and the five safeguards that hold the line when a busy Friday tempts you to cross it.

Follow me on Instagram@sagnikteaches

What Claude does in an agency's week, and what it never does

Take an illustrative agency with four recruiters filling roles in finance and operations, about 12 live vacancies and roughly 30 placements a year. Where do the hours go? A recruiter might spend 90 minutes on a new job spec after a briefing call, an hour on an advert, three hours a week on outreach messages, and a steady trickle of follow-ups, interview-prep notes and client updates. All of that is language work.

Connect on LinkedInSagnik Bhattacharya

The last row is deliberate. If you want to explore AI in screening, the CV screening tutorial for small agencies and the bias test checklist deal with that separately, with the extra care it needs. This tutorial stays on the admin side.

Subscribe on YouTube@sagnikteaches

Set Claude up for an agency, once

Do four things before the first prompt.

1. Use a business plan. Claude Team has a minimum of two seats and, like other major business plans, does not train on your content by default. A personal free account has different settings and retention. Candidate data belongs on the first, never the second. Read Anthropic's current plan and privacy pages before you buy.
2. Create one Project for the agency. Claude Projects lets you attach standing instructions and reference files that every conversation in the project uses, and on a Team plan you can share a project with colleagues as view-only or editable. Add your house style, a short list of banned phrases, your privacy-notice wording and two or three sample outreach messages that performed well. Do not add any candidate files to this shared project.
3. Write the standing instruction. One paragraph, for example: "You help recruiters draft job specs, adverts and candidate messages. Never score, rank or reject candidates. If asked to, refuse and say a recruiter must decide. Write in plain, clear English. Never invent salary, benefits or company facts; mark gaps with [CHECK]."
4. Agree the data rule. A one-line rule on the wall: names and contact details out unless the task needs them; no health, ethnicity, religion, age or other sensitive details ever go into a prompt.

For more on setting up shared projects around a job rather than a person, see Claude Projects for an office manager or EA, which follows the same one-project-per-job pattern.

Job specs: from a rambling briefing call to a signed-off draft

The hiring manager talks for twenty minutes and the recruiter has a page of rough notes. That is the ideal Claude input, because the notes contain no candidate data at all.

The prompt

Turn these briefing notes into a job spec with these sections: role summary (60 words), responsibilities (6-8 bullets),
essential requirements (max 5), desirable (max 4), working pattern, what the first 90 days look like.
Rules: separate essential from desirable strictly - essential means we would not interview without it.
Remove any requirement that is really a preference about personality or background. Flag anything that could exclude people unfairly.
Do not add salary, benefits or perks unless in the notes. Mark gaps with [CHECK].
Notes: [paste]

Illustrative output for an operations analyst role, trimmed: "Essential: two years' experience reporting on operational data; confident with spreadsheets and pivot tables; clear written English. Desirable: experience in a logistics setting. Flag: the notes say 'young, energetic team player'. Age-related wording can exclude people unfairly, so I have replaced it with 'works well in a small, fast-moving team'. [CHECK] Salary band and hybrid days not stated."

The flag is the useful part. Claude catches wording that a tired recruiter writes without noticing. But treat the flag list as a prompt for your own review, not a clearance. You are still responsible for the final text.

What the recruiter edits

Three things, every time: the essential list (hiring managers inflate it, and a long essential list quietly narrows your candidate pool), the salary and location lines, and anything the assistant "tidied" that changed the meaning. A useful test: read the essential requirements and ask "would we really refuse to interview someone missing this?" If not, move it to desirable. Time check: a spec that took 90 minutes takes about 35, including the hiring manager's sign-off.

A before and after of one requirement

Before: "Must be a digital native with a degree from a good university." After: "Comfortable learning new software quickly; degree or equivalent practical experience." The first excludes career changers and mature candidates and tells you nothing about the work. The second describes what the job needs.

Adverts and outreach that read like a person wrote them

Adverts

Feed Claude the signed-off spec and ask for two versions: a 150-word job-board advert and a 60-word social post. Add "Lead with what the role does and what the person will learn, not with a company boast. Avoid 'rockstar', 'ninja', 'fast-paced' and 'dynamic'." Then read both aloud. Anything you would not say to a candidate on the phone gets cut.

Outreach to passive candidates

Here the data rule bites. The safest input is a short, factual summary you write yourself from a public profile, for example: "Operations analyst at a mid-sized logistics firm for four years; recently posted about route-planning software." No full profile paste, no contact details, no personal facts you would not say to their face. The prompt:

Write a short LinkedIn-style message (under 90 words) from [your name] at [agency] to a candidate for the operations analyst role.
Use only the facts below. Mention one specific, relevant thing from their work. Say what the role is and one reason it might interest them.
Ask one low-pressure question. No flattery, no "I came across your impressive profile".
Facts about candidate: [your summary]
Facts about role: [from the spec]

Illustrative output: "Hi [first name], your post about route-planning software caught my eye. I'm recruiting an operations analyst for a mid-sized logistics client: reporting on delivery performance, with room to shape how the team uses data. If that's close to what you'd like to do next, would a ten-minute call this week be useful?" The recruiter changes one line and sends it. The aim is not to send more messages but to send better ones faster. If you send many, the business development tutorial for agencies applies the same discipline to winning client companies.

A mistake that shows up in outreach

Illustrative: a recruiter pasted a candidate's whole profile and asked for a "personalised" message. Claude wrote that it admired the candidate's "time at a company they left in March" and referenced a volunteer role the candidate had not mentioned to anyone at work. The candidate was unsettled, not flattered. The fix is the summary rule above: only facts you would be comfortable saying aloud, chosen by you.

Follow-ups and rejection messages: draft them, never decide them

Claude is good at the polite, tedious messages: interview invitations with three time options, "your application is with the client" updates, and the hard ones, regret emails. The hard rule: a rejection message is drafted only after a named recruiter has made and recorded the decision against the spec. Claude never suggests who gets one.

Write a respectful regret message to a candidate who reached second interview.
Decision made by [recruiter name] on [date] against the spec; reason category: another candidate's experience was closer on reporting tools.
Do not give a detailed critique. Offer to keep their details on file only if they agree. Under 120 words. Plain English.

Notice that the prompt contains the reason category, not the candidate's name, CV or notes. The human has already decided; the model only writes. Keep a one-line decision record in your applicant tracking system ("Rejected after interview 2: closer match on reporting tools, [recruiter], 13 Nov") and you have both the audit trail and a paper trail for any later question.

Interview notes: tidy, don't judge

You may want Claude to turn a page of scribbled interview notes into a consistent format. This involves candidate data, so use the business plan, remove names and replace with "Candidate A", and keep health, family and other sensitive details out of the notes in the first place. A safe prompt: "Reformat these notes under the headings: examples given, evidence against each essential requirement, questions I still have. Do not recommend a hire or rank. Quote my words, don't add inferences." Then compare it to the notes. The model's temptation is to add an inference ("seems enthusiastic"), which is exactly the kind of unrecorded judgement you do not want in a candidate file.

Briefing the hiring manager with a one-page role summary

Before the spec is signed off, many agencies lose a week to back-and-forth. A cheap fix is a one-page "role summary for sign-off" that Claude drafts from your briefing notes and the hiring manager can read in two minutes. Prompt: "Summarise this role in a table: the problem the role solves, three outcomes by month six, essential versus desirable skills, a realistic candidate profile in two sentences, and three questions I still need answered. Plain words, no recruitment jargon." Illustrative output: "Problem: reporting is done by hand and arrives two days late. Month six: weekly delivery report automated; one process improvement shipped. Questions: is a day in the office required? Is the budget fixed or negotiable for the right person?" The hiring manager answers three questions instead of reading a long spec, and your search starts on firmer ground. Sign-off comes back faster because the questions are the only thing you are asking them to think about.

A talent-pool message for people you have already spoken to

Your own database of past candidates is where an agency's best leads sit, but messaging them is a data-protection matter: you need a lawful reason to contact them and a record of what they agreed to. Where you have a sound basis to get in touch, Claude can help with the wording. Give it the role facts and a category ("finance candidates we spoke to in the last 12 months who asked to be kept informed"), not a list of names, and ask: "Write a short, low-pressure message that says why I am getting in touch, what the role is, and how to opt out of future messages. Under 100 words." Then merge names yourself in your own system, outside Claude. This keeps the names out of the prompt and still saves the drafting.

Answering candidates' questions about your use of AI

Candidates are entitled to ask, and increasingly do. Keep a short answer sheet in your Project and in your privacy notice so every recruiter gives the same reply. An illustrative version:

Do you use AI?  We use AI tools to help draft job adverts, messages and tidy interview notes.
Does AI decide who gets a job?  No. A recruiter decides every shortlist and every rejection.
Is my CV put into an AI tool?  Only on a business-grade account with the supplier's data terms checked, and only where needed for the task.
Can I opt out?  Yes. Tell your recruiter and we will handle your application without AI-assisted drafting.

Only promise what you can do. If your recruiters cannot honour "handle without AI-assisted drafting", do not write it. An honest short answer builds more trust than a polished promise you cannot keep.

When a candidate asks what you hold about them

Data-protection law, including the GDPR where it applies to you, gives people the right to ask what you hold and how it is used. If AI drafts have touched a candidate's data, it matters where that data now sits. A simple rule makes this manageable: never let the assistant be the only place a candidate's information lives. Everything that matters goes into the applicant tracking system or your CRM; chat history in Claude is a scratchpad that you clear on a schedule. Agree a retention period with your adviser, for example deleting working conversations that include candidate details after a short fixed time, and write it down. Then a request becomes a lookup in one system rather than a hunt across chats and personal accounts.

A quick sum for a four-recruiter agency

Say each of four recruiters saves four hours a week on writing, as in the illustrative week above. That is 16 hours a week, about 64 a month. At, say, an internal cost of $25 an hour for recruiter time, that is $1,600 of capacity a month against about $100 a month for four Claude Team Standard seats on monthly billing. Do not read that as a profit figure; the saving only counts if the time goes into calling clients and candidates, the part of recruiting that a model cannot do. If it just disappears into longer lunches, you have saved nothing. Track one measure: calls to candidates and client contacts per recruiter per week, before and after.

Five safeguards that keep a human in charge

1. No scoring, no ranking, no filters. Put it in the project instruction and the staff rules. The instruction makes Claude more likely to decline, but your people are the real safeguard.
2. Named decision-maker. Every rejection has a person's name and a date in your system. "Claude suggested it" is not a decision record.
3. Minimise candidate data. Strip names and contact details unless the task needs them; never paste sensitive categories.
4. Review the wording, not just the facts. Read every draft for exclusionary phrasing and invented claims before it leaves your hands.
5. Tell candidates. Add a line to your privacy notice that you use AI tools to help draft communications and that people make hiring decisions. Data-protection law, including rules on decisions made solely by automated means, is a reason to keep this true. Some places also regulate AI hiring tools specifically; ask a qualified adviser what applies to your clients.

Your clients will also ask. Hiring companies increasingly want to know what an agency does with AI, and the questions clients ask a recruitment agency about AI screening are worth answering in advance with a one-page statement of these five safeguards.

A week with the workflow, with numbers

Illustrative week for one recruiter with three live roles. Monday: briefing call for a new role; spec drafted in 35 minutes instead of 90, saving about an hour. Tuesday: advert and social post, 20 minutes instead of 50, saving half an hour. Wednesday to Thursday: 25 outreach messages, each with one personalised line; about 4 minutes each instead of 9, saving roughly two hours. Friday: 12 scheduling and status messages in 30 minutes instead of 80. The weekly saving is around four and a half hours. That is plausible for a recruiter whose current process is mostly typing; it is a smaller saving if the agency already has strong templates.

Two costs to count: the seats (Claude Team Standard seats were $25 a month on monthly billing or $20 on annual billing as of November 2026, with a two-seat minimum; check the pricing page before you buy), and about an hour a month keeping the project's instructions and examples fresh.

Check the workflow is not drifting

Once a month, pull ten messages Claude drafted and sent. Check each against four questions: did a named person send it, did it contain any claim you could not support, did the wording exclude anyone, and did any candidate data go in that did not need to. Also check one random rejection: is there a named decision-maker and a written reason? If a recruiter admits to pasting a full CV into a personal account, treat it as a data incident, not a telling-off: delete what you can, note what happened, and ask your data-protection adviser whether anything needs reporting.

If you would like to go further and add AI to parts of screening, do that as a separate project with a bias test and a clear review point, and consider the review points for recruitment agents. If you would rather map the whole thing with someone, that is what my AI implementation consultation covers. Either way, the writing is the safe place to begin.

Questions recruiters ask about using Claude

Can Claude shortlist candidates for me?

It can summarise a CV so you read faster, but the shortlist should be yours, made against the criteria in the signed-off spec. Ranking or scoring people with a general chatbot is the step that creates bias and data-protection risk, and it is outside what this workflow does.

Is it safe to paste a candidate's CV into Claude?

Only on a plan whose terms you have checked, with the candidate told how their data is handled, and with contact details stripped where the task does not need them. Free and personal accounts are the wrong place for candidate data. Ask your data-protection adviser if unsure.

Do I have to tell candidates that AI wrote the outreach?

There is usually no rule that every message needs a label, but you must not mislead. If a message claims you read someone's profile closely, you must have. A short line in your privacy notice about AI-assisted drafting is good practice and easy to add.

Which Claude plan should an agency use?

A team or business plan, because those do not train on your content by default and allow shared projects. Claude Team runs from 2 to 150 seats, with Standard seats at $25 a month billed monthly or $20 billed annually. Check Anthropic's current plan and privacy pages before you sign, as terms and prices change.

Further reads

• How to Evaluate the AI Features in Your Recruitment Software — Judge the AI features in the ATS you already own.
• AI Screening: Bolt It Onto Your ATS or Switch Systems? — If you ever add screening, bolt it on or switch?
• AI Recruitment Platform for Employer Branding: What to Look For — What to look for in a platform for employer branding.
• Free AI Tool for Recruitment: What a Small Agency Gets for Nothing — What a small agency gets from free tools, and the limits.
• How to Roll Out an AI Policy So Staff Actually Follow It — Get your recruiters to follow the data rule.
• HR Consulting AI: Sell an AI Policy and Handbook Update Package — What a staff AI policy and handbook update contain.
• How Much Time Does AI Save a Recruitment Agency per Role? — Benchmark the hours saved per role, stage by stage.
• AI Sourcing Tools for Recruiters: Seat Costs and Consent Rules Compared — Five sourcing tools priced for three recruiters, with data sources, consent handling and a one-live-search trial you can run in a week.
• Job Intake Form for Recruiters: Fill It From the Call With AI — A call-to-form method with a filled example, extraction prompt and confirmation email for ambiguous role requirements.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Claude pricing page, Anthropic's privacy article on model training for commercial products and the Claude help article on projects, all read on 4 November 2026; data-protection points are general guidance, not legal advice.