Seven recruitment tasks to hand to Claude Cowork first
TL;DR
Start with company research, it is the safest and most immediately useful. Then move to intake notes, reporting, and document piles, all high repetition and low judgment. Save each working task as a skill so the second run costs you one sentence. Keep candidate-facing work for later and keep a human deciding. One task properly automated beats five half-finished experiments.
The hardest part of using an AI agent is not the technology. It is knowing which task to give it first. Pick something too big and you spend an afternoon fighting it. Pick something too trivial and you learn nothing.
These are the seven tasks that work best as starting points for recruiters, in the order we recommend trying them, with the instruction to use and what a good result looks like.
If you are new to the tool itself, read what is Claude Cowork first. This post assumes you have it open.
1. Company research before a pitch
Why start here. It uses only public information, so there is no candidate data and no compliance knot to untangle. It is also the task where the output is instantly, obviously useful.
The instruction. "Research this company before my business development call. Tell me what they do, whether they are hiring and for what, any recent trigger events like funding or leadership changes, the angle I should pitch, and two opening lines. List your sources."
What good looks like. A one-page brief you can read in two minutes and walk into a call with. Not a wall of text about the company's mission statement.
Watch for. Confidently stated facts with no source. Always ask for sources and spot-check one.
2. Messy intake notes into a structured brief
Why it is next. This is where the agent earns its keep, because the input is genuinely messy and lives across formats: an email from one manager, a voice note you transcribed, a pasted chat thread.
The instruction. "Read every intake file in this folder, build one overview spreadsheet with a row per role including the title, must-haves, salary range, location and urgency, and write a clean role brief for each role. Save everything back in this folder."
What good looks like. A spreadsheet you can open and a brief per role, both saved to disk, without you having copied anything.
Watch for. Gaps the notes never contained. A good brief flags "salary not discussed" rather than inventing a range.
3. The weekly report you rebuild from scratch
Why it matters. It is pure repetition, it happens every week, and it is exactly the kind of work that quietly eats a Friday afternoon.
The instruction. "Take my notes in this folder and build this week's report in the same structure as last week's file. Keep the section order, summarise what moved per role, and flag anything that has gone quiet."
What good looks like. A report in your own format that needs a light edit, not a rewrite.
Watch for. Padding. If nothing happened on a role, the report should say so plainly.
4. Working through a pile of documents
Why it works. Volume is where an agent is simply better than a human with a coffee. Contracts, policies, vendor documents, a folder of long PDFs.
The instruction. "Go through every document in this folder and give me a summary per document with the key points, plus one combined overview of the themes and anything that looks like a risk."
What good looks like. A summary you can act on, with the source document named for each point so you can verify.
Watch for. Confidential material you did not mean to include. The working folder is the boundary, so put only what belongs in it.
5. Candidate communication drafts
Why it comes later. Now you are touching candidate context, so the guardrails start to matter. This is drafting, not deciding.
The instruction. "Draft a status update for each candidate in this list, using my notes on where they are in the process. Keep it short and warm, and leave the tone consistent with the examples in this folder."
What good looks like. Drafts you review and send, never messages that go out on their own.
Watch for. Anything that reads like a decision about a person. Rejections in particular deserve your own eyes and often your own voice.
6. Preparing a structured shortlist
Why it is sixth, not first. This is the loaded one. Screening and selection are high-risk use cases under the EU AI Act, and CVs are personal data under the GDPR.
How to do it responsibly. Three rules we teach in every training:
- Allowed and anonymised. Only use data you are permitted to use, and strip what you do not need.
- Human in the loop, always. Claude proposes, you decide.
- Judge the role, not the person. Score against the requirements, never on a name, age, gender or background.
The instruction. "Compare these CVs against the requirements in the vacancy file. For each one, list the evidence for and against on the stated requirements only, and note what is missing. Do not rank or recommend anyone."
What good looks like. Evidence you can check, not a verdict you are tempted to trust. Full detail in Claude Cowork, the GDPR and the EU AI Act.
7. Anything you have now done three times
The real seventh task is not a task. It is the moment you notice you have run something three weeks in a row. That is the signal to stop retyping the instruction and save it as a skill, so next week it costs you one sentence.
This is also the point where most recruiters feel the change. Not during the first impressive run, but on the fourth Monday when the work is already done. See Claude skills for recruiters.
The order in one table
| # | Task | Repetition | Judgment involved | Start here? |
|---|---|---|---|---|
| 1 | Company research | High | Low | Yes, safest first win |
| 2 | Intake notes to brief | High | Low | Yes |
| 3 | Weekly report | Very high | Low | Yes |
| 4 | Document piles | Medium | Low | Yes |
| 5 | Candidate drafts | High | Medium | After the first four |
| 6 | Structured shortlist | High | High | Only with guardrails |
| 7 | Save it as a skill | n/a | n/a | As soon as a task repeats |
The pattern behind the order: automate the moving of information, never the judging of people.
How to actually run your first one
- Make a folder and put only the files this task needs in it.
- Open Cowork and point it at that folder as the working folder.
- Paste the instruction, describing the finished result rather than the steps.
- Approve the steps and watch the first run so you can steer.
- Open the output, correct what is off, and note what you would say differently next time.
- When it works, save it as a skill.
That loop is the whole method. Everything else is a bigger version of it.
If you would rather do this guided, on your own real task instead of an example, that is what the AI Recruitment Engineer Starter is for. About four hours, self-paced, and you finish with one task genuinely off your plate. It is 297 euro, and it includes a plugin with ten ready-made recruitment skills.
For the wider view on what to automate across your process, see the recruitment process automation guide.
Frequently asked questions
What is the best first task to try in Claude Cowork? Company research before a business development call. It uses only public information, so there is no candidate data involved, and the output is immediately useful.
How long does it take to set up a task? The first run of a simple task takes ten to twenty minutes including corrections. Once you save it as a skill, later runs take one sentence.
Can Claude Cowork read PDFs and spreadsheets? Yes. It works on the files in your working folder and can produce spreadsheets and documents as output, saved back to that folder.
Should I use it for CV screening? Only with guardrails. Screening is high-risk under the EU AI Act. Use it to structure evidence against stated requirements, keep a human making every decision, and never select on protected characteristics.
What if the output is wrong? Correct the instruction rather than arguing with the result. Most bad output comes from an instruction that described the steps instead of the finished result.