Job Brief: How to Write One So the AI (and Your Recruiters) Find the Right Candidates

When a search stalls — no candidates, or the wrong ones — the blame goes to the tool, the market or the hiring manager. The cause is almost always upstream: the job brief. It is the one document everything else reads — the recruiter, the semantic engine, the matching score, the outreach message. A vague brief produces a vague shortlist, whatever the quality of the AI behind it. This guide explains what separates a brief from a job description, the seven elements that make a brief usable by a machine and a human, how to phrase it in natural language for a semantic engine, and how to get it out of a hiring manager in twenty minutes.

A job description and a sourcing brief are not the same document

The job description speaks to the candidate: it sells the role, describes the company, lists the responsibilities. The brief speaks to whoever is searching: it describes the candidate, not the role. The two overlap on skills, but their logic is reversed. The description says "here is what you will do"; the brief says "here is how you recognise the right person, and how you recognise the wrong one". Confusing them is the original error: pasting a job description into a sourcing engine is asking it to find someone from a brochure.

The seven elements of a usable brief

1. Context in two sentences. The team, the problem the person will own, why the role is opening now. This is what lets the recruiter — and the model — understand the type of profile, beyond keywords.
2. Three verifiable must-haves, no more. "Three years of Kubernetes in production" can be checked on a profile; "strong on infra" cannot. Beyond three, you are no longer describing a candidate, you are describing an empty intersection.
3. Nice-to-haves, explicitly separated. What tips the balance between two good profiles but disqualifies no one. Mixing them with must-haves is the number one cause of shortlists that come up short.
4. Named dealbreakers. What excludes without discussion — a mandatory language, a clearance, a geographic perimeter. These are the only criteria applied as a hard filter; everything else is weighted.
5. Compensation, location, work mode. A real range, a city, the number of days on site. Without them, the recruiter contacts people who will say no at the third message, and the matching score evaluates a fit that does not exist.
6. Timing signals. Typical tenure in the current role, expected trajectory (second senior role, first lead position…), availability signals. This is what turns "relevant profile" into "relevant profile who will reply".
7. Two reference profiles and one counter-example. Two people — anonymised or public — who embody the right profile, and one who looks like a match but is not. Three examples calibrate a semantic engine better than thirty adjectives.

The natural-language brief for a semantic engine

A semantic engine does not read a bulleted list the way a recruiter does: it reads text and builds a representation of its meaning. The brief that works is therefore a paragraph of three to six sentences, written the way you would describe the role to a colleague, not a table of criteria. Here is the form that gives the best results:

"We are looking for a senior Data Engineer to take over the data platform of a 120-person fintech in London, after the lead's departure. Must-have: at least three years on Spark or dbt in production, one end-to-end migration to a data lakehouse, and professional English. Nice-to-have: Airflow, time at a scale-up, a first mentoring experience. We exclude profiles without production experience and those who cannot be in London two days a week. Range £80-95K. The profiles that respond best are typically two to four years into their current role, in a data team of five to fifteen people."

Three things to avoid in that paragraph: hollow adjectives ("dynamic", "passionate", "rigorous") that the model cannot attach to anything verifiable; lists of fifteen technologies that dilute the signal instead of sharpening it; and contradictory requirements — "junior with ten years' experience", "autonomous but closely supervised" — which the engine resolves by picking at random. This is the brief that then serves as the opening query in the method we describe for combining semantic and Boolean search.

The twenty-minute intake with the hiring manager

The brief is not written alone at a screen: it is extracted from a hiring manager who, often, has not yet articulated what they want. Twenty minutes and eight questions are enough, provided you ask them in this order:

1. What problem must this person have solved in six months for you to be satisfied?
2. Who, in your current or past team, would do this role perfectly — and what characterises them?
3. Whom have you seen fail in this kind of role, and why?
4. If you could keep only three requirements, which ones?
5. What, on a CV, would make you say no in ten seconds?
6. What compensation range have you actually had approved — not the one you hope for?
7. How many days on site, and is it negotiable for the right profile?
8. Which companies do these profiles work at today?

The answers to questions 2, 3 and 8 are the most valuable: they provide the reference profiles, the counter-example and the target pools — exactly what an engine needs and what a job description never contains. An AI copilot can structure that debrief into a brief on the fly; it is one of the everyday uses we describe in what an AI copilot changes for recruiters.

The most expensive mistakes

• The inventory brief. Twelve must-haves, twenty technologies, four languages. Nobody exists, and the engine surfaces the profiles that tick the most boxes — that is, the wordiest CVs, not the best.
• The frozen brief. The first twenty profiles always teach something — an unrealistic criterion, a title the market does not use, a range that is too low. A brief that is not revised after that first read stays wrong until the end.
• Hidden criteria. The hiring manager wants "someone senior" and rejects every profile over forty-five. The real criterion is illegal and unspoken; the official brief is useless. The intake exists precisely to surface those criteria so they can be discarded.
• Missing compensation. A brief without a range loses on average one message in three — the candidate asks about salary, you answer, they decline.
• The ownerless brief. Written by the recruiter alone, it reflects what the recruiter understood. Signed off in one line by the hiring manager, it becomes a shared contract against which the shortlist can be measured.

Writing a job brief in six steps

1. Run the intake before writing. Twenty minutes, eight questions, raw notes.
2. Sort into three columns. Must-have, nice-to-have, dealbreaker. Anything that fits no column leaves the brief.
3. Rephrase each must-have as a verifiable criterion. A duration, a technology, a deliverable, a context. If you cannot check it on a profile, it is not a criterion.
4. Add the frame. Range, location, days on site, timing signals, target pools.
5. Write the natural-language paragraph. Three to six sentences, two reference profiles, one counter-example. This is the version the engine reads and the hiring manager signs off.
6. Revise after the first twenty profiles. Hold the first shortlist against the brief, fix what the market contradicts, date the version.

A good job brief does not make the AI smart: it gives it something smart to read. It is the highest-return investment in sourcing — twenty minutes that decide the quality of the shortlist generated in ten minutes and the relevance of the matching score that ranks it. Want to see what EMILY does with a well-written brief? See how EMILY turns a brief into a shortlist.