How to Write a Job Description That AI Screening and Hiring Managers Can Score Against

A job description used to have one reader: the candidate. Now it has two. Software reads it to decide how to score a few hundred applications, and the hiring manager reads it to decide whether the shortlist is any good. When the description is a list of 20 requirements with no order of importance, both readers guess, and they guess differently. The recruiter then spends the debrief explaining why the top-scored candidate is not the one the manager wanted.

This guide is about writing the description so that it can be scored. It covers what the evidence says about length and wording, how to sort requirements into must-haves and everything else, how to agree weights before the posting goes live, and how to keep the scoring sheet with the posting so the decisions can be explained later. It ends with a worked example for a customer support engineer.

The description now does two jobs

The first job is still to get the right people to apply. Greenhouse's 2025 Workforce and Hiring Report found that 72% of job seekers say a job description did not match the actual role once they got there. That mismatch costs you twice: people who would have been a fit do not apply, and people who do apply leave when the job turns out to be something else.

The second job is newer. With Greenhouse's benchmarks putting the average at 244 applications per job in 2025, most teams run some kind of scoring on the pile, whether that is a recruiter with a checklist or an AI tool. Every scoring method needs a definition of "good" to score against, and the only written definition most teams have is the job description. If it says "strong communicator" and "experience with modern tooling", the scorer has to decide what those mean, and nobody will remember the decision a month later.

Pin's analysis of about 29,000 job descriptions, published in 2026, gives a sense of what scorers are working with. The median description runs about 412 words, and among descriptions that use bullet points, the median requirements list has 20 items. Twenty requirements with equal weight is not a definition of good. It is a wish list.

What the evidence says about length and wording

Three findings are worth keeping in mind before you write anything.

Shorter postings get more applications per view. LinkedIn's analysis of its own job posts, published in 2018, found that posts of 1 to 300 words had an apply rate per view about 8.4% above the platform average, while posts of 301 to 600 words ran 3.4% below it. This is company-reported data from one platform, so treat it as a direction rather than a target. The direction is clear enough: the long requirements list that makes a hiring manager feel thorough is the part candidates stop reading.

Masculine-coded wording shrinks the applicant pool. Joyce He and Sonia Kang published four studies in PNAS in February 2025, covering 37,920 participants. In 576 real job ads with 32,834 applicants, women made up 40.0% of applicants to ads one standard deviation below the mean for masculine language and 34.1% of applicants to ads one standard deviation above it. In a field test on a live ad for a male-dominated role, replacing words like "dominant" and "competitive" with neutral synonyms raised the share of women applying by about four percentage points, and the effect held ten months later. The replacement words were synonyms, so the job did not change. Only the pool did.

Rigid requirements screen out people who could do the job. Harvard Business School and Accenture surveyed 2,275 executives in the US, UK and Germany for the 2021 report Hidden Workers: Untapped Talent. 88% said qualified, high-skill candidates are screened out of their process because they do not match the exact criteria in the job description, such as holding a four-year degree. For mid-skill roles, 94% said so. The report is not about AI rejecting people on its own. It is about the criteria recruiters typed into the filter, which is the part you control.

Step 1: Run the intake as a scoring conversation

The structured interview scorecard guide describes a 45-minute intake meeting that defines what good looks like for the role. Hold the same meeting before you write the posting, and add one question to every requirement the hiring manager raises: "If a resume shows this, how much should it count?"

Three answers are possible, and they map to three lists.

• "Without it, we would not interview them." That is a must-have.
• "It matters, and more of it is better." That is a weighted requirement.
• "It would be nice." That is a nice-to-have, and it goes at the bottom of the posting or nowhere.

Most managers start with eight to twelve must-haves. Push back on each one with two tests. Can a capable person learn it in their first 90 days? Would the job fail without it on day one? If the answer to the first is yes or the second is no, it is not a must-have. Three to five survive in most roles. The ones that fall off are not wasted; they become the weighted requirements that separate strong candidates from adequate ones.

Write the result down as a table before anyone writes prose. The table is the scoring sheet. The posting is a view of it.

Step 2: Write must-haves as evidence a resume can show

A must-have only works if a scorer, human or software, can find evidence for it in a resume. "Excellent problem solver" cannot be found. "Has debugged production issues for a software product used by external customers" can be.

For each must-have, write:

• The skill in plain words.
• The adjacent names it goes by. A resume that says "Zendesk" and one that says "Freshdesk" both show help-desk tooling experience. If your posting names one product, a strict scorer rejects the other.
• The minimum level, stated as something a person has done rather than years. "Has owned a customer escalation end to end" is a level. "Three years" is a proxy for a level, and a poor one.

Be careful with years and degrees. Years of experience is where age bias enters quietly; a cap like "two to four years" tells candidates over 40 not to apply. Degree requirements are where the hidden workers from the HBS report went missing. The Burning Glass Institute's 2022 report, The Emerging Degree Reset, found that 46% of middle-skill occupations and 31% of high-skill occupations saw employers drop degree requirements from a material share of postings between 2017 and 2019. If a degree is a genuine must-have for the role, say which degree and why. If it is a proxy for "can write and reason", test for that directly.

Step 3: Agree the weights before the posting goes live

A scoring sheet with must-haves and weighted requirements still needs to say how the overall score is built. Most teams find that four areas cover almost every role: skills, experience, projects and education. The question for the intake meeting is how much each area should count.

The weights change the shortlist more than people expect. A support engineer role weighted 40% skills, 30% experience, 20% projects and 10% education produces a different top ten from the same role weighted 25/25/25/25. The first favours people who have the specific tools; the second favours generalists with a degree. Neither is wrong, but the hiring manager should choose, and the choice should be recorded with the date.

Two rules keep the weights honest. First, decide them before you see any applications, because weights chosen after you have a favourite candidate are a rationalisation. Second, keep the must-haves outside the weights. A candidate who is missing every must-have should not reach the top of the list because they scored highly on education. The must-have list is a floor; the weights sort the people above it.

Step 4: Cut the posting to what a candidate needs

With the scoring sheet done, the posting is the easy part. Four blocks, in this order, in 300 to 500 words:

1. What the job is for. Two or three sentences on what this person will own and what a good first year looks like. This is the block the 72% of job seekers who felt misled were missing.
2. What you need to bring. The three to five must-haves, written as the evidence statements from Step 2. Label the block "Required" and mean it.
3. What helps. The weighted requirements, in order of weight. Label it "Helpful" or "We would value". Candidates who meet two of five should still apply, and the label should say so.
4. Practical details. Location, hours, pay range where the law requires it and ideally where it does not, the stages of the process and how long each takes.

Then read the whole thing once for wording. The PNAS study used a published list of masculine-coded words; "competitive", "dominant", "aggressive" and "rockstar" are the usual suspects, and every one has a neutral synonym that means the same thing. The reason to replace them is not politeness. It is that the pool gets bigger and the job does not change.

Step 5: Keep the scoring sheet with the posting

The scoring sheet is the record of what "qualified" meant for this role on the day you opened it. Save it next to the posting, with the date and the names of the people who agreed it. When a candidate asks why they were not shortlisted, the answer is in the sheet. When an auditor asks how the AI decided, the answer is also in the sheet, because the AI was scoring against it. The explainable AI resume screening guide covers why that record matters in more detail, and the AI hiring laws guide covers which jurisdictions now expect it.

If the role changes mid-search, update the sheet, keep the old version, and note the date. A shortlist built against last month's weights should be explainable against last month's weights.

A worked example: customer support engineer

A 200-person software company opens a role for a support engineer who handles technical escalations from business customers. The intake meeting produces this sheet.

Requirement	Type	Evidence a resume can show
Has debugged customer-reported issues in a software product	Must-have	Support, QA or engineering role with customer-facing troubleshooting
Can read application logs and basic SQL	Must-have	Any role naming log tools or SQL; a project that used them
Has written explanations for non-technical readers	Must-have	Help-centre articles, runbooks, customer updates, documentation
Experience with help-desk tooling (Zendesk, Freshdesk, Intercom or similar)	Weighted, high	Tool named, or a role that implies one
Experience with the company's stack (web APIs, Postgres)	Weighted, medium	Named in a role or project
Has worked in a company selling to businesses	Weighted, low	Employer type
Degree in computer science or related	Nice-to-have	Education section

Weights: skills 40%, experience 30%, projects 20%, education 10%. The three must-haves form the floor; a resume that shows none of them cannot rank as a strong match regardless of its degree. The posting runs to about 380 words: a paragraph on the escalation queue and what the first year looks like, three required lines, three helpful lines, and the practical details, including that the process is a 30-minute call, a take-home debugging exercise and one panel interview, over about three weeks.

A month in, the hiring manager says the shortlist is too heavy on people from large companies. The team looks at the sheet, agrees that "worked in a company selling to businesses" was pulling in enterprise support staff, lowers that weight, records the change with the date, and re-sorts. Nobody rewrites the posting from memory.

How this maps to HireRabbit.AI

The method above does not depend on any particular tool, but it is the method HireRabbit.AI is built around. Every resume gets four area scores for skills, experience, projects and education, and the overall score is computed from those four with the weights you set, so the score always matches the breakdown. Must-have skills act as a floor: a resume that shows none of them cannot reach a strong match, however much else it has. The scoring sheet, the weights and the cut-off stay with the role, and any move of a single candidate carries a written reason. The how scoring works page walks through the mechanics.

Sources

• Greenhouse, 2025 Workforce & Hiring Report: https://www.greenhouse.com/blog/greenhouse-2025-workforce-hiring-report
• Greenhouse, Hire Standard recruiting benchmarks: https://www.greenhouse.com/recruiting-benchmarks
• Pin, What Makes a Good Job Description in 2026? 29,000+ JDs Analyzed: https://www.pin.com/blog/job-descriptions-analyzed-study/
• LinkedIn Talent Solutions, New job post stats (2018): https://business.linkedin.com/talent-solutions/blog/job-descriptions/2018/new-job-post-stats
• He and Kang, Debiasing job ads by replacing masculine language increases gender diversity of applicant pools, PNAS (Feb 2025): https://www.pnas.org/doi/10.1073/pnas.2409854122
• Fuller, Raman, Sage-Gavin and Hines, Hidden Workers: Untapped Talent, Harvard Business School and Accenture (Sep 2021): https://nahr.shrm.org/sites/default/files/hiddenworkers09032021.pdf
• Burning Glass Institute, The Emerging Degree Reset (Feb 2022): https://www.burningglassinstitute.org/research/the-emerging-degree-reset