
Every staffing system has a request form. Skills, level, years, start date, duration, location, work mode, priority. It looks reasonable when you design it. Then a delivery manager with three escalations and a client call in ten minutes opens it, sees fifteen fields, and sends RMG a WhatsApp message instead: need 2 Java devs asap, JD attached.
We saw this happen with our own first version. We built a step-by-step chat to create a task, and managers told us straight: it is too long. They already have the client's job description. Why should they type it all again?
The fix is not a shorter form. The fix is to treat the JD as the input, pull out what is in it, and ask the manager only for what is not. This post shows how to do that by hand in about ten minutes per JD, what a parser has to get right if you automate it, and the mistakes that make a request look complete when it is not.
A JD and a staffing request answer different questions
A job description is written to attract a candidate. It sells the company, describes the work and lists skills, often generously. A staffing request is written so that somebody can say yes or no about a specific person on a specific date. The second needs fields that are checkable; the first rarely provides them.
- JDs mix the company pitch, the role, perks and requirements in one block.
- Must-have and nice-to-have are rarely labelled clearly. "Experience with Kafka is a plus" and "strong Kafka" mean very different things.
- Important details are missing. Start date, how many people, timezone overlap and duration are often not in the JD at all.
- Skill names are messy. React, ReactJS and React.js are the same skill. Node and Node.js too. A naive job description parser treats them as three different things.
- Some JDs come as scanned PDFs or screenshots of an email.
The fields a request needs
Whatever system you use, a request that can be matched has roughly these fields. If your form has many more, the extra ones are probably reporting fields that can be filled later by RMG, not by the manager.
- Role title and level, in your company's own level names (for example Senior Engineer, not "5+ years").
- Must-have skills, each with the level needed. Three to five is a healthy number; every extra must-have removes people from the pool.
- Nice-to-have skills, kept separate so they rank people rather than exclude them.
- Client and domain, and whether domain experience is truly required or only preferred.
- Headcount: how many people with this profile.
- Start date, duration and allocation (full time, or a share of a person's week).
- Location, work mode and working-hours overlap, written in hours (for example four hours with UK time).
- Priority, so RMG knows which request to work first when two compete for the same person.
Worked example: one JD, read by hand
Here is a shortened, made-up JD of the kind a client forwards: "We are a fast-growing fintech looking for a passionate backend engineer to join our payments team. You will build scalable microservices in Java and Spring Boot. Strong SQL required. Kafka experience is a plus. AWS exposure preferred. 6+ years. Hybrid, Pune. Immediate joiners preferred."
Read it in three passes. First, delete everything that describes the company or the culture (fast-growing, passionate, join our team). Second, sort what is left into must-have and nice-to-have using the JD's own words: required and strong mean must-have; a plus, preferred and exposure mean nice-to-have. Third, list what is missing.
- Must-have: Java (senior), Spring Boot (senior), SQL (strong). Microservices is a way of working rather than a separate skill; keep it as a note unless your skill list has it.
- Nice-to-have: Kafka, AWS.
- Domain: payments / fintech. The JD does not say it is required, so record it as preferred.
- Level: 6+ years maps to whatever your company calls that band. Write the band, not the years.
- Location: Pune, hybrid. Overlap: not stated, probably Indian hours, but ask.
- Missing: headcount, start date ("immediate" is not a date), duration, allocation, and whether the client will interview.
That is five questions back to the manager, not fifteen fields. The manager can answer them in one short reply, and RMG has a request that can be checked against real people and real free dates.
Normalise skill names before you search
Most bad matches from a good JD come from names, not from people. Keep one short list of canonical skill names and the spellings that map to each: ReactJS and React.js to React, Node and NodeJS to Node.js, Postgres to PostgreSQL, "Spring" to Spring Boot only if your teams use it that way. A spreadsheet with two columns, spelling and canonical name, is enough to start. Add a row every time you find a new spelling. If a JD names a skill that is not on the list at all, do not drop it silently; flag it so someone decides whether it is new or a typo.
Mistakes that make a request look complete
- Copying every technology in the JD into must-have. Clients list their whole stack; a request with nine must-haves usually matches nobody.
- Turning "6+ years" into a hard minimum on years instead of a level. Someone with five strong years at the right level is often the better fit.
- Writing "ASAP" as the start date. Availability can only be checked against a date. If the client really means as soon as possible, write a date two weeks out and say so.
- Leaving out overlap hours for an offshore role, then finding that the shortlisted person cannot join the client's morning stand-up.
- Filling unknown fields with guesses that narrow the search, such as a city when the client never said one. A default should widen the search, never narrow it, and it should be visibly marked as a default.
- Keeping personal data from forwarded emails. Client JDs often carry a recruiter's name, email and phone. They are not needed for matching.
If you automate it: what a parser has to get right
Large language models are good at the first reading step, turning messy text into fields. They are not reliable as the last step. A safe design lets the model propose fields and then checks every field in code: skills against your own skill list, clients against your client list, dates calculated in code rather than written by the model, numbers within sensible ranges, and anything that fails a check dropped or flagged instead of trusted. It then asks the manager only the questions that are left, most important first.
Two simple tests tell you whether a JD parser is safe to use: give it a JD with no start date and see whether it invents one, and give it a JD with a skill you do not have on your list and see whether it silently drops it. Either failure means the request will look complete and be wrong.
How FitRank's task intake handles a JD
In FitRank the manager uploads the JD (PDF, Word or an image) or pastes it, or simply describes the task in plain words. An AI model reads it into fixed fields: title, client, domain, level, must-have and nice-to-have skills, start date, duration, allocation and location limits. Code then checks the fields as described above, and the screen asks only about what is still missing, never more than five questions. Defaults never narrow the search and are marked so the manager can change them. Nothing is created until the manager confirms the live preview.
Emails and phone numbers in the document are masked before the AI reads it, and the file is not kept. FitRank is in early access, so treat this as a description of how the product works rather than a track record. You can see the screens on the task intake page, and how the finished request is matched on how FitRank decides. For the step after this one, turning the request into a shortlist, see why finding the right person still takes days.
More from the blog
Be one of the first teams to staff with FitRank
We're taking on a small number of pilot companies. Pre-register for early access, or talk to us about an annual licence.