Magai

How to Use AI to Write Job Descriptions, Week First

How to use AI to write job descriptions from a hiring manager's notes: build the posting around a real week, compare models, and audit every line.

By Dustin W. StoutPublished 20 min read
A dispatch desk at dawn with a headset on a stand, a paper route board with colored magnets and a mug of coffee beside a window

Your hiring manager sent you a job title and a link to the last posting.

That is a label, and it is the reason most people learning how to use AI to write job descriptions end up with the same posting everyone else has.

The fix is better notes, not a better prompt. Get the hiring manager's raw notes about a real week, paste them into one Magai thread, and tell the model it may use nothing else. Have it build the posting around that week. Then put the same instruction to two or three models, keep the best lines, and check every sentence against the notes. Magai keeps the thread whole while you switch models, so nothing gets pasted twice.

A job description is a forecast of the week, and a model can only write the forecast if the notes describe the week.

You have four open roles and four managers who each gave you a title and "see attached." You do not need a faster way to stare at a blank page. You need the part of the job that lives in a manager's head, pulled out and put in order.

AI is very good at the putting in order. It is bad at the pulling out, and it covers the gap with whatever sounds like a job description. That is how a dispatch role ends up "high-pressure" and "dynamic" in a posting that never says who answers the phone at 7:00.

The rest of this post runs one role from the first call to the last check: a dispatch coordinator at Harbor Mechanical, a made-up commercial heating and cooling company in Tacoma, with a made-up service manager named Marcus. The notes, prompts and replies are the kind that land on your desk. Every step ends in something you can paste into your own thread on Monday.

Get the week out of the hiring manager first

The old posting is the wrong place to start. It describes the last person's job, written for the last hiring cycle, and a model will copy its shape faithfully. You want a fresh account of a real week in the manager's own words, and the cheapest way to get it is a fifteen minute call.

Ask the questions below. Type the answers as the manager talks, or paste in the transcript your meeting tool already made. Do not tidy it. Rough notes beat a clean form, because the odd specific detail (the radio app, the Saturday rotation, the person who just left) is the sentence a candidate remembers.

Ask them in this order:

  1. What does the first hour of a normal day look like?
  2. What is the busiest day or season, and what happens to this person then?
  3. What is the hardest decision this person makes alone?
  4. Which tools do they live in all day?
  5. Who do they talk to, and who talks to them?
  6. What will you teach on the job, and what must they arrive knowing?
  7. Where did the last person in this seat go?
  8. What is the exact pay range, schedule, location and benefit start date?
  9. What would make someone good at this in six months?
  10. What would make someone quit in the first month?

Questions 2, 3 and 10 do the most work. They are the ones a manager has never been asked, and the answers are the lines a candidate will quote back to a friend.

Managers also hand you adjectives, and every adjective hides a question. The table shows four common ones, the question that opens each up, and the sentence it lets the posting say.

What the manager says The question to ask What the posting can say
"High-pressure" What happens at 7:00 on a Monday? Calls start around 7 and Monday is the heaviest day.
"Strong communicator" Who is the hardest person they talk to? You will take calls from customers whose heat is out.
"Self-starter" What do they decide without asking you? You decide which no-heat call gets a truck today.
"Team player" Who do they hand work to and take work from? You work with 14 technicians by radio and ticket.

Here is what Marcus dictated for the dispatch role, typed as he said it:

Need a dispatch coordinator. Replacing Dana, she moved to scheduling installs.
Reports to me, Marcus, service manager. 14 techs, 11 trucks, commercial buildings.
Day is 6:30 to 3, Monday to Friday. One Saturday a month on the on-call rotation, time and a half.
First calls hit around 7. Mondays are the worst, phones triple. First cold snap of the year
is worse. You're deciding who gets the no-heat call today and who waits till tomorrow.
Tools: ServiceTitan for tickets, a radio app called Zello for the techs.
Doesn't need HVAC experience, we teach the vocabulary. Needs dispatch, scheduling, or a
front desk with heavy phones. Has to stay calm with an angry customer.
Spanish helps, maybe a quarter of calls.
Pay 24 to 28 an hour. Overtime during cold snaps. Health benefits after 30 days, 401k match 3 percent.
In the office in Tacoma. Not remote.

A recruiter's desk with rough hiring manager notes and a phone, the starting point for how to use AI to write job descriptions

Those notes are messy on purpose. Before you move on, underline two things in your own copy: every number the manager gave you, and everything that is an impression rather than a count. "Phones triple" is an impression. Marcus did not count the calls. You will need that underline in the audit step.

Set the rules in the thread before you ask for a draft

Open a new thread in Magai and name it for the role. Do not paste the notes yet. Start with the rules, because a model that has not been told what to avoid will fill every gap with the most common sentence in its memory.

A bare request for a dispatch coordinator posting tends to open with "We are seeking a detail-oriented candidate to join our dynamic team," and to add requirements nobody asked for, like a degree. Nothing in Marcus's notes mentions a degree. The model adds one because most postings for most roles have one.

Paste this first:

I'm a recruiter writing a job posting from a hiring manager's notes.
These rules apply to this whole thread.

1. Use only facts that appear in the notes I paste. Do not invent
   responsibilities, qualifications, benefits, pay, tools or team details.
2. If the posting needs a fact the notes do not give, list it under
   "Gaps" at the end. Do not fill it in.
3. Keep must-haves and nice-to-haves in separate lists. A must-have is
   something the notes say the person needs on day one.
4. Put the pay range and location in the first five lines, exactly as
   the notes give them.
5. Banned phrases: "we are seeking", "dynamic", "high-pressure", "rockstar",
   "self-starter", "team player", "passionate", "excellent communication
   skills". Say what the work is instead.
6. Write in plain words, to "you". No exclamation marks.

Reply "ready" and nothing else.

Each rule closes a specific hole. Rule 1 stops invented duties. Rule 2 turns a model's urge to fill a gap into a list of questions for the manager, which is the most useful thing it will produce all day.

Rule 3 keeps a nice-to-have from quietly becoming a requirement that turns qualified people away. Rule 4 puts the number where a candidate looks first. Rule 5 bans the phrases that make a posting interchangeable with a thousand others, and it works because the model has to replace each one with a fact.

Make rule 5 yours. Pull the last ten postings your company put out, circle the phrases you would not say to a candidate's face, and add them to the list. The banned list is the one part of this prompt that should differ from company to company.

This is the first place Magai earns its keep in this job. The rules live in the thread. Every model you bring in later reads them and follows them, so you set them once and they hold across the whole comparison in the fourth step. If you have ever pasted the same style guide into three browser tabs, you know what you are saving.

The same move works anywhere a model must stay inside a set of facts. It is how a real estate listing description stays honest, and how a consulting proposal stays inside what the discovery call produced.

Use AI to write the job description as a week, then fix the padding

Now paste the notes under a structure. Here is the first version of that request, exactly as typed:

Here are the hiring manager's notes for the role. Write the posting.

Structure:
- Title, location, pay range
- Two sentences that open with the work this person owns
- "A week in this job": Monday through Friday, plus the Saturday rotation
- Must-haves
- Nice-to-haves
- What we offer (only what the notes state)
- Gaps

[paste Marcus's notes]

The reply opened well:

Dispatch Coordinator, Harbor Mechanical Tacoma, WA, in the office. $24 to $28 an hour. Overtime during cold snaps.

You decide who gets a truck today. Harbor Mechanical runs 14 technicians and 11 trucks across commercial buildings, and every call, ticket and radio message about those trucks lands on you first.

That is a good opening. It leads with the decision, carries the numbers, and says nothing about a dynamic team.

The week section is where it went wrong:

Monday, 6:30: You start before the phones do. First calls come in around 7, and Monday is the heaviest day, when the phones triple. Tuesday through Thursday: Steady days of scheduling follow-up visits, returning customer callbacks and updating tickets in ServiceTitan.

Marcus said nothing about Tuesday, Wednesday or Thursday. The prompt asked for Monday through Friday, the notes covered Monday, a cold snap and a Saturday each month, and the model filled the middle of the week with plausible work. Rule 1 did not catch it, because the structure told the model to produce five days and the notes could not supply them.

There were two other problems. "The phones triple" came through as fact, and it is Marcus's impression. And the draft listed "calm under pressure" as a must-have, which is fair, but it hung that phrase on its own with no example of what the pressure is.

The fix is to change the instruction so the structure cannot demand what the notes lack. Send this:

Rewrite the "A week in this job" section. Only describe days and
situations the notes actually cover. For any weekday the notes do not
describe, say nothing about it, and add it to Gaps instead. Where the
notes give an impression rather than a count (like "phones triple"),
describe it as the manager's view or soften it to what the notes
support. Replace "stay calm" with the situation it refers to.

The rewritten section came back shorter and better:

Monday, 6:30: You start before the phones do. First calls come in around 7, and Monday is the heaviest day. When the first cold snap hits: The calls pile up and you decide which no-heat call gets a truck today and which waits until tomorrow. Overtime applies. One Saturday a month: You take your turn on the on-call rotation, paid at time and a half. Every day: You work in ServiceTitan for tickets and Zello for radio, and you take calls from customers who are angry because their heat is out.

The Gaps list at the bottom now carried what the manager still owes you:

Gaps: Is the Saturday rotation required or voluntary? What is the overtime rate on weekdays during a cold snap? Is there a start date? What does a normal Tuesday through Thursday look like?

Read that last question again. The model was right to ask it. A posting that cannot describe Tuesday has told you the manager has not finished thinking about the job.

If a hiring manager cannot describe a week, no model fixes that. The notes come back thin, the Gaps list runs long, and your job is to send the list back to the manager. That is the system working, and it is a better outcome than a polished posting for a job nobody has defined.

Two printed postings side by side, one a grid of identical gray blocks and one with varied hand-marked lines

Run the same draft through two or three models and keep the best lines

You may be thinking that candidates already spot an AI posting from three lines away. They do, when it is built from adjectives. A posting built from Marcus's Monday has nothing for a reader to spot, because the model had nothing to embroider.

Models still differ in what they smooth over. One tends to keep the odd specifics. Another tends to round them toward something more general. You do not know which will do which on your notes until you run them, and reading three drafts takes less time than rewriting one weak one.

In Magai, this is one thread. You do not open three tabs or copy the notes three times. Name each model in the message with a slash, and the thread carries the rules and the notes to every one of them:

/Claude Write the posting from the notes and the thread rules only.
Do not copy phrasing from any earlier draft in this thread.

Then the same message for another model:

/GPT Write the posting from the notes and the thread rules only.
Do not copy phrasing from any earlier draft in this thread.

And a third, if you want one, with /Gemini. The last line matters. The thread is shared, so a second model can see the first model's draft, which is a gift for the stitching pass and a contamination problem for the comparison. Telling each one to ignore the others keeps the drafts independent.

Score the three drafts on the same short list. This table gives you the five checks, what to look for, and what a pass looks like:

Check What to look for A pass
Numbers Pay, headcount, hours and schedule copied exactly Every figure matches the notes
The week Only days and situations the notes cover Gaps lists the days it skipped
Invented requirements Degrees, years of experience or tools nobody mentioned None
Opening Starts with the work the person owns No "we are seeking"
Impressions Manager's guesses labeled or softened "Phones triple" is not stated as fact

Do not pick a winning draft. Pick winning sentences. Take the opening from one, the week from another, and the benefits block from whichever kept the numbers straightest. Then hand the pieces to one model to stitch:

/Claude Combine these pieces into one posting. Opening from draft 1.
"A week in this job" from draft 2. Benefits from draft 3. Change
nothing else and add no facts. Keep the Gaps list.

This is what the comparison buys you. You choose the answer, not the vendor. The three models share one conversation, so the stitched posting is built from the same notes and rules all the way through. For a wider test of which model suits which writing job, the five-task method for choosing a chatbot works on the same principle, and switching between models in one workflow covers the mechanics.

Cost matters when you run three models on every role. Magai plans start at $20 a month, and the pricing page lists what each plan includes. Separate subscriptions to the big chat apps add up to about $100 a month, and none of them remembers what you told the others.

Three printed drafts of the same document fanned across a desk with one line circled in pen

Audit every line against the notes before it goes out

A stitched draft reads smoothly, and smooth is dangerous. Every sentence now sounds equally sure of itself, including the ones nobody said.

Make the model check its own work, and make it show the evidence. Send this:

Audit the final posting. Make a table with one row for every factual
claim in it: pay, hours, schedule, tools, benefits, duties, requirements.
Columns: Claim, Quote from the notes that supports it, Status.
Status is one of: Supported, Impression (the manager's view, not a
count), or Not in notes. Do not rewrite anything yet.

The table is the useful artifact. Here is how three rows came back:

Claim Quote from the notes Status
$24 to $28 an hour "Pay 24 to 28 an hour" Supported
Monday is the heaviest day "Mondays are the worst, phones triple" Impression
Health benefits start after 30 days "Health benefits after 30 days" Supported

A claim marked Not in notes comes out. A claim marked Impression gets confirmed with the manager or softened. Ask the model to rewrite only after you have read the table.

This pass catches what you will not. A model that wrote the sentence will defend it on a casual reread, and a model asked to find the supporting quote cannot invent one without it showing. If a row has no quote, the claim has no source.

The notes cannot tell you whether the posting is lawful or fair, so the second half of the audit is yours. Treat what follows as the paperwork side of the job and not as legal advice. Your employment counsel or your state labor department has the final word.

Pay ranges come first. A growing number of states require the range in the posting itself. Colorado's Equal Pay for Equal Work Act is the long-standing example, and California, New York and Washington have their own versions with different employer thresholds. If your role could be filled by someone in a covered state, check the rule before you post. Do not ask a model what your state's law says and treat the answer as settled. A model's description of a statute is a lead to verify at the source.

Age language comes next. Federal regulations say help wanted notices may not contain terms that limit or deter older applicants, and they name "young," "college student" and "recent college graduate" as examples. The text is in 29 CFR 1625.4. A model asked for an "energetic" posting will sometimes reach for those words, so search your draft for them.

Gendered wording matters as well. Gaucher, Friesen and Kay found in 2011 that job ads written with more masculine-themed words, such as "leader," "competitive" and "dominant," made the jobs less appealing to women and lowered their sense of belonging, as reported in the Journal of Personality and Social Psychology. You can ask a model to flag those words, then decide whether the job really needs them. Most of the time the concrete version of the sentence does the same work.

Last, requirements. Federal rules under the Americans with Disabilities Act treat a written job description prepared before the role is posted as evidence of what a job's essential functions are, which you can read in 29 CFR 1630.2. That is one more reason the Saturday rotation line has to be true. If the posting says the person works one Saturday a month, that line may later be read as part of the job.

Ask the model for a final pass on language only:

List every word or phrase in the posting that signals an age
preference, a gender coding, or a requirement the notes do not
support. Quote each one. Do not rewrite.

Then you decide.

A recruiter's hand holding a blue highlighter over one line of a printed page

Send the week back to the hiring manager and finish the posting

Your posting now makes claims on behalf of someone who has not read it. Fix that before a candidate does.

Write Marcus a short note with the Gaps list and the impressions. Let the model draft it:

Write a short, friendly note to the hiring manager. List the open
questions from Gaps as numbered questions, then the two
claims marked Impression, asking him to confirm or correct each.
Attach the posting below. Keep it under 150 words.

The note that comes back asks four things: whether the Saturday rotation is required, the weekday overtime rate, the start date, and whether "phones triple" is a fair description or a rough feeling. Marcus can answer in two minutes, because every question is a yes or a number.

If your email is connected to a Magai agent through one of the 150+ integrations, the note can land in your drafts folder instead of your clipboard. That saves a paste. The part that matters is the questions, because they are what turns the posting from your version of the job into his.

Here is the kind of answer you get back, again made up for the example. Saturday is required, and it comes around once a month because four people share the rotation. Weekday overtime during a cold snap is time and a half after 40 hours. The start date is as soon as the offer is accepted. "Phones triple" was a feeling, so call Monday the busiest day. And Tuesday through Thursday is scheduling follow-up visits and moving trucks around as calls come in.

Paste his answers into the thread and ask for one more revision. The thread still holds the rules, the notes, the three drafts and the audit table, so the model changes only what his answers touch. Then run the audit table once more. It takes a minute and catches the sentence that drifted.

The finished posting, in the parts that changed most, now reads like this:

Dispatch Coordinator, Harbor Mechanical Tacoma, WA, in the office. $24 to $28 an hour, with overtime at time and a half after 40 hours during cold snaps.

You decide who gets a truck today. Harbor Mechanical runs 14 technicians and 11 trucks across commercial buildings, and every call, ticket and radio message about those trucks lands on you first.

A week in this job: You start at 6:30, before the phones do, and the first calls come in around 7. Monday is the busiest day. Tuesday through Thursday you schedule follow-up visits and move trucks around as new calls arrive. When the first cold snap of the year hits, you decide which no-heat call gets a truck today and which waits until tomorrow. One Saturday a month you take your turn on the on-call rotation, shared among four people and paid at time and a half.

You need on day one: Experience in dispatch, scheduling or a front desk with heavy phones, and the ability to stay level with a customer whose heat is out. Nice to have: Spanish, which about a quarter of calls use. You do not need HVAC experience. We teach the vocabulary.

Compare that with the label you started with. A candidate who reads it knows the pay, the start time, the hardest day, the hardest decision and the Saturday, and can decide in two minutes whether to apply.

Save the rules once and the next role takes ten minutes

You have three more open roles. Do not repeat the setup four times.

Turn the thread rules into an agent. An agent in Magai is a set of reusable instructions that runs on any model, so the same rules apply whether you ask for a draft from /Claude, /GPT or /Gemini. Give it a plain name like "Job description from notes" and put this in its instructions:

You write job postings from a hiring manager's notes.
Use only facts in the notes. List anything missing under "Gaps".
Open with the work the person owns. Build "A week in this job" only
from days and situations the notes describe.
Keep must-haves and nice-to-haves separate. Put pay range and
location in the first five lines exactly as given.
Never use: "we are seeking", "dynamic", "high-pressure", "rockstar",
"self-starter", "team player", "passionate".
Label a manager's impression as an impression. Never present it
as a count.
After the posting, produce an audit table of every factual claim
with a supporting quote from the notes.
Our company line, equal opportunity statement and application
instructions go at the bottom exactly as follows: [paste yours]

The last line is the part most recruiters forget. Your equal opportunity statement and your application instructions do not change from role to role, so the agent holds them and you never retype them. If you want the full build, how to build an AI agent without writing code walks through it.

Keep manager facts out of the agent. Marcus's 14 technicians belong in his thread, not in instructions that every role will read. The agent holds the rules and your company's words, and each thread holds one job's facts.

Then give each role its own thread and keep the threads together. Folders and Collections in Magai organize chats, so a folder named for the quarter's open roles holds the dispatch coordinator, the next three, and the notes behind each. When a hiring manager calls back in March asking what the posting said about Saturdays, the answer and the notes that justified it are one click away.

A thread per role also protects you from contamination. The notes for a front desk role never leak into the notes for a technician role, because each thread has only its own facts.

The next role then starts with one message. Open a new thread with the agent, paste the manager's notes, and type:

Notes from [manager's name] for the [role title] are below.
Write the posting, then the audit table.

Read the Gaps list first, send it back to the manager, and have a finished posting by the time the answers arrive.

Open the thread before the next manager calls

Next Monday, a manager will send you a title and a link.

Reply with the ten questions. Type the answers into a new thread in Magai. Paste your rules or call your agent, ask for a week, run it past two or three models, and audit every line. Send the Gaps back. Post when the manager has answered.

Start with the role whose manager is hardest to pin down. If you can pull a week out of that one, the other three are easy. Book the fifteen minutes today, before the posting goes live with the old words and the old link.

Prices start at $20 a month, and every plan carries a 30 day money back guarantee, so the first posting can be the test.

Get the notes. Write the week.

More in AI for Business