How Do I Decode a Job Posting With an AI Agent?
Build an AI agent in Gumloop that reads a job posting and the company's own pages, then ranks what the role asks for.
Give a Gumloop AI agent the job posting and the company's own pages. It reads both and hands back the work in order of weight, what the first 90 days likely hold, the words the posting repeats, and the questions the posting leaves open.
Every bullet in a job posting sits at the same size on the screen, whatever its weight.
An AI agent can read the posting for weight. How Do I Use an AI Agent in My Job Search Without Letting It Contact Anyone? covers what a job search AI agent never does, and this one follows all of it. Career Lab has its own Decode for a job description you paste in. The AI agent you build in Gumloop is your own, and it reads the company's pages too.
What does the posting decoder hand back?
The work in order of weight. A posting lists duties in the order someone typed them. The AI agent ranks them by how much of the posting leans on each one, and it quotes the line that earned the rank.
The first 90 days, marked as a guess. A posting never says what month one holds. The AI agent works it out from the weighted duties and from what the company's pages say it's building, and it labels every line "read" or "inferred".
The words the posting repeats. A word that shows up in the summary, the duties and the requirements is the word to use in your resume summary and in your first answer.
The questions the posting leaves open. Reporting line, team size, what success looks like. The AI agent writes each gap as a question you can ask at the interview.
What do you paste into the instructions for the posting decoder?
Fill the brackets with your own words.
You are my job posting decoder AI agent. I'm looking at the [ROLE TITLE] role at [COMPANY], and I'll give you the posting text or its link. Read the posting. Then read [COMPANY]'s own pages that I link or that you can open without logging in: the about page, the careers page, the newsroom and the leadership page. Hand me four things, each under a short heading. 1. The work in order of weight. Rank the duties by how much of the posting leans on each one, and quote the posting line that earns each rank. 2. The first 90 days. Say what the role likely holds in that time. Mark each line "read" if a page says it or "inferred" if you worked it out, and give the link for every "read" line. 3. The words the posting repeats. List each repeated word or phrase and where it appears. Tell me which ones to use in my [RESUME SUMMARY] and my first answer. 4. The questions the posting leaves open. List what it doesn't say, such as reporting line, team size and what success looks like after [NUMBER OF MONTHS] months. Write each one as a question I can ask. Never guess what a posting doesn't say. Write "not stated" and turn it into a question. You contact no one. You never apply, email, message, connect with or follow anyone, and you never open a page behind a login.
How do you build the posting decoder in Gumloop?
- 01
Open Gumloop and say what the decoder is for.
Open Gumloop, create a new AI agent and tell guided setup in one sentence that you want a job posting decoded. Then paste the block above into the instructions box on the tab named Agent, and press Save, because nothing saves until you do. Checked on 2026-11-09 against Gumloop's AI agent documentation.
- 02
Set what it reads.
Web Fetch is on by default and reads the content of a specific URL, which covers the company's pages. Web Search is on by default too. For the posting itself, paste the text into the task when the page asks you to sign in or the posting is a PDF, because the documentation doesn't promise PDFs. Turn Image Generation off. New custom AI agents can start with the browser on, so open the Browser section of the configuration and turn it off if you don't want it browsing. Checked on 2026-11-11 against Gumloop's AI agent documentation and Gumloop's browser documentation.
- 03
Connect nothing, then turn Edit Own Instructions off.
Leave the Connectors list empty and the Triggers tab alone. If you add a connector later, Gumloop's default approval is Always allow, so change it to Ask for writes/deletes first. Edit Own Instructions sits right under the instructions box. It's on by default and keeps no version history, so copy your instructions into a note and switch it off. Every posting is a different document, and a correction for one posting, like "ignore the benefits paragraph", would turn into a standing rule on the next posting. Checked on 2026-11-09 against Human in the loop and the self-improving instructions section of Gumloop's AI agent documentation.
- 04
Check it against a role you know.
Use the posting for a role you hold now, or held, read the four parts against what you know happened, and mark any line that's off.
I am an ambassador for Gumloop. Code JOBSEARCHAGENTS takes 20% off your first month.
What should the first three postings tell you?
First, ask someone who does the job, or did it, whether the top-ranked duty is the one that takes up the most time. If it isn't, the posting and the real job weigh things differently.
Next, search the posting yourself for the word the AI agent says repeats most, and count. Last, open the link behind each "read" line in the first 90 days. If a page doesn't say it, the line was inferred, and it should have been marked that way.
Check your resume against the posting with How Do I Compare My Resume to One Job Posting With an AI Agent?.
Frequently asked questions
What if the posting is a PDF or sits behind a login?
Paste the text into the task. Web Fetch reads the content of a specific URL, and the documentation doesn't promise PDFs.
What does it do when the company's pages say nothing about the role?
It writes "not stated" for the gap and turns it into a question you can ask. Lines in the first 90 days that no page supports stay marked "inferred".
Why is Edit Own Instructions off for this one?
Every posting is a different document, and a fix for one would become a rule for the next. Switch it on only after three postings in a row come back right, and copy your instructions out first because there's no version history.