The two prompts behind an autonomous job-search agent
Article·12 min read
Published · By Ofer Regev
A skill makes Claude excellent at one task, but only when you ask. It does nothing on the day you forget, and it never hands its result to the next step. So you end up being the glue: remembering to check, opening a chat, pasting context, copying the result into your tracker.
A job search is a good example of that problem, so I used it to build two scheduled agents on top of the job-search skills. One runs every morning, one every Sunday. They read the inbox and a Notion tracker, judge what they find, add rows, and send a short summary you read on your phone.
Below are the two real prompts. The personal details (the profile, the database links, the companies to skip) are replaced by brackets you fill in; everything else is exactly as it runs. After each one I explain what every part does and why it's there. At the end there is a template you can copy for any repeating process, not only a job search.
The setup in one minute
A tracker the agents can read and write. Three Notion databases: Applications, Target Companies and Networking Contacts. The free job-hunter plugin sets them up for you ("Set up my job tracker"), in Notion or as local files.
Two scheduled tasks in the Claude app, each with a prompt, a schedule and the connectors it needs (Gmail and Notion).
That's it. No code, no server. The tracker is the agents' memory: each run starts with no memory of yesterday, so anything it needs to know has to be in the prompt or in the tracker.
Agent 1: the daily job-alert check
Runs every morning at 8:13. Its only job: go through yesterday's LinkedIn job alerts and say which roles are worth your time.
You are helping a candidate run their job search. Every morning, triage the LinkedIn job-alert emails they received in the last 24 hours and tell them which roles are worth their time. The goal is that they never waste time on roles they do not fit.
Steps:
1. With the Gmail connector, search: (from:jobalerts-noreply@linkedin.com OR from:jobs-noreply@linkedin.com) newer_than:1d. Read each message with get_message (PLAIN_TEXT). Each email lists jobs (title, company, location, LinkedIn job link). Collect every distinct job; dedupe by LinkedIn job id.
2. With the Notion connector, query the Applications Tracker ([link to your Applications database]; columns Company, Role / Company, Status). Drop any job at a company already in the tracker with any status (mention separately only if it is a NEW role at a company where they are still Applied / in process). Never surface [companies you already ruled out].
3. For each remaining job try WebFetch on the LinkedIn view link or the company's careers page to read the description; if it cannot be read, judge from title + company and say so.
4. Run a fit analysis against the candidate's profile: [years of experience; current role and company; 2-3 things you shipped, with numbers; your domain background; the tools you build with; the size of the teams you led]. Gaps: [the domains and skills you have no evidence in]. Target: [level, focus, location]. Verdict per job: Strong fit, Possible (stretch), or Skip (long shot). Skip: junior roles, non-product titles, [sectors or languages you rule out], product-analyst/data roles, Principal/Head/Team Lead roles whose core depends on a domain they have no evidence in, and non-AI roles at non-tech companies unless clearly senior. Be honest, do not inflate.
5. For each Strong fit or Possible job, add a row to the Applications Tracker: Role / Company, Company, Source LinkedIn, Status Not Started, Next Step "find a contact at <company>, then apply", Next Step Date within 3 days (Strong earlier), Notes with the verdict, one or two evidence lines, and the link. Do not add Skip jobs to the tracker.
6. Final message (this is the phone notification, keep it short): list only Strong fit and Possible roles, best first, each as one line: Company, role title, rating, one-sentence reason (the gap if any), and the link. End with one line: how many roles were reviewed and how many were skipped as not a fit. If nothing is a Strong fit or Possible, say 'No good fits today' and the count reviewed.
Do not apply to anything, do not message anyone, do not modify Gmail, and only add new rows to Notion (never edit existing ones). Write in English.
The opening sentence gives a goal, not just a task. "Triage the job alerts" is the task. "So that they never waste time on roles they do not fit" is the goal, and it's what the agent falls back on when a case isn't covered by the rules. When in doubt, it leans towards protecting your time.
Step 1 names the exact source. Not "check my email" but the exact Gmail search, how to read each message, and what to pull out. "Dedupe by LinkedIn job id" matters: the same role often shows up in two or three alerts.
Step 2 checks the tracker before doing any work. This is what makes it an agent rather than a daily report. Anything already seen, applied to or rejected is dropped, so you never read the same role twice. The exception is useful: a new role at a company where you're already in process is worth knowing about, so it's mentioned separately. The "never surface" list is for companies you've already decided against. It grows over time.
Step 3 reads the real job description, and says when it can't. LinkedIn often blocks automated reading. Without "say so", the agent would quietly guess from the title and present it as a finding. With it, you know which verdicts are solid and which are guesses.
Step 4 is the rubric, and it's the heart of the prompt. It holds the profile, the real gaps, the target, and an explicit list of what to skip. Writing down the gaps is what makes the verdicts honest: without them, every AI role looks like a "strong fit". "Be honest, do not inflate" is there because models like to please. Three fixed verdicts (Strong, Possible, Skip) keep the output comparable from day to day.
Step 5 writes the result where the next step can use it. Only Strong and Possible roles become rows, each with a next step ("find a contact, then apply") and a date. Strong roles get an earlier date. The notes carry the verdict and the evidence, so when you open the row a week later you know why it's there. Skipped roles aren't added: the tracker stays a to-do list, not a log.
Step 6 is written for a phone. The final message is the notification you get, so it's short, best first, one line per role, with the gap spelled out. The closing count ("reviewed 14, skipped 11") tells you it actually ran, even on a day with nothing good.
The last paragraph is the guardrails. No applying, no messages, no changes to Gmail, and in Notion only new rows, never edits. That is what makes it safe to leave alone: the worst a bad run can do is add a row you'll delete.
Agent 2: the weekly network check
Runs every Sunday at 8:47. The daily agent handles what comes in. This one handles what needs doing on your side: who to talk to, who hasn't been followed up, and which target companies have new roles.
You are helping a candidate ([your level and role], [your city]) run their job search. Every week, audit their Notion job-search workspace, find new roles at target companies, screen them for fit, and tell them what to do. Use the Notion connector. Databases (data source ids): Target Companies = [link to your Target Companies database] (columns: Company, Tier [Dream/Strong Fit/Backup], Category, Stage, Target Level, Status [Not started/In progress/Done], Warm Contact, Why It Fits, Careers URL, Location, Notes); Networking Contacts = [link to your Networking Contacts database] (Name, Company, Role, Relationship, Status [To Reach Out/Reached Out/Call Scheduled/Call Done/Ongoing], Last Contact, Notes); Applications Tracker = [link to your Applications database] (Role / Company, Company, Status [Not Started, Applied, Recruiter Screen, Professional Interview, Pending Response, Home Assignment, Final Round, Offer, Rejected, Withdrawn], Source, Next Step, Next Step Date, Date Applied, Notes). Fetch each data source first to confirm the schema, then query all rows.
FIT RUBRIC (use for every role, every run). The candidate's profile: [the same profile as in the daily agent]. Gaps: [the same gaps]. Target: [level, focus, location]. Verdicts: strong, stretch, long_shot. long_shot means: level clearly above scope (Principal/Head/Team Lead with deep domain requirement), or the core of the role depends on a domain they have no evidence in ([the domains from your gaps]) with no offsetting link to their strengths, or junior/non-product/[languages you rule out]. Be honest, do not inflate.
Do this:
1. Coverage gaps: list target companies (prioritise Tier Dream, then Strong Fit, Status Not started or In progress) that have NO row in Networking Contacts (match on Company, case-insensitive) and an empty Warm Contact. For the top 5, say what kind of person to find there (a PM or Head of Product, an ex-colleague from [your previous companies], a member of [a community you're in], a recruiter) and give a one-line suggested way to reach them.
2. Stalled outreach: contacts with Status To Reach Out for more than 7 days (use createdTime), or Reached Out with Last Contact more than 7 days ago. List them with a suggested next message.
3. Pipeline check: companies already in the Applications Tracker with Status Applied, Pending Response or Recruiter Screen that still have no contact in Networking Contacts. These are the best referral opportunities.
4. Refresh the target list: count target companies that are Status Not started AND not present in the Applications Tracker. If fewer than 10, research (web search) and add 5-8 new companies to Target Companies as complete rows (Company, Category, Stage, Tier, Target Level, Location, Why It Fits, Careers URL, Status Not started). Fit criteria: [your location], AI-native or AI-heavy product, growth-stage, with Senior PM openings plausible, ideally near [your domain background]. Do not add companies already in Target Companies or the Applications Tracker (including rejected or withdrawn ones). Skip [companies you already ruled out].
5. New role search: for each target company with Status Not started or In progress, check its careers page (web search / fetch) for open Product Manager, Senior PM or AI PM roles in [your country]. Skip companies already in the Applications Tracker with that same role. Never add [companies you already ruled out]. If a careers page cannot be read, list it as unreadable, do not say it has no roles. For each new role, run the FIT RUBRIC. Add it to the Applications Tracker as a row (Role / Company, Company, Source Direct Apply, Notes with the verdict, 1-2 evidence lines, role link and the date seen). If verdict is strong or stretch: Status Not Started, Next Step "find a contact at <company>, then apply", Next Step Date within 3 days (strong earlier than stretch). If verdict is long_shot: Status Rejected, Next Step "Closed: screened out as not a fit", no date, and the reason in Notes starting "FIT: long shot, screened out by me, not by the company". This is intentional: the candidate does not want to spend time on roles they do not fit.
6. For any Applications Tracker row you touch or notice with no Next Step Date and an open status, set a sensible Next Step Date (an interview date if known, otherwise a follow-up date about 7 days after applying or the last contact).
Do not message anyone. Do not change the Status of existing rows, except the new rows you create in step 5. [A company to leave alone] is deliberately low priority, leave it alone. Finish with a short summary the candidate can read on their phone: new roles worth their time (strong/stretch, best first, with a one-line reason), how many roles were screened out as long shots (names only), what needs action this week (max 5 bullets), and what you added or changed. Write it in English.
How it works, part by part
The first paragraph is a map of the workspace. Every database, every column, and every allowed value of every status. The agent starts each run knowing nothing, so it needs the map every time. "Fetch each data source first to confirm the schema" protects against the day someone renames a column: it checks before it writes, instead of failing silently.
The rubric is repeated in full, not referenced. It's the same profile and gaps as the daily agent, with labels named for the job (strong, stretch, long_shot). It would be shorter to write "use my usual criteria", but there is no "usual" for an agent with no memory. Every prompt has to carry everything it needs.
Steps 1 to 3 are about people, not jobs. Most roles are filled through someone who knows someone, so the agent looks for the gaps in the network:
Coverage gaps: dream companies where you don't know anyone yet, with a suggestion of who to look for (an ex-colleague, a community member, a recruiter) and how to reach them.
Stalled outreach: people you meant to contact and didn't, or contacted and never followed up. "More than 7 days" turns a vague feeling into a clear rule.
Pipeline check: companies where you've already applied but have no contact. That's where a referral helps most, so it gets its own step.
Step 4 keeps the target list full. It only researches new companies when fewer than 10 are left untouched, so it doesn't flood the list every week. Each new company is added as a complete row with "Why It Fits", so you can judge it in a second. It's told not to bring back companies that were already rejected or withdrawn.
Step 5 is the role search itself, with two important rules.
"If a careers page cannot be read, list it as unreadable, do not say it has no roles." Without this rule, an agent that can't read a page will happily report "no open roles", and you'll never check again.
Long shots are recorded, not hidden. A role that isn't a fit still gets a row, marked Rejected, with a note starting "FIT: long shot, screened out by me, not by the company". That way the agent won't suggest it again next week, and anyone going through the rejected rows knows it was your call, not the company's.
Step 6 makes sure nothing open is left without a date. An open application with no next date is one you'll forget. The agent adds a sensible follow-up date, which is the only change it may make to existing rows.
The guardrails are narrower than they look. No messages, no status changes to existing rows (except the new ones it just created), and one company it's told to leave alone. It can add a lot, but it can't undo your decisions.
The summary is a to-do list for the week. New roles worth the time, long shots by name only, at most five actions, and what it changed. Reading only that message on Sunday morning is enough to know what to do this week.
Why two agents and not one
Different rhythms. Job alerts arrive daily and go stale fast. Networking and target research work weekly. One big daily agent would be slow, noisy and expensive.
The tracker connects them. The daily agent adds roles, the weekly one finds contacts for those same companies, and both skip anything already there. Neither knows the other exists. They share state, not memory.
Each one is small enough to trust. When something goes wrong you know which prompt to fix.
Lessons from running agents like these
Run it by hand first. Press Run now, approve the tools it asks for, read the result, and fix the prompt. Only then let it run on schedule.
Write down the gaps, not just the strengths. Without them, every verdict is "strong fit".
"Say so if you can't" beats "do your best". Silent guesses are the most expensive mistakes an agent makes, because you don't know to check them.
Add-only is the best guardrail you'll write. If the agent can only add rows, a bad run costs you a minute of deleting.
The exclusion lists grow. Every time a run surfaces something you've already decided against, it goes into the prompt. The prompt gets better because you correct it, not because the agent learns.
The template, for any repeating process
Both prompts follow the same six parts. Copy them into a new scheduled task and fill in the brackets:
GOAL: You are helping me run [my process]. Every [day/week], [what to do] so that [the outcome I want].
WHERE THINGS LIVE: New items come from [source: a Gmail search, a list of sites, a form]. My tracker is [link] with the columns [list them, with the allowed values]. Read the tracker first to confirm the columns.
HOW TO JUDGE: Here is who I am and what I want: [profile, goals]. A good fit means [...]. My real gaps / limits are [...]. Skip anything that [...]. Give each item one verdict: Strong, Possible or Skip, with one line of evidence. Be honest, do not inflate.
STEPS:
1. Collect the new items. Dedupe them.
2. Drop anything already in the tracker.
3. Judge each one with the rules above. If you cannot read something, say so instead of guessing.
4. Add a new row for each Strong or Possible item, with the verdict, the evidence, the link, a next step and a date.
GUARDRAILS: Do not send, post, apply, pay or delete anything. Only add new rows, never edit existing ones. Never touch [anything off-limits].
REPORT: Finish with a short message I can read on my phone: the best items first, one line each, then how many you reviewed and skipped. If nothing is worth my time, say so.
The same structure works for following up on leads every morning, screening applicants against a role, or checking what competitors released this week. Swap the source, the tracker and the rules for judging.
To create it: in the Claude desktop app, open Scheduled in the sidebar and choose New task. Paste the prompt, pick the schedule, connect the tools it needs, and press Run now once before you let it run on its own. Or just tell Claude in a chat: "make this a scheduled task that runs every weekday at 8:00".