Joel EvertJE ADVISORY

Workshop take-home · no signup

Rebuild the
routine at home.

Everything from the AI job search workshop in one place: the weekly routine, the prompts, two worked examples, and a starter for anyone who wants to go further. The prompts are pulled live from the same deck I present from, so this page stays current as the deck changes.

Synced to the live deck Copy-ready prompts No signup

Setup, then four exercises

The weekly routine.

This is the order the workshop teaches it in. Exercise 1 is a one-off setup. Exercises 2 to 4 repeat every week. Follow-up runs alongside all of it, every time you apply.

  1. 01

    Start your source of truth

    A private factual reference document: work history, skills, education, target roles, and constraints, in your own words. Not a CV, and not a place for AI to round anything up. One-off setup, done once.

  2. 02

    Build your search profile

    Primary roles, stretch roles, exclusions, and a weekly volume cap, saved somewhere your AI reads automatically: a ChatGPT project or custom instructions, or a Claude project.

  3. 03

    Evaluate a real ad

    Give it your source of truth and let it find one real, currently open ad. Get an honest fit assessment, with anything unclear marked as a gap instead of guessed at.

  4. 04

    Set up a routine

    Turn the search and the evaluation into one repeatable process, run it by hand first, and only schedule it once it holds up more than once.

  5. 05

    Follow up

    Log every application the day you apply, with proof, and let it draft, never send, a nudge once the follow-up date comes due.

Official guides

  • Claude scheduled tasks

    Part of Claude Cowork. Needs a paid plan. Runs in the cloud, even with the app closed.

    Guide
  • ChatGPT scheduled tasks

    Available on Plus, Pro and Team. Set up at chatgpt.com/schedules.

    Guide
  • Gemini scheduled actions

    Rolling out gradually, so you may not have it yet. Best for simple daily summaries.

    Guide

Paste-ready

The prompts.

These load straight from the workshop deck, so if a prompt changes there, it changes here too. Paste one into ChatGPT or Claude and follow the instruction above it.

Loading the current prompts from the deck…

Exercise 2, continued

Where your profile lives.

Save the search profile once, somewhere your AI reads it automatically, so you're not re-pasting it into every new chat.

ChatGPT

  1. Open Settings from your profile menu, then Personalization, then Custom instructions. Paste your search profile there so every new chat can see it.
  2. Or create a Project instead, and add your source of truth and search profile as project files. Every chat inside that project reads them.

Menu wording changes over time. Look for Custom instructions or Projects in Settings if this doesn't match exactly.

Screenshot coming

Claude

  1. Open Projects in the sidebar, then start a new project for your job search.
  2. Open its Project knowledge and add your source of truth and search profile as files. Every chat inside that project reads them automatically.

Menu wording changes over time. Look for Projects in the sidebar if this doesn't match exactly.

Screenshot coming

Illustrative only

Two worked samples.

Short examples of Exercise 1 and Exercise 2 output, built around the same illustrative Financial Controller candidate used in the workshop's own live demo.

Not a real person. This is a made-up candidate, used only to show the shape of the two documents. Copy the structure, not the facts.

Example source of truth

Work history
Financial Controller, mid-size Danish manufacturer, 4 years. Senior Accountant before that, 3 years.
Skills
Month-end close, IFRS reporting, advanced Excel, Power BI, Dynamics 365 Business Central.
Education
BSc, Business Economics.
Target roles
Financial Controller, Finance Manager. Not interested in pure bookkeeping roles.
Constraints
Copenhagen or remote within Denmark. Three-month notice period.
Proud of
Cut month-end close from nine days to five. Built one reconciliation automation that removed a recurring manual task.
Things I do not claim
No formal accounting certification. No external audit experience at a listed company. No hands-on SAP experience.

Example search profile

Primary roles
Financial Controller, Finance Manager.
Stretch roles
Head of Finance, small team.
Locations
Copenhagen, or remote within Denmark.
Exclusions
Roles requiring a CPA or ACCA license. Pure bookkeeping. Employers under ten people.
Recency window
Last 14 days.
Worth-showing threshold
Strong or possible fit only. Skip weak fits.
Weekly volume cap
15 ads reviewed per week.
Things I do not claim
Not fluent in Danish business correspondence, conversational only. No SAP experience, so roles listing SAP as a hard requirement are excluded rather than stretched for.

Make it standing

Turn the rules into a skill.

Three small skills, built from everything above: the job-search discipline itself, an independent review before anything gets sent, and the follow-up tracker. Claude can import a skill folder directly, drop it in your skills folder and refresh. ChatGPT and Gemini do not support that yet, so paste the plain-text version into your project instructions or custom instructions instead. Same rules, either way.

Skill 1 of 3

Job Search Discipline

The workshop's rules for evaluating an ad, building a search profile, and logging an application, as a standing skill instead of something you re-explain every time.

Job Search Discipline. Apply these rules whenever you help me with a job search: evaluating an ad, building a search profile, or logging an application.

The one fact source: use only my source of truth document and what I tell you directly in this conversation. Never invent work history, skills, dates, or achievements. Never fill a gap with something that merely sounds plausible.

Keep a things-I-do-not-claim list: every profile or assessment you produce should carry an explicit "Things I do not claim" section, skills, certifications, or experience I do not have or cannot support with evidence. Before any draft goes out, scan it for every item on that list. A match is a hard failure to fix, not a style note.

Mark gaps, do not guess: when a requirement or fact is not covered by what I gave you, write NEEDS_USER_INPUT next to it instead of filling it in.

Verify before you score: before assessing fit against any ad, confirm the employer is real, the role is still open, and you have the full advertisement, not a summary. If you cannot positively confirm the role is open, meaning the title is visible and there is a working apply action, treat it as unverified. Unverified is not open. Do not score it and do not apply.

Overriding a poor fit: if I tell you to apply despite a poor fit, keep the honest verdict and the named gaps in the record, note that I overrode it, and proceed without strengthening any claim. Rewording a true fact into a stronger claim counts as inventing. Requirements I factually fail, work authorization, a required language level, or a required certification, cannot be overridden.

One job per employer and title: same job title plus same employer within the last 30 days counts as one job, no matter which site or URL it came from. Treat the employer's own page as the copy of record. The same ad appearing much later is a repost. Treat it as new, but never reapply to anything already applied to or unconfirmed.

Proof or unconfirmed: an application only counts as applied once there is proof, a confirmation screen or a receipt email. A confirmation only counts if it identifies this application, the role or employer on the confirmation screen, or a receipt email for it. A generic thank-you page is not proof. Without proof, set its status to unconfirmed and do not let a second application get logged for the same role. Resolved means proof arrived. If proof never arrives, the role stays unconfirmed permanently. Never apply to it again. A possible duplicate application costs more than a missing confirmation.

Follow-up on a schedule: set the next follow-up date the same day an application gets logged, fourteen days out. A sent nudge or a snooze resets that date another fourteen days out. When I report any response, rejection, interview, or offer, clear the row's follow-up date. Never draft a nudge for a role that has answered or closed. Draft nudges. Never send them.

Pacing is mine to set: this does not cap how many applications go out in a week. My own routine caps around fifteen a week. Help me pick a cap and hold to it, rather than sending as many as the search turns up.

Nothing goes out without a yes: drafting, logging, and tracking can happen freely. Sending, applying, or posting requires my explicit yes for that specific action, every time.
Download skill folder 5.2 KB · MIT License · Claude skill import
Skill 2 of 3

Review Before Send

An independent check before anything gets sent, an application, an email, a post. Checks claims against your facts, hunts invented detail, and judges voice against real writing samples you supply.

Review Before Send. Before any draft you write for me (an application, an email, a post, or anything else about to reach another person) goes out, run this review as a separate pass.

Why a fresh context: if you can start a new conversation or a separate agent with no memory of the drafting conversation, run the review there. The check should not just be the same reasoning agreeing with itself.

What to check:
1. Claims against facts. Compare every factual claim in the draft, skills, dates, numbers, achievements, quotes, against the source facts I actually gave you. Flag anything you cannot trace to a specific fact. Numbers and percentages must match the source exactly.
2. Invented detail. Read for anything that sounds plausible but was not given: a rounded-up number, an upgraded title, a smoothed-over gap, a detail that reads like it filled itself in.
3. Voice. Ask me for two or three things I actually wrote, a real email, post, or message. Judge the draft against those samples, not against itself or a generic idea of good writing. Keep the samples on file and reuse them for later reviews, so I only have to supply them once.
4. Structure and completeness. Confirm the draft actually does the job it is meant to do, and that nothing I asked for is missing.
5. No-claim list. Compare the draft against my things-I-do-not-claim list. Any match is an automatic reject, not a suggestion.

Permission to reject: you are explicitly allowed to say none of the drafts are good enough. Report failures plainly. Do not soften a rejection into a list of minor suggestions to avoid saying no.

One hostile pass: after accepting a draft, make one hostile pass, try to disqualify it. Look for the one reason it should not go out. Only once it survives that pass is it ready.

Output: give me a verdict, ready to send, needs changes, or reject entirely. If changes are needed, list each one, traced to a specific claim or a specific voice mismatch. If no voice samples are available, say so in the verdict. Never report ready to send with an unchecked voice.
Download skill folder 4.1 KB · MIT License · Claude skill import
Skill 3 of 3

Follow-up Tracker

The take-home follow-up prompt as a standing skill: log an application with proof, get a due list, and draft, never send, a nudge when it's time.

Follow-up Tracker. Keep a running, factual log of my job applications. This is not a one-time task, update it every time I report something new.

Keep a table with these columns: Employer, Role, Date applied, Status, Next follow-up date, Proof captured (yes or no).

When I tell you I applied to a role:
1. Record the employer, role, and date applied exactly as I state them.
2. Ask whether I have proof, a confirmation screen or a receipt email. A confirmation only counts if it identifies this application, the role or employer on the confirmation screen, or a receipt email for it. A generic thank-you page is not proof. Without proof, set status to unconfirmed, tell me to check the employer's portal, and do not let me log a second application for the same role. Resolved means proof arrived. If proof never arrives, the role stays unconfirmed permanently. Never apply to it again. A possible duplicate application costs more than a missing confirmation.
3. Set the next follow-up date to fourteen days from today, the day I log it, not from the date applied. If I am logging an old application, do not backdate the clock.

When I ask what is due, show me only the rows whose follow-up date has arrived or passed.

When a row is due, draft a short nudge, two or three sentences, in the same language as the original job ad, referencing my application date. Do not send it. I will send it myself.

When I report any response, rejection, interview, or offer, clear that row's follow-up date. Never draft a nudge for a role that has answered or closed.

If I tell you I sent a nudge, or I tell you to snooze a row, reset that row's follow-up date to fourteen days from today.

Rules you must follow: never record an application as applied without proof, log it as unconfirmed instead. Never invent a status, a date, or an application I did not give you. Never send, submit, or post anything on my behalf. A drafted nudge is as far as you go.
Download skill folder 3.4 KB · MIT License · Claude skill import

Prefer the complete set?

Get all three skills.

One small ZIP with all three skill folders, the paste versions, and the license.

Get all three 9.7 KB · MIT License

For the ambitious

Build a stronger search with JobSpy.

JobSpy is a free tool that checks several job sites in one search. You don't need to write any code yourself, your AI writes it for you. Paste this prompt to get started.

Paste-ready

JobSpy starter prompt

I want you to build me a small script using the JobSpy Python library (github.com/speedyapply/JobSpy) that searches for [job title] roles on [job sites, for example LinkedIn, Indeed, Glassdoor] and saves the results to a CSV file.

Before you write any code, tell me plainly:
- JobSpy's coverage is not guaranteed. It can miss jobs, or return results it can't confirm are still current.
- Jobindex is not on JobSpy's official list of supported sites. Do not rely on this script for Jobindex results, I will search that separately myself.

Build the script so that:
- It removes duplicates. Same job title plus same employer counts as one job, no matter which site it came from or what the URL is.
- It does not log into LinkedIn, send a connection request, send a message, or contact anyone automatically. This script only searches and saves a list. Nothing goes out to anyone.

Once it runs, walk me through the results. For every row, remind me to open the full ad on the employer's own site before I trust anything in it. The CSV is a shortlist to check myself, not a verified list.