Magda Kufrej - Work Ally
May 20, 2026
AI has wedged itself between job seekers and employers, and learning to work with it now sits closer to a survival skill than a nice extra. This session moves through eight things. They are why the job market has changed, how to set up a paid account, the ACE prompting model, using AI responsibly, researching roles before you apply, building a job-hunt project, pushing into agents and skills, and the other tools worth knowing. It closes on the one move AI cannot make for you.
Recruitment has quietly turned into a machine-to-machine process, and that shift is what makes AI literacy matter for anyone looking for work. The old picture of a candidate, a recruiter, and a job ad in a newspaper is gone. Now there are two machines in the middle. You use AI to translate what you want into an application, and the recruiter uses AI as a gatekeeper so that only some applications ever reach a human desk.
Companies keep pouring money into recruitment AI even though the results are shaky. A report presented at Web Summit in Lisbon in November 2025 put numbers on the strain. 63 percent of employers say finding talent is harder than ever, 70 percent of job seekers struggle to land good jobs even when they're well qualified, and 87 percent of recruiters say resumes are failing them as a way to evaluate candidates. On the employer side, AI now handles candidate vetting through LinkedIn keywords, resume screening, pre-interviews, and in some cases interviewing you before any human sees you. Applicant tracking systems run on semantic matching, skill assessments, and automated screening. Oddly, many of the people building these tools are strong on AI and weak on hiring. Asked why a system filters applicants at an 80 percent profile fit, one builder had no reasoning beyond "just like that," because plenty of them have never been recruiters. If the company's side is automated, matching machine to machine is the rational response from the candidate's side too.
The argument for getting AI literate is generational. Not learning it now is compared to a grandparent who never picked up email or a smartphone, and office work that runs on a computer is the most exposed. The numbers come from a few sources. The World Economic Forum projects a net gain of 78 million jobs by 2030, made up of 170 million new jobs created against 92 million displaced. The IMF estimates 60 percent of jobs in advanced economies will be exposed to AI, where exposed means they will need to work with it rather than be erased. The WEF also expects 39 percent of core skills to be transformed or outdated by 2030 and 59 percent of workers to need retraining. An Anthropic report mapping where AI could theoretically replace human work against where it is actually used today shows the widest gap in management, business and finance, computing and math, and especially office and administrative work.
Takeaway The job you want is being filtered by software before a person reads it, so building basic AI fluency is less about chasing a trend and more about not getting screened out. The job hunt itself is the project to practice on.
Before any clever prompting, two setup moves do most of the early work, a paid subscription and a few changes to the default settings. A paid plan, around 20 dollars a month, buys better data protection, the ability to create projects and agents, access to the newest models, and more usage.
Two settings matter on the first day. First, turn off the training-model option so your inputs are not fed back into the provider's general models and instead stay with you. Second, write custom instructions that constrain the model. Tell it not to invent answers, to say when it does not know, to flag the parts you should verify, and to link to a source or document when something is confirmed. One reassurance comes with this. The newest models, like Opus 4.7 on Claude, the thinking models, Gemini, and ChatGPT's latest, hallucinate far less than they did a year ago, so the old fear of a model making things up wholesale has eased.
Takeaway Pay for the tool and change two settings before anything else. Turning off model training protects your data, and a custom instruction that forbids invented answers heads off most of the trust problems people complain about.
ACE is a three-part structure for writing a prompt, Aim, Context, and Example, meant to leave the model as little room to guess as possible. The aim is the action you want, stated plainly, meaning what task and often what role. Adding "act as" a specific expertise shifts the model's perspective, so you can ask it to act as a recruiter who is brutal but constructive, an adversarial editor, a project manager, a hiring manager with limited bandwidth, or a supportive friend. Pairing the action with your objective, do this because I want to achieve that, sharpens it further.
Context is the verified background you hand over. That covers you, the recruiter, the role, the sector, the situation, and what you have already tried. The more accurate detail you give, the less the model interprets on its own. The example then tells it what output you expect, including the format, whether that is a document, bullet points, a short summary, or a long report. Showing beats telling. You can paste a cover letter that worked before, share samples of your writing style, or point to a model to imitate, even quoting a book like Nonviolent Communication when you want a negotiation plan to carry that tone. One worked example follows a senior-level persona, Maria, preparing for a salary negotiation. The prompt asks the model to "run a step-by-step coaching session" as "a highly experienced negotiator," with the job description attached, and requests a strategy modeled on Chris Voss, the former FBI hostage negotiator.
Beyond the basics, a handful of habits raise the quality. Ask the model to refine your prompt before running it, break a big task into smaller steps, brainstorm several options, and challenge the answers rather than accept them. When something feels off, ask it to explain its assumptions and list the sources it used. The phrase for this is the "art of ping pong," the refusal to settle for the first answer. A Web Summit slide put rough numbers on why that matters. By 2026, about 70 percent of what these tools tell you is useful, 15 percent is unhelpful, and 15 percent is wrong. Useful, but not a final authority.
Takeaway A good prompt names the aim, loads in real context, and shows an example of the output you want. After that, treat the first answer as a draft, push back, and ask the model to defend its reasoning.
Two cautions sit underneath all of this, protecting your own data and reckoning with the technology's environmental cost. On data, the rule is to keep sensitive personal information out, no matter what boundaries you have set. Strip identifying details like the phone number on your resume before you paste it in. The test is simple. If you would be uncomfortable with it leaking, do not share it.
On the environment, data centers draw heavily on both electricity and water, and the International Energy Agency expects AI electricity demand to more than double by 2030. One visible consequence is tighter usage limits, which is a reason to be efficient and not ask the same thing five times over. The flip side has two parts. Renewable energy is drawing more investment, and data centers themselves are a hot climate subsector with real hiring needs. The honest version is that this is a double-edged sword, so the advice is to use AI consciously and skip it when you do not actually need it.
Takeaway Do not feed the model anything you would hate to see leak, and do not burn tokens out of habit. The energy cost is real, even as data centers turn into one of climate's busier job markets.
Long before writing a resume, AI is most useful as a research partner, and using it well starts with knowing the difference between a chatbot, a project, and an agent. The three modes climb in capability. A chatbot is a single conversation with the memory of a goldfish, good for quick rewrites, brainstorming, and exploring roles. A project, called gems or workspaces depending on the provider, is a focused environment with its own files, instructions, and memory, which makes it a natural job-search hub. An agent, or Co-Work in Claude, has the autonomy to create files and take actions across your tools, suited to advanced workflows, though it should always run with your approval rather than free rein.
The research itself splits into five use cases. The first is not knowing which roles fit you. Feed in your resume, skills, and success stories, describe your passions, goals, and red and green flags, then ask for a list of possible roles and explore the four or five most interesting ones in their own chats, asking about day-to-day reality, typical career path, essential skills, location, pay, and future outlook. The second is exploring a subsector, say data centers, by asking for a report on its current state, major players, and hiring opportunities, and keeping a single living summary that gets updated each time a new source appears. The third is finding organizations to target, where you describe a niche, hand over your value proposition, and ask for a list of, for example, 20 companies. One real case paired an epidemiology background with full-stack development to find US food-waste startups. The fourth is researching a single organization in depth, from its recent hiring and the expertise it wants to its growth, events, media coverage, and stance on ethics and sustainability. The fifth is turning a vague sense of a role into the actual keywords and job titles to search, which then feed your LinkedIn profile and your filters.
A useful trick sits inside the third use case. Newsletters that track a subsector often list startups right after they raise funding, which is weeks before hiring begins. Perplexity, an AI search engine that returns sources, can surface those newsletters, and from there you can reach out early, even with a simple note of congratulations, before any job is posted.
Takeaway Most of AI's value in a job hunt happens before you apply. Use it to widen the list of roles you would consider, map a subsector, and catch funding news early enough to network before the posting goes up.
A project turns scattered one-off chats into a single workspace that remembers your history and gets better at producing your materials over time. The setup is to create a project, name it something like "job hunt assistant," and load it with background files and a set of instructions. The files are the project's brain, and a strong set includes your master resume in its fullest form rather than an AI-written one, a repository of skills and success stories, your values and constraints, tone-of-voice guidelines built by having AI analyze your past writing, old cover letters, your downloaded LinkedIn profile, personal-branding keywords, salary expectations, and best-practice notes for resumes and cover letters.
Inside the project, the discipline is one topic per chat, so sector research, role research, a top-20-companies list, and each individual application each get their own thread. The project files act as a shared bible that every chat can draw on, while per-chat attachments add specific context like a single job description or a company report. For substantially different targets, separate projects act as buckets, the way distinct streams of work such as career coaching, facilitation, and public speaking each warrant their own.
Each application follows the same workflow. Paste the job description into a fresh chat and ask the model to analyze it for the top required skills and compare them against the profile already in your files. Ask for company research on mission, projects, and strategy. Then have it tailor your resume and draft a cover letter using everything it already holds. The claim is that this roughly halves the time an application takes. If you are shortlisted, the same thread carries into interview prep, generating likely questions from the job description, deeper company research, a view on which STAR stories to feature, and a personal pitch. Throughout, the output still needs checking for invented facts and for AI slop.
Takeaway One project with your real documents loaded in beats a hundred scattered chats. Give it a master resume, your stories, and your voice, then run one application per chat so each draft starts from everything the model already knows about you.
Agents push past chat into action, letting AI read and write files directly on your computer and even save a repeatable process as a skill. In Claude, this lives in the Co-Work tab on the desktop app. You point a new project at an existing local folder, for instance a resume and cover letter ghostwriter folder holding job descriptions, your full resume, old cover letters, and a portfolio. Once connected, a single prompt can tell it to read your resume and past letters, learn your writing style, and generate a tailored resume and cover letter for each job description in the folder, with constraints like a two-page resume, a short cover letter, at most three matching success stories, and each pair saved in its own subfolder named by date and company. The result is a set of new subfolders the model builds while you step away, after which the real work of reviewing them begins.
The next step is to ask it to convert what you just did into a skill, named something like PrepareCV Cover Letter. The model writes a skill file and stores it both in the local folder and in the Co-Work project. From then on, dropping a new job description into the folder and calling that skill produces another resume and cover letter, usually better than the last because earlier feedback is baked in. Skills can accumulate over time, for research, pitching, portfolio building, even making slides, until the simple project grows into a full agent-driven job search.
The hard limit is never fully automating. Recruiters can spot AI slop, the vague three-line paragraphs, which is why every output gets reviewed and tweaked before it goes out. One example ran an AI-drafted resume through many rounds of edits and a custom one-page template, then showed the model that template as the standard for future drafts.
Takeaway Agents can draft a folder full of tailored applications while you get coffee, and a saved skill makes that repeatable. But the model never hits send, and it never ships a draft you have not reworked into something that sounds like you.
Beyond the big chat assistants, a handful of specialized tools cover specific parts of the hunt, and beneath all of them sits the one advantage AI cannot give you. The supporting cast each does one thing well. Perplexity is an AI search engine for reliable, sourced answers. NotebookLM turns reports into podcasts and analysis for heavy research. Jack and Jill and Sonara are AI job-search engines that take your resume and preferences and ping you, sometimes over WhatsApp, when matching roles appear. Quick CV and Teal handle resume optimization and application tracking. For mock interviews, ChatGPT's voice model is the pick. A practical caution rides alongside. Always read the employer's instructions, because some explicitly forbid AI in applications or interviews, and how a company handles that says something about its culture.
The closing point cuts against everything technical. AI makes the process faster and easier, but it cannot win it for you, because the real move is to out-human the machine. That means genuine human connection rather than avatar-to-avatar messaging or templated outreach. Talk to real people, go to events, build a network, and show enthusiasm and conviction. Be audacious rather than picture-perfect. The decisions are still made by humans, and that is the part worth pouring effort into.
Takeaway Specialized tools cover search, research, and interview practice, but none of them is the edge. The edge is human contact, the networking and the conviction a model cannot fake, since people still make the final call.
How do you get notified when a startup gets funded so you can reach out before it hires?
The mechanism is newsletters that track a given subsector and publish weekly lists of newly funded startups. You can let AI read those emails and flag the relevant ones, or simply skim them yourself. Either way, the value is timing. When a startup's raise hits the press, recruiting usually has not started, which gives you a few weeks to a month to begin networking and break the ice with a quick note of congratulations before any role is posted.
With data centers running heavily on gas power, how do you justify using AI while pursuing sustainability work?
There is no clean answer. Part of it is restraint, since you do not need AI for everything, the same way you would not drive a car for every errand when you could walk or bike. Part of it is potential, because AI can analyze data and find patterns that may help solve sustainability problems, and if the people who most want to help refuse the tools on principle, it is unclear who steers them toward good ends. And part of it is self-interest, since avoiding AI entirely can be a disadvantage when employers, even climate ones, expect it. The honest position is a balance, using it where the use is justified rather than at either extreme.
When a job post says it uses only human reviewers and no AI, what does that signal and how should you respond?
It signals an organization willing to spend more to find the right person, one that values authenticity and alignment over keyword-matching. Even Anthropic asks applicants not to use AI. The practical effect is freeing. You can drop the ATS keyword game and write a real, creative application that tells your story and connects on a human level, while still using AI for light cleanup like grammar.
How do you choose which model tier to use for a task when stronger models cost more?
Sort tasks into bike, car, and spaceship. A spaceship, like a large and complex research report, justifies the most advanced models, the thinking or deep-research modes and something like Opus 4.7. A car, like drafting or redrafting a resume or cover letter, runs fine on moderate default models. A bike, like asking for a simple recipe, needs nothing fancy. For most of a job search, mid-tier models are enough, with deep research saved for genuine deep dives.
AI literacy — The practical ability to work with AI tools, framed here as a baseline job skill rather than a specialty.
Hallucination — When a model produces confident but false or invented information, reduced by custom instructions and newer models.
ACE model — A prompting structure with three parts, Aim, Context, and Example, designed to minimize how much the model has to guess.
Custom instructions — Account-level settings that constrain a model's behavior, such as forbidding invented answers and requiring sources.
Chatbot — A single AI conversation with no lasting memory, suited to quick, one-off tasks.
Project — A focused AI workspace with its own files, instructions, and memory, called gems or workspaces by some providers.
Agent / Co-Work — An AI mode that can act across files and tools with some autonomy, available in Claude's desktop Co-Work tab.
Claude skill — A saved, reusable process an agent can call again later, created by asking the model to convert a completed workflow into a skill.
AI slop — Generic, low-effort AI output, often vague three-line paragraphs, that recruiters can recognize.
ATS (Applicant Tracking System) — Recruitment software that screens applications using semantic matching, skill assessment, and automated filtering.
Deep research — A higher-capability mode that pulls from many sources to produce a thorough report.
STAR stories — Structured accounts of past achievements covering Situation, Task, Action, and Result, used in applications and interviews.
Master resume — The most complete record of your experience, kept as a source file rather than the version you submit.
Out-human the machine — Competing on the human strengths AI cannot replicate, mainly genuine connection and networking.
Work Ally — Career coaching practice focused on professional and climate career transitions.
The AI-Augmented Job Market — Article making the case for upskilling in AI for the job search.
Custom instruction examples — Sample custom instructions for reducing hallucinations.
Job hunt project instructions — Starter prompt for setting up a job-search project.
Work Ally AI podcast — Series on using AI for the job search.
Magda Kufrej on LinkedIn — Professional profile and ongoing writing on AI and careers.
Book an intro call — 30-minute Climatebase Fellows introduction call.
Perplexity — AI search engine that returns sourced answers, used for deep research and finding niche newsletters.
NotebookLM — Tool that turns reports into podcasts and analysis for heavy research tasks.
World Economic Forum — Source of the job creation, displacement, and skills-transformation projections.
International Monetary Fund — Source of the estimate that 60 percent of jobs in advanced economies are exposed to AI.
International Energy Agency — Source of the projection that AI electricity demand could more than double by 2030.
Anthropic — Author of the report mapping AI's theoretical capability against its current use by occupation.