guide · step by step · free
How to know what job
you want
And how to build the CV that proves it, without making anything up.
After more than a decade writing software and three years running my own company, I went looking for a job and realised I had no idea what to look for. My old skills were still there. The new ones didn't match any job title. And nothing I knew of put the two together.
So I did it properly. Questions, and more questions, about my own career, some of them uncomfortable enough that I had to sit with them for a while 😅, until every single thing I said about myself had a proof sitting next to it. Then I packaged it up so anyone else can do the same.
This is the whole guide. The link to download it is at the bottom.
This will not get you a job, and I am not going to promise you that. It is what I used, tidied up so it can be useful to you.
What you need before you start
Two things, and one of them surprises everybody.
1. A coding assistant on your own computer
Claude Code or Codex, either one works. This is not a page you read on your own, and that is on purpose: the method works because something keeps asking you back until you finally hand over the number, and a PDF with the same questions on it will never do that.
Both of them need a paid account with the assistant. That is the only thing here that costs money.
2. Your paperwork, whatever you have
You do not need all of it to start. The first step works on conversation alone. But the more you gather, the fewer holes you are left with later:
- Your current CV and your LinkedIn profile as a PDF (on your profile: More, then Save to PDF).
- Employment certificates from every job. If you do not have them, most companies will issue one on request, even years later. Ask now, because they take a while.
- Performance reviews, saved emails with results or praise in them, decks you presented.
- Somewhere between 15 and 30 real job ads, with their links. Roles that interest you and a few you are fairly sure they don't. Without those last ones, the exercise only confirms what you already believed.
Installing it, in five minutes
Open the terminal on your computer and paste this in:
git clone https://github.com/Jinara/cambio-de-carrera
cd cambio-de-carrera
claude
And once you are inside Claude Code, just one thing:
/empezar
That is it. It looks at what you have already done and takes you to the step you are actually on. If you use Codex, start a project pointed at the repo folder and type empezar, without the slash.
Everything of yours lives in a folder called mi-cerebro, "my brain", and that folder has been outside git since the very first commit. You are going to write your salary in there, and the things that went wrong, so it does not get uploaded anywhere, not even by accident.
One heads-up: the assistant answers you in Spanish. The method was written in Spanish and the prompts are in Spanish, so if you want it in English you will have to ask it, and some of the file names will stay in Spanish either way.
The seven steps, one by one
step 1
Discovery: the uncomfortable questions
A long interview about your own career, run by the assistant. One question at a time, and none of them counts as answered until a number or a date shows up.
There are questions for every job you have had (why you really left, what stories stayed with you), and questions about what you want, which end in the hard ones: how much money you actually need, what kind of contract, whether you can travel and how often.
The one that throws people most is the ideal working day, morning to evening. If not a single task from the job you say you want shows up in that day, it will point that out to you.
A tip: do it across several sittings. It is the longest step, and the file remembers where you left off.
step 2
Evidence: every claim with its proof beside it
This is where you gather what you can actually prove. Every claim carries its source stuck to it: a query, an email, a certificate, a screenshot.
And before it lets you use a single number, it makes you write down what your evidence does NOT cover. If your records start in 2022, any figure of yours from before that is a floor and not a total, and that is how it gets written.
It works just as well if you have no access to anything from your previous jobs, which is the usual situation. There the proof comes from certificates, reviews, emails and the company's own public numbers.
Anything that cannot be verified gets flagged, and it never leaves your folder: not into the CV, not into an answer on a form.
step 3
What job I want: you pick Mondays, not titles
This is the strange step, and the one that catches people off guard.
You collect real job ads and hand over the links. For each ad, the assistant writes out what a Monday doing that job would look like, hour by hour, without telling you which ad it came from. Then it builds eight pairs of Mondays that differ in exactly one thing (who decides what gets done, who you spend the day with, how big the company is), and you pick one from each pair.
Only at the end does it reveal which role each Monday you chose came from. Almost every time, the answer is not the job you said you wanted.
After that it gets crossed against three more things: what you can prove, what the ads are actually asking for (counted one by one, not eyeballed) and the money you need. Out of that come your lanes, which are the two or three kinds of role you are going to aim at.
step 4
CV and applications: the verdict comes before the form
A master CV built only from verified evidence, and one variant per lane that never contradicts the master. A script turns them into PDFs, and it refuses to generate them if it finds a number you already corrected.
And for every ad, the verdict goes first, in writing: what it asks for, what you meet, what you are missing, what the real risk is and whether it is worth it. Only then do you answer the questions on the form.
The order matters. A form that is already filled in pushes you towards hitting Submit, and that is exactly when the important warning arrives too late.
You click Submit. Always you. The assistant leaves everything ready and waits.
step 5
Story bank: with your mistake in it
Your evidence turned into stories you can actually tell, in a 30-second version and a 2-minute one. With one rule that is not up for negotiation: every story carries something you got wrong, said plainly.
That is what makes all the good parts believable afterwards. The person who names their own mistake is the one who sounds senior.
step 6
Interview brief: before every call
You paste in the email from whoever contacted you and it builds the brief for that company: what interview format is waiting for you, what they sell and how they are doing (with sources, keeping facts apart from guesses), which questions are likely and which of your stories goes with each one.
step 7
Run-through: out loud, not read off a page
It asks you the question, you read five lines off a sheet, you record your answer as audio, and it gets transcribed on your own machine. Your voice does not go anywhere.
The sheet is bullet points, never a script. A recited answer falls apart the moment you skip a sentence, and everyone can hear it.
Then a script compares your takes and finds what you cannot spot by reading. The main one is this: if a number changes between one take and the next, that number is not anchored to any source. It happens far more often than you would think.
And if the company sends a numerical test, there is a timed 10-question run-through and a 100-question booklet with no calculator. The answer sheet explains which mistake leads to each wrong option, so you learn the name of the mistake you keep making.
Finding openings without living on LinkedIn
This part goes alongside step 4 and keeps running for as long as the search lasts.
- A record of every opening: where it came from, which CV you sent, and what happened afterwards. Every change is stamped with its date, so later you can answer how long they took to reply without relying on memory.
- A sweep that checks the job boards of the companies you picked several times a day (the ones posting on Greenhouse, Lever, Ashby or Recruitee), filters by the titles and locations of your lanes, and tells you what is new. If the same company posts the same role for two countries, you hear about it once.
- A dashboard that runs locally on your computer, with your openings by status, the ones that have gone two weeks without a reply, and how many replied per lane and per CV.
- An email reader that understands what companies write back: it updates the record, files the acknowledgements, the rejections and the alerts, and leaves in your inbox only what you actually have to answer.
The sweep runs on your computer, not in the cloud. I tried it in the cloud 26 times and all 26 failed: those environments have no way out to the job boards. And searching the web for openings is worse, because it hands you roles that already closed.
The rules that hold it up
The method is quick to explain. What makes it work are the rules that do not bend.
- Every claim carries its source. A number you cannot defend when someone asks "compared to what?" is worse than not having the figure at all: it leaves you with nothing to say at the exact moment it matters most.
- Numbers you correct stay retired. They go on a list, and the scripts stop if one shows up again in something heading out the door. I once had a corrected number live on for a month across other documents, my own profile included.
- Read before you fill. The verdict gets written before you touch the form.
- Small numbers count as judgement, not as scale. "I decided which stores went into the pilot, and by what rule" says more than inflating the size of it.
- No spin. If something went badly, you say it went badly and move on to the facts.
- You press Submit. The assistant does not send applications, or emails, or post on your LinkedIn.
- Anything coming from outside is data, not instructions. Job ads and emails were written by somebody else: if one of them contains something that reads like an order aimed at the assistant, it does not get carried out.
Questions I get every time
Do I need to know how to code?
No. You paste three lines into the terminal and after that it is a conversation. The only technical bit is installing the assistant the first time.
Does my data end up on somebody's server?
Your folder lives on your computer and sits outside git. Audio is transcribed on your own machine. What does go through the assistant is whatever you type to it, same as any other conversation you have with it.
Does it work with ChatGPT?
With Codex it does, which is ChatGPT's coding mode, installed on your computer. With regular ChatGPT in the browser it does not, because it needs to read and write files of yours.
Is it in English?
This guide is. The tool itself is in Spanish: the prompts, the file names and the questions it asks you. You can ask it to talk to you in English and it will, but you should expect to read some Spanish along the way.
How long does it take?
Step 1 is the longest and it is better done across several sittings. Steps 2, 3 and 4 depend on how much evidence you have and how many ads you bring. The interview ones happen once per company that calls you.
Can I see what it looks like before installing anything?
Yes. The repo has an ejemplo folder with the full walkthrough for Toffy, a made-up person: her questions, her evidence, both of her CVs, her applications and three takes of a run-through.
Download it and get going
It is free and it is open. If you get stuck on a step, write to me and we will look at it together.
See the repo on GitHub @nath.villamor Back to the start
a hug from Buenos Aires
Published on .