Career guides thatrespect your time

Building a job search automation tool

A build order for your first one, where each step produces something usable and each checkpoint tells you whether to continue.

An agent that fills application forms

Filling a form is a mapping problem with a sharp edge: some fields must never be answered by a machine, and telling which is the actual work.

Building a job finder on APIs

APIs are the boring, legal, reliable way to get postings — and the schema differences between them are where the real work turns out to be.

Detecting duplicate job listings

Two postings for the same job rarely share a single identical field. Two postings for genuinely different jobs often share almost all of them.

Tracking every application

Tracking is the least glamorous component and the only one that makes the whole system improve. It is also where most implementations stop.

Scoring job descriptions

Scoring is two jobs pretending to be one: understanding what the posting asks for, and judging whether it is worth this candidate’s time.

Human approval, designed properly

An approval step nobody reads is worse than none — it provides the appearance of oversight and the accountability that comes with it.

Stopping the wrong application

Agents rarely apply to a randomly wrong job. They apply to a job that was right an hour ago, or to one of several that looked identical.

Prompt injection in job agents

A job agent reads documents written by people who want something from it, and can send email and submit forms. That combination is the whole problem.

Agent permissions and sandboxing

You cannot make an agent refuse reliably. You can make the thing it was asked to do impossible, which is a different and far more dependable guarantee.

Giving an agent browser access safely

Your browser holds every session you are signed into. An agent given that browser holds them too, and it takes instructions from web pages.

Protecting credentials in job automation

If your product signs into job boards for users, you are running a credential store — whether or not you designed one.

What happens to your resume data

A CV is a dossier: employment history, education, location, contact details, often more. Uploading one to a tool is a bigger decision than it feels like.

Making agents reliable

Ordinary software fails by stopping. Agents fail by continuing, which is why the usual reliability toolkit does not catch them.

Securing a job application platform

A platform that acts for many candidates concentrates two things attackers want: personal data at scale, and the ability to act under someone else’s name.

Securing a browser agent

A browser agent reads attacker-controlled content by design. Security is therefore about what it can do afterwards, not about what it reads.

Authentication in browser agents

The agent needs to be signed in. It does not need to know the password — and those two requirements are easier to separate than most implementations assume.

Designing the approval experience

Approval is a product design problem before it is a safety one. Ask for too much and it is ignored; too little and the gate protects nothing.

Scripts or agents?

The real answer is almost never one or the other. It is a script for the paths you know and an agent for the ones you do not.