Job hunt AI pipeline
The AI writes the prose, which was always the cheap part. What took the work was one store of verified evidence and a step that is allowed to refuse to write.
I set out to stop rewriting the same four paragraphs about myself every week. The letter was never the slow part. Reconstructing what I did seven years ago, in numbers I can stand behind, was. So I built one place where that truth lives once.
One profile, and a list of what I must never claim
Everything starts from one repository of verified evidence about me. Next to it sits the most useful file in the system: the things I must never claim. Every step reads both before it writes a word.
Run a job through it
I paste a link into a chat, it lands in a queue, and a pipeline takes it from there. Three kinds of thing act in it: plain scripts, agents that read and decide, and me. Try both jobs.
The step that refuses to write
The gate is the part I would keep if I had to throw the rest away. Any model can turn a job ad into a confident CV. This one stops, asks me, and waits.
What needs me today
Every job is one row, and every view of it is generated. The number that matters is how long each application has been silent. The tracker once stopped for fifteen hours and looked exactly like a quiet week, so now it says when it breaks.
If you want to build your own
- One store of verified evidence about you. The facts a CV gets written from.
- A list of things you will never claim. This is the load-bearing one.
- A queue, not a conversation. The chat only drops the link in. Checking for work should cost nothing.
- A requirement-to-evidence table per job. Quote the ad; never summarise it.
- A step that is allowed to stop. It asks you instead of inventing.
- One record, and generated views. Never edit a view.
The AI writes the prose, which was always the cheap part. The work was deciding what is true and building something that refuses to walk past a gap.