System architecture
Not a chatbot I ask things. A working system with its own memory, its own research, and a rule that it checks before it ships.
A job to apply for, a company to understand, a deck to build, an email to write.
Everything I have ever corrected it on. How I write, what I will not claim, which numbers are confirmed and which are not. It starts from what it already knows about me.
Crawls the site, pulls the public sitemap, checks every claim against the live page, and writes the findings to a file with a source on each line.
The CV, the tailored version of it, the deck, the tracker, the site. All rendered from one profile file, so a number changed once is changed everywhere.
Twenty-eight labelled cases with deliberate traps. Below ninety per cent accuracy the build stops. It also tells me which claims it could not verify, instead of quietly filling the gap.
Nothing is sent, published or posted without me. Drafts only, every time.
When two readings of a request would produce different work, it stops and asks rather than picking one and hoping.
Tell it once that a phrase is not how I talk, and it is written into memory. It does not need telling twice.
Built in Claude Code. The research, the build pipeline and the eval suite are separate pieces that hand off to each other, which is why any one of them can be replaced without breaking the rest.