Part 1 of 4 · 1 min read
Overview
Agent projects are easy to demo and hard to verify. A reviewer cannot rerun your system in a minute, so the write-up has to carry the evidence: a number they can trust because it is regenerated, a baseline that shows the agent earns its complexity, and limits that show you know where it breaks.
About the example, and what is real
Real: the template renderer, the README linter, the number-provenance check, the planted-fault and drift experiments, the four templates (README, architecture note, evaluation report, demo script) and 20 tests in code/mini-portfolio. Assumed: the project ("TicketDraft"), its simulated evaluation (150 synthetic items; a rules baseline and a simulated agent), the token prices ($3 and $15 per million), the reading depths and the 5-command and 1,200-word limits. Not real: no real product, customer tickets or repository; the checks are my own rules and have not been validated against real READMEs.