A resume builder bans its AI from adding nine named kinds of fact to a user's bullets

In the developer's account, the model invents because invention is what a strong resume looks like. The developer wrote that strong resumes have numbers, tools and outcomes, so when a model sees a vague bullet, the most resume-like completion includes a number, a tool and an outcome, whether or not they are true ^[2]. In the post's example, "worked on the website" comes back as "Spearheaded a website redesign that boosted conversions by 35%" ^[3]. According to the post, that line falls apart the first time an interviewer asks how the 35% was measured ^[3].

The developer wrote that vague instructions, such as telling the model not to make things up, are not enough ^[4]. The prompt names nine categories the model may not invent ^[5]. Two other lines mattered more than the developer expected, and both are good engineering ^[6]. The first bans placeholders such as [X%] and "N users". Without it, models "politely" offered [X]% for the user to fill in, and the user was left to make up the number ^[6]. The second gives the model something to do when a bullet has no measurable result: make it clear and specific with the facts given ^[7].

"Prompts are guidance, not guarantees," the developer wrote ^[8]. The count rule stops the model slipping in an extra bullet with an invented achievement. It also stops the model merging or dropping lines, so the user can compare each bullet before and after ^[9]. When the check fails, the user sees an error. "I'd rather fail loudly than show something plausible but wrong," the developer wrote ^[10].

Put the post's own example through those layers. The count check passes it: one bullet went in, one came out. The scan for digits, % or technology names missing from the input would flag "35%" ^[12]. It would not flag "Spearheaded" or "redesign". Those two words claim a leadership role and a project the user never mentioned. The developer calls the scan "crude" ^[12]. The post lists it among the tests the developer ran and does not say it screens live responses ^[12]. As described, an invention written in plain words reaches the user. The last check is the interface telling people to read the result before they use it ^[13].

The post does not report how often either check fired, or a fabrication rate before and after the rules went in. By the developer's description, JSON output mode is there for reliable parsing and a temperature of 0.4 keeps rewrites conservative ^[11]. No measurement comes with either setting. For the approach to carry over, invention in another product would have to look the way it does here: numbers and tool names in short bullets. Where invented claims are mostly words, a lexical scan catches less.

The same habit of testing what a machine reads turned up a second defect. The PDF export is the browser's print-to-PDF of a single-column HTML preview. That keeps the text selectable, which matters because many employers run applicant tracking systems that read it ^[15]. Extracted text came back as "SKI L LS" and "CERTI F ICATIONS" because wide letter-spacing on uppercase headings made the extractor treat the gaps as word breaks ^[14]. The developer wrote that where a human sees the heading SKILLS, a parser may see three words ^[14]. Removing the letter-spacing, plus a thin divider so the headings still stand out, gave clean extraction ^[16].