--- title: "What a Resume Roaster Actually Checks Before It Says a Word" description: "Before any AI writes a joke about your CV, a few dozen lines of plain pattern matching have already scored it. Here is exactly what those checks look for — buzzwords, vague verbs, dates, length — and why each one is a real problem rather than a punchline." canonical: "https://www.razi.pro/blog/things-ai-roast-finds-in-every-resume" date: "2026-09-07" tags: ["Resume", "Career", "ATS", "AI", "How-To"] source: "razi.pro" --- # What a Resume Roaster Actually Checks Before It Says a Word Before any language model sees your CV and writes something unkind about it, a much dumber process has already run: a few dozen lines of pattern matching that count words, look for specific strings, and score what they find. That boring layer is the interesting one, because it is roughly what an applicant tracking system does to you as well. The jokes are downstream. The patterns are the substance. Here is what [the CV roaster](https://www.razi.pro/games/cv-roaster) actually looks for, taken from its checks rather than from what it says afterwards, and what each one is really diagnosing. ## Buzzwords, in three distinct flavours The detector holds a list of roughly forty terms and counts every occurrence of each. What makes the list interesting is that it contains three different kinds of failure, and only one of them is the one people expect. **The classic empty adjectives.** *Results-driven, team player, go-getter, self-starter, detail-oriented, passionate, innovative, dynamic, proactive, strategic.* These are not lies. They are unfalsifiable. No candidate has ever described themselves as careless and reactive, so the words carry no information and cost you a line that could have carried some. **The bravado nouns.** *Rockstar, ninja, guru, thought leader, 10x engineer, growth hacker, hustler, visionary, disruptive, paradigm shift.* These date a CV badly, and not in your favour. **The technology words that have expired.** This is the category people miss, because these were genuinely impressive to write once: *blockchain, web3, metaverse.* Listing them now signals when you last updated the document rather than what you can do. And a fourth group worth watching, currently on the list precisely because it is on the way to becoming the third: *ai-powered, agentic, multi-modal, rag pipeline, fine-tuning, prompt engineer.* These are live terms today. The reason they earn scrutiny is that they are becoming the new default garnish — sprinkled on a CV to signal currency rather than to describe work that was actually done. The test for all four groups is the same. **Delete the word. If the sentence still says what you did, the word was decoration.** ## Vague verbs, which are the real problem The second list is only ten phrases, and it is more diagnostic than the buzzwords: > improved efficiency · increased productivity · managed team · responsible for · worked on · helped with · assisted in · contributed to · participated in · involved in Read those again and notice what they share. Every one describes **proximity to work rather than work.** "Responsible for" tells a reader you were in the room. "Contributed to" tells them somebody else could also claim the same achievement. They also, almost without exception, have no number attached. "Improved efficiency" is a claim with the evidence removed. By how much? Over what period? Compared to what? The fix is mechanical and it is the single highest-value edit on most CVs: > **Before:** Responsible for improving efficiency of the deployment process. > > **After:** Cut deployment time from 90 minutes to under 5, across 20 services. Same fact. One of them can be interrogated in an interview, which is exactly why it is believed. ## Length, measured crudely and on purpose Page count is not read from the file. It is estimated: pages = ceil(wordCount / 500) Five hundred words per page is a rough number and it is deliberately rough, because the actual question is not how your PDF paginates. It is whether you have written more than someone will read. A CV that estimates to four pages is making a claim on a stranger's attention that most careers do not support. ## Common misspellings, checked by name There is a short list of specific misspellings rather than a general spell-checker: > experiance · managment · recieve · occured · seperate · definately · accomodate · occassion These are the ones that survive a spell-check because several of them are near-misses people re-approve without reading, and because "experiance" in particular tends to appear in the section header where the eye stops looking. ## The date check, and the flaw worth knowing about Job count is derived by matching date ranges. The pattern looks for a four-digit year, a dash, and either another year or the word "present" or "current". Which means these are counted: 2019 - 2023 2021 – present March 2020 - June 2022 And these are not: 2019 to 2023 03/2020 - 05/2023 Jan '21 – Present If your dates are written in one of the second group, the roaster concludes you have zero jobs. That is a bug in a toy. **The same brittleness in an applicant tracking system is not a toy**, and it is the reason this whole category of tool is worth paying attention to: your CV is being read by pattern matchers with narrow expectations, and formatting choices that look like taste are actually parser input. Write your dates as a four-digit year, an en dash or hyphen, and either a four-digit year or the word "Present". It costs nothing and it removes a class of failure you would never otherwise see. That thread is picked up properly in [what an ATS actually rejects in your resume](https://www.razi.pro/blog/what-ats-actually-rejects-in-resumes), which covers the parsers that decide whether a human ever opens the file. ## What the roaster does not know Being honest about the limits, since the whole point of a roast is that it is blunt rather than that it is right. It does not know your industry norms — academic CVs are long on purpose, and a list of publications is not padding. It does not know which of your vague phrases is covering genuinely confidential work. It cannot tell a buzzword you chose from one a previous employer put in your job title. And it has no idea whether you are actually good at the job. It is checking how the document reads, which correlates with getting an interview and not at all with deserving one. ## Where your CV goes Worth stating precisely, because you are about to paste your employment history into a website. The pattern matching above runs in your browser. The text is sent to the server for the AI portion — that part cannot happen locally — but **the CV text itself is not stored.** What gets recorded is counts: word count, page estimate, how many buzzwords matched, how many date ranges were found, the resulting score. The document is not kept. ## A short checklist If you do nothing else with this page: 1. Delete every word that survives being deleted. 2. Replace each of the ten vague phrases with a number. 3. Write dates as `2021 – Present`. 4. Read the section headers out loud, where the typos hide. 5. Cut it until it estimates to under two pages. Then, if you want it said less politely, [the roaster](https://www.razi.pro/games/cv-roaster) is free and needs no account. Two related tools, if the roast turns something up: if your CV reads as machine-written, [the text humanizer](https://www.razi.pro/tools/ai-humanizer) is the relevant fix, and if you are converting between formats, [the Markdown converter](https://www.razi.pro/tools/markdown-converter) handles the plain-text version that parsers prefer.