Should I vibe code
Score a resume against transparent rules and a user-supplied job description
You can ship the score by Sunday. You cannot ship a reason to believe it.
?
Their verdict, the Pro (monthly) price and the build-time estimate come from their entry, MIT-licensed. Checked 2026-08-04.
?
Our verdict, the regret score and everything below it. Editorial and unsponsored — nobody can pay to be moved.
The honest answer
why the verdict is what it is
Two products live inside this one and they are not remotely the same difficulty. The first is a diff: paste a job description and your CV, pull the terms out of each, show what is missing. That is an afternoon, it is genuinely useful, and the Jobscan entry on this site already says so in as many words. The second is a score — a number out of 100 with line-level criticism attached, presented as though it knows what recruiters do. You can ship something that emits that number by Sunday evening. What you cannot ship is a reason to believe it. Resume Worded's number is at least anchored to a fixed rubric applied at volume; yours will be anchored to whatever the model felt like at temperature 0.7, and it will move by ten points on a rerun of the same file. Nothing here leaks, nothing here costs money, nobody else is affected. The failure is quieter than that: you will rewrite a real document to satisfy a metric you invented on a Saturday, and then send it to forty companies.
What actually breaks
not "if". the specific failures.
- Stability. The same PDF scored twice, twelve points apart, and no way to tell which run was right — a scorer that is not deterministic is a mood, not a measurement
- The rubric, which rots. Advice about keyword density and "ATS-friendly formatting" written for 2019 parsers is actively wrong against a 2026 one, and nothing tells you it has drifted
- PDF extraction, where a two-column layout interleaves into nonsense, ligatures become mojibake, and a header rendered as an image contributes nothing at all
- The advice loop, which optimises for your own scorer: stuff the keywords, watch the number climb, hand a human a document that reads like a shopping list
- Any LinkedIn scoring feature, which means either a manual paste or scraping a profile behind a login, and the second option breaks constantly and violates the terms you agreed to
- Calibration, which is the whole product and the one thing you have no data for — the real version's number comes from having seen an enormous number of resumes and outcomes
Is that you?
the verdict is a default, not a law
- You build the diff and skip the score: terms in the job ad, terms in your CV, the gap between them, and no number pretending to be objective
- It runs locally over files you already have, with nothing uploaded and no account
- You treat the output as a checklist to argue with rather than a grade to chase
- The number is going to change what you send. An invented metric applied to a real document is worse than no feedback, because it feels like feedback
- You would run it for other people — friends, a jobseekers' group, a paid side hustle — because then somebody else's search is being steered by your Saturday heuristics
- You are going to scrape LinkedIn profiles to score them, which is against the terms of service and breaks every time the markup shifts
- You expect it to predict whether an application succeeds, which is not a thing this class of tool can do, paid or homemade
If you build it anyway
the checklist, then the prompt that enforces it
- Make the scorer deterministic. Fixed rules, fixed weights, temperature zero if a model is involved at all, and a test that asserts the same input scores identically twice.
- Show the working. Every point gained or lost names the rule that caused it, so the output is inspectable rather than oracular.
- Version the rubric and stamp the version onto every score, because a number from a rubric you have since changed is not comparable to today's.
- Separate the two features in the interface: the job-description diff is evidence, the overall score is an opinion, and mixing them makes the opinion look like evidence.
- Extract text from the PDF and show it to the user before scoring anything — half the useful signal is discovering your CV parses as one scrambled paragraph.
- Never auto-apply a suggested rewrite. Suggestions are proposals a human accepts one at a time, because it is their employment history.
- Keep everything local. There is no reason for a resume scorer to have a server, and every reason for it not to.
I am building a resume scorer: it takes my CV and optionally a job description,
and tells me what to fix. The risk is not a leak, it is that I will believe a
number I made up. Build it so the number can be audited.
1. Start with the part that is real: extract the text from the PDF and show it
to me exactly as a parser sees it, before any scoring. If it comes back as
one scrambled column, that finding is worth more than any score.
2. Then build the job-description diff — terms present in the ad, terms present
in my CV, the difference, with the surrounding sentence for each. No
scoring, no grade, just the gap.
3. Only then a score, and it must be deterministic. Explicit rules with
explicit weights in a config file. If a model is involved, temperature zero
and a fixed prompt, and write the test that asserts identical input yields
an identical score across runs.
4. Every point of the score names the rule that produced it and shows the line
of my CV that triggered it. No unexplained totals.
5. Version the rubric and record the version alongside every score. Scores from
different rubric versions must never be shown on the same chart.
6. In the interface, keep the diff and the score visually separate and label
the score as an opinion generated by rules I can read and edit.
7. Never rewrite my CV automatically. Suggest, one item at a time, and require
an explicit accept — this is my employment history and a plausible sentence
is not the same as a true one.
8. Refuse to reward keyword stuffing: if a suggestion would raise the score by
inserting terms that do not describe something I actually did, say so.
9. Keep it entirely local. No accounts, no uploads, no telemetry. My CV is a
file, not a record in someone's database.
10. Out of scope, and say it rather than faking it: predicting whether I get an
interview, LinkedIn profile scraping, and any claim to replicate a specific
company's applicant tracking system. If I want a benchmark grounded in real
outcomes, tell me that is what the paid tools are actually selling.That one keeps you out of trouble. For the prompt that actually builds it, canivibecodeit.com has one.
their build prompt ↗Or don’t build it
the boring option, and the way back out
If what you want is a number anchored to something outside your own head — a rubric applied consistently across a very large number of resumes, with the ATS-parsing rules kept current — then that grounding is the product and it is the one thing you cannot generate. It is a steep subscription for a job search, though, and the cheap and honest move is to build the job-description diff yourself and spend the money on one round of human feedback instead.
There is nothing to exit. The CV is a file you already had, the rubric is a config file, and the whole tool is a script that reads both. Keep the rubric in version control and you can rerun any past score and see what changed about your opinions, which is more than the paid product offers.
Open-source resume and job-description matcher that surfaces keyword and skill gaps locally.
Active open-source resume builder with structured data and PDF export.
Questions
Jobscan is rated ship-it and this is demo-only. Why the difference?
Jobscan is scored as the diff — two documents in, a comparison out, nothing invented. Resume Worded's headline feature is a grade with authority attached, and the authority is exactly the part that does not survive being homemade. Build the half that is a comparison and you are in ship-it territory; build the half that is a verdict and you have a demo.
What if I just use an LLM as the scorer?
Then pin it: temperature zero, a fixed prompt, a versioned rubric in the prompt, and a test that the same file scores the same twice. An unpinned model is a scorer that changes its mind between Tuesday and Thursday, and you will not notice, because a plausible number looks exactly like a correct one.
Is there anything here genuinely worth a weekend?
Yes — the parse preview. Extract the text from your own CV and read it back in the order a parser sees it. Two-column layouts, header images and text boxes routinely turn a careful document into gibberish, and that discovery is worth more than any score on this page.
Every week, someone ships something they shouldn’t have.
New verdicts, the worst thing that landed in the trap, and the occasional incident report. No other email, ever.
Comparing your CV to a job ad is text analysis on two documents you already own.
Ask a model for achievement bullets and it will invent the metric, because metrics are what good bullets have.
Tracking your own applications is a table you will actually maintain because it is yours.
last reviewed 2026-08-04 · verdict is editorial and unsponsored · shared entry data from canivibecodeit under MIT · not legal advice