Give the model your facts, tell it plainly that it may only rephrase what you've given it, and then check every line it returns against what you actually did. That's the whole method. The model is good at wording and bad at knowing what's true about you, so your job is to supply the truth and audit the wording.
Where people get into trouble is the reverse: pasting a job description and asking for "a resume for this role". The model will happily produce one, with confident numbers and skills you've never used, and some of it will end up in front of an interviewer.
What the model makes up, and why
Language models write the most plausible next words. A plausible resume bullet has a metric in it, so when you ask for a "stronger" bullet and haven't given a metric, it supplies one. "Improved API response time by 40%." You never measured that. Nobody did.
The same goes for skills. Paste a posting that mentions Kubernetes and ask it to tailor your resume, and Kubernetes may quietly appear in your skills list or inside a bullet about a deployment you did with a shell script.
The usual inventions, roughly in order of how often they slip through:
- Percentages and numbers that sound right
- Tools and skills lifted from the job description
- Scope inflation: "led" when you contributed, "architected" when you implemented
- Outcomes nobody measured, like "improving customer satisfaction"
None of this is malicious. It is what the model does when it's asked to make something sound impressive without enough facts. And it's a problem for you, because an Indian tech interview is very often a walk through your resume, line by line. "You reduced latency by 40%. How did you measure it? What was it before?" If the number came from a chatbot, you have nothing to say.
There's also background verification at many companies, which checks titles, dates and employers. An inflated title is a much bigger problem than a weak bullet.
Start with a facts file
Before you open any chat window, write down what you actually did, badly worded is fine.
Say you're a backend engineer two years into a services company. Your raw notes might read:
Worked on the claims processing module for an insurance client. Java, Spring Boot, Oracle. Wrote the batch job that processes claims overnight; it used to take about 6 hours and after I changed the queries and added parallel processing it took under 2. Fixed a lot of production bugs. Onboarded two new joiners. Did the migration of 3 services from WebLogic to Docker on the client's servers.
That is a gold mine. Every fact in it is yours and every number you can explain. The model's job is to turn it into four clean bullets and add nothing.
If you don't know a number, leave it out of the notes. Don't estimate one for the model to polish. "Much faster" in your notes should stay "much faster", or better, describe what changed: "the batch now finishes before the business day starts".
Prompts that keep it honest
The instruction that matters most is the constraint. Something like this works with ChatGPT, Claude or any of the others:
Here are my notes about my current role. Rewrite them as 3 to 5 resume bullets. Use only the facts in my notes. Do not add numbers, tools, technologies or outcomes that aren't in the notes. Do not upgrade my role (for example, "contributed" must not become "led"). If a bullet would be stronger with a detail I haven't given, ask me for it instead of guessing.
That last line changes the conversation. Instead of inventing a metric, a well-behaved model will ask "how many claims does the batch process?", and you can answer or say you don't know.
For tailoring to a specific posting:
Here is a job description and my current resume. Suggest how to reword my existing bullets so they use the job description's terms where they genuinely describe the same work. Don't add any skill or experience that isn't already in my resume. Then list, separately, any requirements in the job description that my resume doesn't show, so I can decide what to do about them.
The separate list is the useful part. It tells you where you're a weaker fit, which is information you'd want before the interview, rather than papering over it. Our guide to tailoring your resume for each job goes into where rewording ends and fabrication starts.
Check every line
When the bullets come back, read each one and ask three questions.
Is every fact in it something I did? Is every number one I can explain, including how it was measured? Could I talk about this bullet for five minutes if an interviewer asked?
If any answer is no, edit or delete. Also watch the verbs. Models love "spearheaded", "orchestrated", "drove" and "championed". Interviewers notice when a two-year engineer has spearheaded six initiatives. Plain verbs (built, wrote, fixed, moved, reduced) read as more credible.
And read it as a whole. An AI-polished resume tends to have every bullet in the same shape, same length, same rhythm, each ending with an impact clause. Real resumes are a bit lumpier. Vary a few, and cut any line that is there only because it sounded good.
Where AI genuinely helps
Used well, it's very good at some resume chores:
- Turning messy notes into tight, parallel bullets.
- Spotting vague lines ("worked on various features") and asking what you actually did.
- Suggesting the industry's name for what you did, so "made the pages load faster by caching" can become "added Redis caching for product listing pages".
- Cutting a two-page draft down without losing the important parts.
- Checking grammar and consistency of tense and date formats.
Item 3 deserves a caution. Using the posting's vocabulary for work you actually did is fine and sensible. Using it for work you didn't do is the problem we started with. More on making bullets specific in writing bullet points that get read.
A word on privacy
Your resume has your phone number, email, current employer and sometimes your CTC in it. Before you paste it into any chat tool, consider removing the contact details; the model doesn't need them to rewrite a bullet. If your notes mention a client by name, check whether your employment contract is fine with that leaving the building. Swapping "an insurance client" for the real name costs you nothing.
If you'd rather not write prompts
OpenRoles builds a tailored version of your resume for each posting by re-saying what you've already done in the posting's words. If you ask it to add something that isn't on your resume yet, it does, and tells you what it assumed so you can correct it. Whatever tool you use, the audit step is still yours: read every line before it goes out under your name.