The 60-80 percent rule is simple: let AI take your application from nothing to a structured draft, then write the final fifth yourself. Anything you cannot verify, personalise, or defend in an interview does not belong in the version you send, and that boundary is what keeps an AI resume too generic problem from reaching the hiring manager at all.
The rule exists because automation advice usually arrives in two forms. Some pages promise one-click applications. Others warn against any AI use without explaining where the line sits. Neither tells you what to keep and what to cut. Below is the split, an editing pass you can run in one sitting, and the checks that tell you whether a draft is finished.
What the first 80 percent actually covers
Drafting, not deciding. The tool does the structural work that takes longest and matters least to authorship: turning a job description into a role summary, converting messy bullet fragments into parallel action verbs, compressing a ten-year career into two lines, and catching the gaps where you left something out.
That work is genuinely useful, and it is where a well-prompted tool earns its place. The problem is that this 80 percent reads as finished because it is grammatical and complete. Nothing in the text signals where the model filled a gap.
So should i use ai for my resume becomes the wrong question. The useful question is which parts of the document you can defend line by line in a conversation with a stranger.
Draw the line at evidence, not at sections
Keep what you wrote. Cut what the model supplied. If a claim traces back to your memory, your files, or your manager's confirmation, it can stay. If it appeared because the phrasing "results-driven" was nearby and the model completed the thought, it goes.
| Line type | Source | Keep or cut |
|---|---|---|
| "Reduced invoice errors by 30 percent" | your memory, confirmed with your team | keep, verify the number first |
| "Spearheaded digital transformation" | model completion | cut, unverifiable |
| "Managed a team of eight" | model inferred from "mentored juniors" | cut or downgrade to what is true |
| Bullet restructured from your rough notes | you, reworded | keep |
| Company name and dates | you | keep |
AI application honesty is a documentation problem, not a confession problem. You cannot verify what you did not do, so it does not go in the document.
The editing pass that finishes the last 20 percent
Once a draft exists, run these steps in order. Ten minutes is enough for a two-page résumé, and you need them in this order because each one depends on the previous.
1. Cut every line you cannot source
Read the draft with a pen and mark each claim as remembered, verified, or unknown. Unknown lines are deleted rather than softened, because a softened version is still a claim.
2. Replace abstractions with the specific thing
Abstract language survives because it is unfalsifiable. Replace it with the actual name, number, tool, scope, or constraint. If you cannot supply the detail, the line is not specific enough to earn its place.
3. Re-anchor the opening to your actual situation
The first three lines carry the most weight and are the most likely to have been smoothed into generality. Rewrite them around the role you are targeting and the thing you have actually done that matches it.
4. Read it as the interviewer who will ask about it
Anything you would hesitate to explain for ninety seconds is a liability in the room. Rewrite it into something you would be glad to talk about, or remove it.
5. Compare against the posting, not against other résumés
Check that the requirements you meet are visible within the first half of the document. Recruiters scan top-down, and a matching qualification buried on the last line reads as absent.
The prompt that produces the draft, and its refusals
Rewrite my résumé bullet points so they are specific and outcome-shaped. Here is the job description: [paste it]. Here are my raw notes, including unfinished fragments: [paste them]. Use only facts present in my notes. Keep each bullet under 30 words. Start most bullets with a past-tense action verb. Do not add metrics, percentages, team sizes, budgets, or client names that I did not write. Do not upgrade my title. Do not create a skill I have not listed. Where you would need a number to complete a sentence, write [CONFIRM NUMBER] instead. After the rewrite, list every line where my notes were ambiguous and tell me what you assumed.
Fill in these details
- Job description: paste the actual posting, including the must-have list, not a summary of the role.
- Raw notes: messier is better. Fragments and half-sentences give the model something real to work from.
- Ambiguity flag: keep it in the prompt. It moves the model's guesses into a list you can correct, instead of burying them in prose you might not notice.
A worked example of the last 20 percent
Suppose a candidate's AI-assisted draft opens with five years of experience in backend engineering, then lists skills including Kubernetes, and describes leading a migration. The candidate's own notes mention helping with a migration and running three services.
| Draft line | Problem | Fix |
|---|---|---|
| "Five years of experience in backend engineering" | Accurate but says nothing | "Backend engineer, five years, payments and order services" |
| "Kubernetes, Docker, AWS" | Docker and AWS are from their notes, Kubernetes is not | Remove Kubernetes |
| "Led a cloud migration project" | Overstates helping | "Ran the service cutover for three services during a cloud migration" |
The fix column is where the last 20 percent lives. Every change traces back to something the candidate knows and can discuss.
How to judge whether a draft is finished
Run this check before you apply. If any answer is no, the draft is not ready.
- Can you defend every line? If a claim would surprise you to be asked about, it is not yours.
- Does anything read as written for every company? A paragraph that could sit in any application is doing no work.
- Are the numbers traceable? Every figure should come from a source you could name in the interview.
- Does the document match this posting? Generic means the same thing as tailored, not less impressive.
- Have you checked the employer's stated policy? Some organisations state rules about AI-written applications. Read the current wording and follow it rather than assuming.
Common mistakes when using AI on applications
- Editing from the top down. Style edits at the top let you reach the bottom without noticing invented claims.
- Accepting improvement you cannot verify. Fluent rewording of a vague line makes the vagueness harder to spot.
- Letting the model add a skill list. Keywords help in some systems, but skills you cannot discuss in an interview will surface later.
- Using one draft for every application. A single generic document is the exact output a hiring manager recognises.
- Treating length as the goal. Padding with responsibilities you did not own dilutes the parts that work.
Frequently asked questions
Should I use AI for my resume at all?
Yes, for drafting, restructuring, and tightening your own material. The risk is not the tool, it is the claims you cannot verify. Keep the evidence standard high and the tool's contribution narrow.
Does using AI on an application disqualify me?
Rules vary by employer and change over time, and most say nothing about it. Check any stated policy where you apply, and never let generated text claim experience you lack.
How do I make an AI-written resume stop sounding generic?
Generic output means the prompt had nothing specific to work from, so feed it the posting, your rough notes, and an explicit instruction not to invent metrics or skills. Then edit the opening lines yourself, since that is where specificity lives or dies.
Draft once, then defend every line
Write your raw notes, run the draft prompt, then spend ten minutes cutting unsourceable claims and rewriting the opening in your own words. Send the version you would be happy to defend out loud. For more prompts for the whole application process, browse the career development topic collection.