A formula for resume bullets that survive the recruiter scan
I have read more than a thousand data resumes from the hiring side. The ones that get interviews are rarely better written. They are structured so that a tired person can extract the point in six seconds.
By Lin S., founder and coach at Open Loop
When I screen a stack of resumes I am not reading them. I am scanning for a reason to keep going. On a busy week that is forty resumes in an evening, and most of them look identical, because most of them were written the same way: open the job description, mirror its language, describe what you were responsible for.
Responsibility is not what gets evaluated. Everyone in that stack had roughly the same responsibilities, because they all had roughly the same job title as you. What separates two candidates with the same title is what changed because one of them was there.
A job description says what you were assigned. A resume should say what happened.
What a weak bullet is actually missing
Here is a bullet I see some version of nearly every week, next to the version I would have written after twenty minutes of asking the candidate questions.
Before
Responsible for building dashboards and reports to track key business metrics for the marketing organization.
- No baseline, so there is nothing to compare against
key business metricscould mean anything at allresponsible fordoes not tell me it ever shipped
After
Replaced 14 hand-maintained marketing reports with one dbt model and a single dashboard, cutting weekly reporting work from 6 hours to 20 minutes and becoming the view the CMO opens on Monday mornings.
- Names the scale: 14 reports, not
multiple - Gives a before and an after, so the number means something
- Ends on who behaves differently now
The second version tells me this person understood why anybody wanted the dashboard. At the screening stage that is close to the only signal I am after, because it predicts how they will behave when the requirements are vague, which is most of the time.
The four parts
Break that rewrite apart and there are four moves in it. I am not going to pretend this is a law of nature, but I have yet to see a bullet fail a screen when it has all four, and I see bullets fail constantly when they have one.
Replaced 14 hand-maintained marketing reports
Scope. How big was this? Fourteen is a number I can picture. Multiple and various are words people reach for when the real number is two.
with one dbt model and a single dashboard,
Method. Your tools belong in a sentence a human reads, not only in a keyword block at the bottom of the page.
cutting weekly reporting work from 6 hours to 20 minutes
Delta. A before and an after. One number on its own is unreadable, because I cannot tell whether it is good.
and becoming the view the CMO opens on Monday mornings.
Consequence. Who does something differently now. Adoption by a named role beats any percentage you could invent.
You do not need all four in every line. Forcing it makes a resume read like a filled-in form, and by the fifth identically shaped bullet I stop believing any of them. What you do need is scope and consequence somewhere in the top third of the first page, because that is genuinely as far as I get before deciding.
Three more, with the reasoning
Before
Built a churn prediction model using XGBoost achieving 0.85 AUC.
- AUC without a baseline is a number, not a result
- Nobody outside the team knows if 0.85 is good here
- No sign the model was ever used by anyone
After
Built a churn model that lifted AUC from 0.71 to 0.85 over the existing rules engine, then worked with lifecycle marketing to target the top two deciles, which cut monthly involuntary churn by 0.4 points on a 1.2M subscriber base.
- The baseline is the old system, which is the honest comparison
- Names who shipped it with you, which reads as collaboration
- 0.4 points on 1.2M is a real number I can size
That last bullet is long. Longer than the advice usually allows. I would rather read two lines that tell me something than four that tell me nothing, and so would every hiring manager I have compared notes with.
Before
Ran A/B tests on the checkout funnel and analyzed results to provide recommendations to stakeholders.
- How many tests? Over what period?
provide recommendationsavoids saying whether anyone listened- No hint of what you were testing or why
After
Designed and read out 23 checkout experiments in 18 months, including the guest checkout test that shipped to 100% and added roughly $2.1M in annualized revenue, and the one-click upsell that I recommended killing after it moved conversion but cut 30-day retention.
- Volume establishes you have actually done this repeatedly
- One win and one kill is far more credible than two wins
- The kill shows you look past the primary metric
Before
Leveraged AI tools including ChatGPT and Copilot to improve productivity and accelerate analysis workflows.
- Reads like a line added because everyone is adding one
improve productivityis not measurable- No sign you know where these tools are wrong
After
Built an internal LLM helper that drafts first-pass SQL against our warehouse schema, with a validation step that runs the query and flags row-count anomalies before a human reviews it. About 40 analysts use it weekly and it cut median time-to-first-query from 25 minutes to 6.
- Shows judgment about verification, which is the actual interview question
- Adoption number proves other people found it useful
- Specific about what the tool does and does not decide
But I do not have numbers
This is the objection I get most, and roughly four times out of five the numbers exist and the candidate has not gone looking. Fifteen minutes of digging usually produces enough. Some places to look:
- Your BI tool's own logs. Looker, Tableau, and Mode all track viewers per dashboard. If 60 people opened your dashboard last month, that is a number and it is on your side.
- Git and ticket history. Count the models you own, the tables downstream of you, the tickets you closed in a quarter. Dull, verifiable, better than nothing.
- Time, before and after. How long did this take a person before you automated it? Ask the person. They remember, usually with feeling.
- The denominator. You may not know your feature's revenue, but you probably know the order volume of the surface it sits on. “On a checkout flow handling 40k orders a day” sizes your work honestly without claiming credit for the whole thing.
- A stated estimate. “Roughly $2M annualized, based on the 90-day holdout and flat seasonality” is completely acceptable. Interviewers respect a number with its assumptions attached far more than a suspiciously round one with none.
If you genuinely cannot find a number, name the decision instead. “Recommended sunsetting the loyalty tier after finding the retention lift did not survive controlling for tenure, and the tier was retired that quarter” has no percentage in it and is still a strong bullet, because something happened.
The team-impact problem
Senior candidates get stuck here in the opposite direction. The work was real but it was six people over a year, and writing I feels like stealing. So they write the team and hand me a bullet with no author.
Name the team result, then name your slice inside it. “On a five-person effort to rebuild attribution, I owned the identity resolution layer and the backfill of 3 years of history” is honest and specific, and it tells me your scope, which is the thing I am trying to level you on. Vagueness reads as modesty to you and as unverifiable to me.
What to cut
Include
- A one-line summary only if you are switching fields or levels, and only if it says something a bullet cannot
- Tools named inside the bullets where you used them, plus one compact skills block for the keyword screen
- One project you argued against, or one that failed and taught you something
- Numbers with their denominators attached
- Scope markers: team size, data volume, who consumed the output
Cut
- “Detail-oriented team player with a passion for data-driven insights”
- Skill bars, star ratings, and any chart rating your own Python ability
- Coursera certificates, once you have real work to show
Utilized,spearheaded,leveraged, andresponsible forin every case- A photo, your full address, and your GPA if you graduated more than three years ago
- The objective statement, which has said nothing since about 2009
How to actually run the rewrite
Do not start with the resume. Start with a blank page and write down everything you did in the last two years, unfiltered and badly. Then go through and ask one question of each item: so what? Keep asking until you reach something a person outside your team would care about. That answer is your bullet, and the original item is just the setup.
- 01
Dump everything, badly
30 min
Every project, migration, dashboard, and fire drill from the last two years. No editing. Aim for thirty items, not ten.
- 02
Ask `so what` until it stops working
40 min
For each item, keep asking who cared. Some items die here, which is the point. A resume with eight strong bullets beats one with twenty even ones.
- 03
Go find the numbers
20 min
Dashboard view counts, ticket counts, timing before and after. Write down the assumption next to any estimate so you can defend it later.
- 04
Write for scope, not for length
20 min
Two lines is fine if both lines earn their place. Put your largest-scope bullet first inside each role, not the most recent one.
- 05
Read it in six seconds
10 min
Look away, look back, and read only what your eye lands on. If that is not the strongest thing about you, reorder until it is.
One last note, since it comes up in every teardown I run. Nobody is rejecting you over a template, a font, or one page versus two. Those debates are loud because they are easy. The reason a strong candidate gets screened out is almost always that the bullets described the job instead of the person doing it, and that is a fixable problem in an afternoon.