YOUR RESUME DOESN'T NEED A BETTER PROMPT. IT NEEDS A SOURCE OF TRUTH.
- Jul 28
- 4 min read
THE FAILURE MODE ISN'T BAD WRITING. IT'S CONFIDENT INVENTION.
I have generated something like 87 résumés with AI this year, and most of that time I was doing it wrong in a way that felt right. Paste the job posting. Paste the résumé. Ask for a tailored version. What came back was always well written. It was also, maybe one time in three, quietly false — a certificate marked "in progress" that I had already earned, a freelance brand promoted to LLC, a date nudged a month.
Never a hallucinated employer. Just small, confident wrongness, in a different place every time.
The variance is the actual problem. I could not predict which detail would drift, so I had to proofread every line of every version as if I had never seen it, which defeats the entire point of the exercise.
Meanwhile the thing on the other end has gotten less forgiving:
The interesting wrinkle is why those systems misfire: they lean almost entirely on self-reported candidate information, which makes them susceptible to inaccuracies, and they will drop a genuinely qualified person whose profile doesn't match the posting's exact criteria. Newer research is moving off keyword matching altogether — Bevara et al. (2025) built a transformer-based matcher that embeds the whole résumé and job description and beat conventional systems by up to about 16% on agreement with human reviewers.
So stuffing the document with the posting's vocabulary buys less than it used to. What earns you the read is a résumé that describes you accurately and says the same true things every single time.
That is not a prompting problem. That's a job aid problem.
STEP 1: WRITE THE CAREER SUMMARY BEFORE YOU WRITE ANYTHING ELSE
I started over in the least glamorous place available: a Word document called My Career Summary. Every role, every date, every number I can actually defend. The credential language written exactly once, the way it should appear, so there is no second version for the model to drift toward.
It took an afternoon. It is easily 80% of why this works now.
Which is embarrassing, because this is just front-end analysis wearing a different hat. You don't design the intervention before you establish what's true. I have spent 12 years saying that to clients and apparently needed to hear it about my own résumé.
STEP 2: PUT IT IN A PROJECT, NOT A CHAT WINDOW
Then I built a Claude project and dropped the summary into it. Boring step. Also the one people skip. A project is persistent context — the file sits there across every conversation, so I am not re-uploading it or re-explaining myself at the start of every application.
That matters more than it sounds, because of how skills actually load. Anthropic's documentation describes progressive disclosure: only a skill's name and description sit in context at rest, and the full instructions load only when a request matches (Anthropic, n.d.). The instructions can point at reference material — but only if that material lives somewhere durable. The project is that somewhere.
STEP 3: THE SKILL IS MOSTLY GUARDRAILS
Here is what surprised me: Almost none of the skill is about writing.
The frontmatter description does the discovery work. It is the only part loaded at startup, and it's what my request gets matched against, so it has to say both what the skill does and when to use it. "Tailors a résumé" is not enough. You want the phrases you would actually type.
The body is rules, and most of them are negative. Never call the freelance brand an LLC. Never append "Expected" to the certificate I already hold. If a posting names a tool that isn't in my career summary, stop and ask me instead of guessing — and once I confirm it, write it back into the summary so the source of truth grows instead of going stale.
The rule I think does the most work is the dullest one: within a given mode, use the same layout every time. Consistency is not a side effect of a good prompt. You have to ask for it out loud.
And the skill asks me one question before it builds anything — ATS, or a human recruiter? Those are two different documents. ATS mode strips every column, icon, and brand mark, because a text box is where a résumé goes to die. Recruiter mode gets the branding back.
TAKE THE FILE:
It's linked at the bottom of this post. I pulled my own history out and left the structure, the guardrails, and the workflow in, so you can fill in the brackets and run it.
One caution about any skill, mine included: Anthropic's own guidance is to use skills only from sources you trust and to audit everything bundled inside before running it, because a skill directs the model to take actions on your behalf (Anthropic, n.d.). Read it before you use it. That isn't modesty. It's the correct default.
So I'll put it to you — if the system reading your résumé is weighing meaning rather than counting keywords, what's actually sitting in your source of truth, and when did you last check that it was still true?
Comments