Claude can tailor a resume well if you treat it as an editor working from a fixed set of facts, not as a writer. The method that works is a short loop you repeat for every posting: store your facts once in a Claude Project, paste one job posting per chat, ask Claude for a numbered list of proposed changes (a diff) instead of a rewritten resume, make it quote the source for each change, then accept or reject each line yourself. That keeps the speed of a Claude resume edit without letting a confident sentence slip a number, title or tool onto the page that you cannot defend in an interview. Below are the setup, the prompts and the checks, written so you can copy them.
If you have read our ChatGPT resume guide, this is a different workflow, not the same prompts with a new name. That guide extracts requirements and rewrites bullets in two passes. This one is built around two features Claude handles well: a standing fact sheet in a Project, and an edit list you approve line by line.
What Claude does well on a resume, and where it slips
Tailoring is mostly reading and rewording, and Claude is strong at both. It can separate a posting's real requirements from the mission statement and benefits, spot that your "cut month-end close from eight days to five" answers their "improve close efficiency", and shorten a long bullet without losing its point.
It slips in the same place every language model does. Anthropic's help center says plainly that Claude "can write things that might look correct but are very mistaken" and that users "should not rely on Claude as a singular source of truth". Anthropic's developer guide on reducing hallucinations ends with the same caution: its techniques "significantly reduce hallucinations, they don't eliminate them entirely."
On a resume the slips are small and plausible, which is what makes them risky:
| Slip | What it looks like | Why a full rewrite hides it |
|---|---|---|
| Number drift | "about 40 accounts" becomes "45+ accounts" | The new figure reads naturally in a new sentence |
| Scope creep | "supported the migration" becomes "led the migration" | The verb change is one word in a paragraph of changes |
| Tool import | The posting wants Tableau, and a bullet now "built Tableau dashboards" | It matches the posting, so it looks like a good edit |
| Title polish | "Analyst II" becomes "Senior Analyst" | Titles are easy to skim past |
| Role merge | Two roles at one employer become one line | The page gets shorter, which you asked for |
The pull toward these edits is not only on the model's side. In Huntr's survey for its 2025 job search trends report, 458 of 1,049 respondents (about 44%) said they would use AI to exaggerate their qualifications. That is a self-reported attitude from one vendor's survey, not a measure of what anyone did, but it shows how easily "make this stronger" turns into "make this up". The workflow below is designed so that neither you nor Claude can do that by accident.
Set up once: a Project or an incognito chat
You have two sensible places to do this work in Claude, and they trade convenience against privacy.
| Claude Project | Incognito chat | |
|---|---|---|
| What it keeps | Files and instructions you add, used in every chat in the project | Nothing in your history or memory |
| Best for | Applying to many roles with the same fact sheet | A one-off edit, or anyone who prefers not to store a resume |
| Limits | Free accounts can create up to five projects | Only available outside projects |
| Used for training | Follows your account's model training setting | Not used for training |
| Retention | Follows your account's settings | Kept for 30 days by default |
The details come from Anthropic's help articles on creating and managing projects and using incognito chats. Two lines from them matter for this workflow. Anything in a project's knowledge "will be used across all of your chats within that project", and "context is not shared across chats within a project unless the information is added into the project knowledge base". So the fact sheet goes into project knowledge, and each posting gets its own chat, which keeps one employer's wording from leaking into the next application.
Check your training setting either way. Anthropic's August 28, 2025 update to its consumer terms says that on Free, Pro and Max plans, data is kept for five years if you allow it to be used for model training, and for 30 days if you do not. You can change the choice at any time in Privacy Settings, and a deleted conversation will not be used for future training. Whichever you choose, strip your phone number, home address and email from the resume first. Claude does not need them.
Prompts that lock your facts
The core of the method is a fact sheet: your master resume rewritten as short, numbered lines, each with a stable ID. IDs let Claude point to a fact and let you check it in seconds. Here is a short sample, written for this post and not a real person's history:
<facts>
E1 Employer: Northwind Freight | Title: Operations Analyst | Jan 2022 to present
E1.1 Built a weekly carrier scorecard in Excel and SQL used by 6 regional managers
E1.2 Cut average invoice dispute time from 19 days to 11 days in 2024
E1.3 Supported the move from a legacy TMS to a new routing platform (testing, training)
E2 Employer: Lakeside Retail | Title: Inventory Coordinator | Jun 2019 to Dec 2021
E2.1 Reconciled cycle counts for 2 warehouses, about 40,000 SKUs
EDU B.S., Supply Chain Management, State University, 2019
CERT None held
SKILLS Excel (advanced), SQL (intermediate), Power BI (basic), SAP MM (user)
</facts>
Two lines in that sample do a lot of work. "CERT None held" stops Claude from filling an empty section, and the level next to each skill gives it something honest to say about tools you have touched but not mastered.
Put the fact sheet in project knowledge, then paste these rules into the project instructions so every chat starts with them:
You are editing my resume for one job posting per chat.
Rules that always apply:
1. The only facts about me are the lines in <facts>. Treat them as fixed.
2. Never change an employer, title, date, school, degree, certification or any
number. Never add a number that is not in <facts>.
3. You may reword a line, reorder lines and use the posting's terms for things
a line already describes.
4. If a change would need a fact I have not given you, do not make it. Add it
to a list headed QUESTIONS instead.
5. If the posting asks for something <facts> does not cover, say "not covered".
That is a good answer, not a failure.
6. Every proposed change must cite the fact ID it comes from.
Rule 5 comes straight from Anthropic's guidance: its first technique for reducing hallucinations is to "explicitly give Claude permission to admit uncertainty". Rule 6 makes every edit traceable to a line you wrote yourself.
The per-posting loop
With the project set up, each application is the same five steps in a new chat.
- Paste the posting first. Anthropic's
prompting guide
advises putting long documents near the top, above the instructions, and
wrapping each kind of content in its own XML tag, because tags help Claude
"parse complex prompts unambiguously". So wrap the posting in
<posting>tags and put your request after it. - Ask for the evidence map. One table: each requirement in the posting's words, the fact ID that covers it, or "not covered".
- Ask for the diff. A numbered list of proposed changes, each tied to a fact ID and a posting phrase. The next section has the prompt.
- Approve line by line. Reply with the numbers you accept and the ones you reject, with a reason for each rejection.
- Assemble it yourself. Apply the accepted changes to your own resume file, in your own template. Claude never produces the final document.
The evidence map prompt is short because the rules already live in the project:
<posting>
[paste the full job posting]
</posting>
Ignore the company description, pay, benefits and legal text. List each
requirement in the posting's exact words. For each, give the fact ID that
covers it and quote that fact, or write "not covered". Do not propose any
changes yet.
Read the "not covered" rows before going further. Each one is your decision. If you have the experience and left it off the fact sheet, add a new line to the fact sheet in project knowledge, in your own words, so it is available for every future posting. If you do not have it, it stays off.
Asking for a diff, not a rewrite
This step is what makes the loop checkable. If you ask Claude to "tailor my resume", you get back a whole new page, and you have to find every change by reading two documents side by side. If you ask for a list of changes, every edit is visible and can be refused on its own.
Propose changes to my resume for this posting as a numbered list. Do not write
out the full resume.
For each change give:
- Fact ID
- Current text (copied exactly from <facts>)
- Proposed text
- Posting phrase it answers (quoted from <posting>)
- Risk: NONE if it only rewords or reorders, CHECK if it changes a verb of
responsibility, adds a posting term, or changes a skill level
Limit yourself to the 8 changes that would most improve the match.
End with the QUESTIONS list.
A sample of what good output looks like, written for this post:
| # | Fact | Current | Proposed | Posting phrase | Risk |
|---|---|---|---|---|---|
| 1 | E1.2 | Cut average invoice dispute time from 19 days to 11 days in 2024 | Reduced freight invoice dispute resolution time from 19 to 11 days (2024) | "freight audit and dispute resolution" | NONE |
| 2 | E1.1 | Built a weekly carrier scorecard in Excel and SQL used by 6 regional managers | Built weekly carrier performance reporting in SQL and Excel for 6 regional managers | "carrier performance reporting" | NONE |
| 3 | E1.3 | Supported the move from a legacy TMS to a new routing platform | Supported TMS implementation: testing and end-user training | "TMS implementation experience" | CHECK |
Row 3 is the one to think about. "Implementation experience" is a stronger claim than "supported the move", and Claude marked it CHECK, which is what the rule is for. If an interviewer asked you to walk through the implementation, could you? If yes, accept it. If you only ran test scripts, reject it with a reason: "Reject 3, I tested and trained, I did not implement. Keep 'supported'." A rejection with a reason also teaches the rest of that chat where your line is.
Then ask for the skills line the same way, as a list of changes rather than a new line. If the posting wants Power BI and your fact sheet says "basic", the honest edit is to move Power BI earlier in the line, not to drop the level.
Checking the output
Anthropic's guide describes one more technique that suits resumes well: after a draft, have Claude find a supporting quote for each claim, and "if it can't find a quote, it must retract the claim". Run it on your assembled text, not on the change list:
<resume>
[paste the final resume text you assembled]
</resume>
Go through the resume line by line. For each line, quote the fact from <facts>
that supports it. If any part of a line has no supporting fact, mark it
UNSUPPORTED and say which words. Do not suggest fixes.
Then check it yourself, because Claude is still reviewing its own work:
- Compare the files. Save your untailored resume and the tailored one as Word files, then use Review, Compare, Compare Documents, as Microsoft's support page describes. Every change is marked in one view, including any you forgot you accepted.
- Read every number. Each number on the page should appear in the fact sheet exactly as written.
- Read the skeleton. Same employers, same titles, same dates, one line per role.
- Run the interview test. For each bullet and skill, ask whether you could talk about it for two minutes with someone who knows the subject.
Our ChatGPT guide covers a parse test for the exported file, and our step by step guide to tailoring a resume works a full example by hand.
Cover letters with the same loop
The same fact sheet and the same citation rule make a reliable cover letter. Ask for a draft in which every sentence that states something about you carries its fact ID, so you can see where each claim came from:
Write a cover letter for this posting, under 250 words, in three short
paragraphs: why this role, two pieces of evidence, a plain close.
Every sentence that says something about me must end with the fact ID it
uses, in brackets, like [E1.2]. Sentences about the company must quote or
paraphrase <posting> only; do not use outside knowledge about the company.
No new numbers.
Check the brackets, then delete them. The line about outside knowledge matters: Anthropic lists "external knowledge restriction" among its techniques, and a letter that praises a product launch Claude half remembers is an easy way to look careless. For what a letter written for one posting should contain, see our post on AI cover letters written per posting.
Where a chat model stops
Even done well, this loop covers one part of an application. In this workflow, Claude does not watch employer boards for new postings, check that a posting is still open, or know whether you already applied. You assemble the file you upload, so the layout, export and parse test are on you. And the employer's form is still yours to fill: the work authorization questions, the screening questions and the upload fields. Each posting costs you a new chat, a few rounds of approval and a careful read, and that time adds up across a search.
ApplyEze takes on more of that chain. It watches the career boards of more than 100,000 employers directly, and it has duplicate detection and a posting lock built to prevent applying twice. For the documents, it writes a resume for each posting and a cover letter when the form asks for one, as PDF and DOCX, and reads each file back with a parser check before upload. It brings the posting's named skills and tools onto your skills line, even when your own record does not list them. There is no step to remove one before a live run, so use a dry run to see what it writes, and be ready to discuss anything on the page. A certification the posting asks for is not added unless you hold it. Your employers, job titles, dates, schools, degrees and certifications come from your own record, and code rejects a rewrite that adds a figure, employer, school or credential your record does not hold. Our comparison of the best AI for resume tailoring covers how other tools draw these lines.