Draft a cover letter from a job posting and your own CV and evidence base — making only claims that trace to something you actually said, and stopping when the evidence stops rather than filling a word count.
Part of claude-skills.
A cover letter is a volatile document. A mediocre one is ignored; a wrong one disqualifies you before anything true about you is read. Bounded upside, unbounded downside — so this skill is built to reduce the cost of producing a true letter, not to produce a persuasive one.
- Tells you when not to bother. It reads the posting for stated hard requirements — a language, a permit, a years-of-management threshold — and says plainly when one fails, before you spend an hour on an application that screens out.
- Separates stated claims from derived ones. "Reduced findings from 180 to 12" is in your CV. "Sixteen years of contracting" is arithmetic on a date range and you never said it. Both can appear; the derived ones are surfaced for you to confirm.
- Refuses to diagnose the employer. If the posting says their cloud spend is outpacing revenue, that is theirs and quotable. If it does not, guessing is the failure that cannot be recovered — you would be explaining someone's business to them, wrongly, in paragraph one.
- Has no target length. The letter ends when the sourced claims end. Three claims in three paragraphs is a finished letter.
- Checks the letter against the CV you are actually sending. Two documents agreeing on specifics is expensive to fake, which is what makes it worth anything now that fluent tailored prose is free.
- Hands you the decisions it cannot make — why this employer, relocation, notice period — marked in the draft rather than invented around.
Works on any posting: a URL, a pasted advert, an email. Not only LinkedIn.
They come from measured outcomes rather than career-advice convention. The synthesis, with
citations and the grading of each source, is in references/evidence.md.
- Composition beat tailoring. Detail, clarity and structure predicted more interviews and a shorter search; tailoring predicted nothing (Wingate et al., 2025, n = 183, real outcomes).
- Mirroring the posting has stopped signalling. After an AI writing tool launched on a large labour platform, the correlation between textual alignment and callbacks fell 51% (Cui et al., 2025) — replicated independently on different data (Galdin & Silbert, 2025).
- A signal works only if it is costly to fake (Spence, 1973). An adjective is free. A checkable fact is not.
- Editing the draft correlated with hiring success, so a finished letter is the wrong output.
The statistics you will see quoted everywhere — "83% of recruiters read cover letters" — range from 83% to 26% on the same question and are all published by companies selling cover-letter tools. This skill does not cite them and neither should you.
Yes — and it is measurably better with the other two.
On its own, with just a posting and a CV, every rule still works: it triages hard requirements, separates stated claims from derived ones, refuses to invent, selects rather than pads, and checks the letter against the CV. The letter it produces will be true.
What it cannot do alone is reach for facts your CV does not contain — and a CV is built to leave things out.
Measured on one real application, the same role and the same person, with and without an evidence base:
| opening line | |
|---|---|
| CV only | "I'm applying for the Engineering Manager role in Europe." |
| CV + evidence base | "Between 2004 and 2008 I built and ran the infrastructure team at a mid-size ISP — five direct reports, three of whom I hired myself." |
The second fact was not on the CV. It was in the evidence base, because a CV compresses
roles beyond roughly fifteen years and this one had dropped it. For an engineering-management
application whose CV showed only "mentored two junior engineers", that fact was the whole case —
and without cv-evidence-base there was nothing better to open with than the fact of applying.
The evidence base also carries something a CV never can: your own conclusions about what you cannot claim. In another run it declined an entire application because the evidence base held a recorded finding — "Kubernetes is operate-around, not own-the-platform" — against a posting whose must-have was deep Kubernetes ownership. Nothing in the CV would have stopped that letter.
So: no hard dependency, and no pretending. Used alone it says when the letter came out thin because the source was thin, and points upstream rather than padding. The family in order:
cv-evidence-base— recovers what is true, including what the CV dropscv-and-human— tailors the CV itselfcv-cover-letter— this one, drawing on both
- Give it the posting, your CV, and your
evidence-base.mdif you have one. It reads what you supply and does not go hunting for files. - Let it tell you not to apply. That is the feature, not a failure to be talked out of.
- Check the derived claims. They are usually right; they are still things you did not say.
- Expect it to be shorter than you think it should be. The urge to fill a page is the convention it is deliberately ignoring.
- Answer the open decisions rather than deleting them. They are the parts only you know, and the evidence says the editing is where the value is.
- Run
cv-evidence-basefirst if you have not. A thin evidence base makes a thin letter, and the fix is upstream of here.
- It does not invent a metric, a date, a duration or a scope. Ever, under any pressure.
- It does not diagnose the employer. No problem is attributed to them that they did not state themselves.
- It does not pad to a length, and it has no target length.
- It does not write "passionate about" or substitute claimed enthusiasm for evidence.
- It does not cite recruiter-survey statistics.
- It does not write or edit your CV — that is
cv-and-human. - It does not decide whether you should apply. It reports a failed hard requirement and leaves the choice with you.
- It does not promise an outcome. Nothing here predicts a callback, and any tool that tells you otherwise is selling something.
None to install. Three inputs, all supplied by you: the posting, the CV you are sending, and
optionally evidence-base.md from cv-evidence-base. Without the evidence base the CV alone is
the source of truth and the claims table will be thinner.
Uses clear-and-human for register if it is available, and web search, if
available, to read a posting from a URL — otherwise paste the text.
MIT, like the rest of this repo. Structure adapted from
Paramchoudhary/ResumeSkills (MIT); see the
Provenance note in SKILL.md for what was deliberately not carried over, and why.