Open source · Claude skill · v0.1
Paste a job link. Dossier reads the posting, tailors your resume and cover letter from facts you have verified, scores them against the posting's keywords, and stops. Nothing ships until you approve every change. Then you get a package per job and a tracker that keeps itself current.
| Missing term | Where in posting | Decision |
|---|---|---|
| marketing automation | Requirements | added · Marketo admin at three companies |
| dashboards | Requirements | added · weekly COO pipeline readout |
| power bi | Requirements | left out · not true (Tableau only) |
| salesforce marketing cloud | Requirements | left out · not true |
| agentic | Preferred | added · agents ran on scoped API permissions |
Justified by profile lines 41 and 44. Scope, number, and title unchanged.
The rule: every claim in the output traces to a fact you stated. When a posting asks for something that is not in your file, Dossier asks you. It never infers, rounds up, or invents. That is what makes the result safe to send with your name on it.
It reads your current resume, then asks the five to ten questions a resume leaves out: which tools you owned versus touched, the number behind each bullet, projects, things you never want claimed, and your format rules. The answers become the fact bank every later stage draws from.
Paste links, or give a title and a city. It reads your target companies' own job feeds (Greenhouse, Lever, Ashby, SmartRecruiters, Amazon) rather than a search engine, so the list is complete and every link is the employer's own posting. Repeat runs show only what is new, ranked by how well your current resume already covers each role. You pick.
Each posting is saved as it read that day, then broken into must-haves, nice-to-haves, keywords, and the specific signals a cover letter can use. Must-haves with no supporting fact are flagged as gaps before anything is drafted.
Three to five people at the company who share one of your networks: same school and program first, then former colleagues, then communities. It points you at LinkedIn's alumni view in your own account, supplements with public bios, and drafts a note under 120 words in your voice. Public professional profiles only; you send the note.
Selection and translation, not invention. Bullets are reordered and rewritten in the posting's language where the underlying fact supports it. A scorer weights the posting's terms (requirements count triple) and reports coverage; the loop stops when every remaining gap is either untrue of you or irrelevant.
You see the score, every change against your master resume with the fact that justifies it, and the full text of both documents in the chat. One question: approve, or what should change?
Nothing final is produced before this step
Final Word and PDF files, the ATS report re-run on the shipping file, a tracker row with score, file names, and a follow-up date, and one zip per job.
I built this during my own search, after a layoff, while in the first year of an MBA. Over five weeks I tailored around thirty resumes by hand. The tailoring itself was never the hard part. The hard part was trust: every tool that offered to do it for me rewrote my experience into something I could not defend in an interview, and the reviews of those tools say the same thing.
So the product is the approval gate. The scorer, the Word builder, and the tracker are plumbing. What matters is that you see a before and after for every bullet, the specific fact that justifies each added keyword, and an honest list of the gaps that stayed gaps. A 71% with no padding beats a 95% you have to explain away.
Because the workflow needed to be proven before the interface did. A skill ships in a day, runs inside a tool my classmates already use, and costs nothing to try. The prompts and the scoring logic are the product; if people come back to it, the browser extension and the app are the next step, and the pieces port straight over.
A weighted keyword-coverage proxy. It extracts tools, skills, certifications, acronyms, and repeated phrases from the posting, counts requirements-section terms triple, knows the common synonyms (GTM and go-to-market, SFDC and Salesforce), and checks a Word file for the things that break parsers: tables, text boxes, multi-column layouts, contact info hidden in the header. It is not a prediction of an interview. It is a list of things to fix, ranked.
Dossier is a folder of instructions and five small Python scripts. Claude runs it; you do not need to.
Settings → Capabilities. (On a Team or Enterprise plan, an owner enables it under Organization settings.)Customize → Skills, click +, then Create skill, then Upload a skill, and choose the zip.Using Claude Code instead? Unzip and drop the dossier folder into ~/.claude/skills/. The README inside covers everything else, including how to run the scorer on your own against any posting.