Imported from slavshik/cv (AGENTS.md). Install upstream with npx skills add slavshik/cv. Copyright stays with the author.

AGENTS.md

CLAUDE.md is a symlink to this file — one set of instructions, two names.

The repo

My CV at slavshik.me/cv. Vite, TypeScript, published to GitHub Pages from a GitHub Actions workflow. The repository contains only source: dist/ and the PDF are never committed.

• content/resume.json — the CV itself, in the JSON Resume v1.0.0 schema. The single source of content. content/TODO.md says what is still missing from it and why.
• src/resume.ts — the schema as types, plus parseResume, a narrow check that fails the build loudly rather than printing undefined onto a CV.
• src/render.ts — resume.json → HTML. A pure function: no DOM, no clock, no I/O. It runs inside vite.config.ts at build time.
• src/main.ts — the only script the CV loads: the time-of-day accent and the theme button. Nothing else.
• content/jobs/<date>.json, src/jobs.ts, src/jobs.css, jobs/index.html — the second page, at slavshik.me/cv/jobs. One file per sweep, written by jobsweep publish; the renderer is pure the same way render.ts is. Read docs/adr/0006 before touching it, and note what the page must never carry.
• content/resume.json also carries projects — the showcase. It renders twice, on screen both times: as a gallery of thumbnail cards in the CV's footer, and in full at slavshik.me/cv/work, from src/work.ts, src/work.css and work/index.html. Not one project reaches the PDF — print hides the rows and keeps a single line pointing at the work page. One array, two surfaces; read docs/adr/0007 before touching either, and note that this page is indexed where the jobs page is deliberately not.
• src/styles.css — the whole design of the CV, screen and print. src/tokens.css is the palette, the type stacks and the reset, shared with all three pages; none of them may fork it.
• index.html — markup and metadata, and nothing else. <!--resume--> and <!--jsonld--> are where the build injects.
• scripts/pdf.mjs — renders the built page to dist/, under the name the page itself asks for in the download attribute of its own link. That name is built by pdfFileName from basics.name and meta.pdfRole.
• docs/adr/ — why this repo looks the way it does. Read before changing the build, the deployment or where content lives.

Working here

• npm ci, then make dev → http://localhost:5173/cv/. make help lists everything else. Editing content/resume.json reloads the page.
• English. README, code comments and commit messages are all English here, unlike the neighbouring slavshik.github.io, which is Russian. That is on purpose — do not "restore consistency".
• Content changes are data changes. Anything that appears on the page comes from content/resume.json. If something cannot be expressed there, the fix is a field, not a hardcoded string in the renderer.
• Never invent CV content. The data came from a LinkedIn export that is lossy in places. Where a sentence was cut off, the fragment was dropped and logged in content/TODO.md. Leave gaps as gaps.
• The phone number is not in this repository. It goes into the PDF at print time from CV_PHONE — a repository secret in CI, .env.local here. It must never reach content/resume.json or the served HTML; there is an e2e test that says so.
• The jobs page is public, and is written as if it were. The repository is public, so /cv/jobs/ and every content/jobs/*.json are readable by anybody who has the URL — noindex and robots.txt are obscurity, not privacy. Rows are titles, companies, locations and links to public postings. Fetched description bodies are somebody else's text and never leave the gitignored runs/. The CV's rules do not apply there: it has no print stylesheet, no screenshot baseline and no weight budget.
• The sweep is run by hand. make jobs sweeps, publishes the day's file and commits it; you push when you want it on the page, and CI builds the site as it always did. There is no schedule. Scheduling it under launchd was tried and abandoned — macOS refuses a LaunchAgent access to ~/Documents, and the grant is per binary rather than per process tree, so allowing /bin/sh still left go refused. docs/adr/0006 has the measurements.
• Two faces, one per surface. The screen is set in a humanist sans and the PDF in Charter; --body is the token, and the print block in src/styles.css is the only place it is overridden. Both stacks end in faces the pinned Playwright image already ships — Bitstream Charter, Liberation Sans, Liberation Mono — so the build installs no fonts. Palatino and the fonts-texgyre apt step were removed on 2026-09-15; do not put either back without checking fc-list in the container first.
• The PDF has four things the screen has and it does not: the reading marks, the portrait, the showcase, and — the other way round — the phone number, which exists only there. All four are one rule each in @media print and all four are load bearing. The showcase is the newest: twelve titles and their studios is a directory, not a document, and it cost a sheet and a half. The last page of the PDF is settled — if a change to the showcase moves it, the change is wrong. .showcase keeps exactly section's margin in print for that reason, and make pdf-ci is how you check.
• The first page of the PDF is a budget. Header, summary and Skills have to fit on it, with the history started under them. That is what the section order in renderResume and the .facts rules in the print stylesheet are for — both look like fussiness until you render the PDF and count. make pdf-ci.
• To test the download button, render the PDF first. It is a build artefact, so make dev alone has nothing to serve: make pdf once and the dev server hands over the real file, or make preview for exactly what goes to Pages. Asking for it before either answers 404 in words rather than 200 with the page, which is what it used to do — and an HTML file saved under a .pdf name looks like a corrupt PDF, not like a mistake.
• make check before calling anything done (types, lint, format), plus make unit. Run make test when a change could move a pixel or the byte count.
• Screenshot baselines are exact. Tolerance is zero pixels and they only run inside the pinned Playwright container (make e2e) — macOS and CI Linux render type differently. A diff is a real change; look at it before reaching for make e2e-update.
• Prettier does not touch HTML (.prettierignore). The markup is aligned by hand and its comments sit next to what they explain.
• The marks are a small syntax, not markdown. **action**, __architecture__ and ==scale== in the prose of content/resume.json become <strong>, <span class="term"> and <mark>. Only prose is parsed — summaries, highlights, project descriptions — and escaping happens first. They are stripped in print on purpose: the PDF must stay a plain document. Keep roughly a sixth of an entry marked; past that they stop meaning anything. See docs/adr/0004.
• Writing or revising the prose is its own job. What may be claimed, how a line is phrased, what to cut, and which prose the renderer throws away are in the cv-prose skill (.claude/skills/cv-prose/SKILL.md). Read it before touching content/resume.json, and before acting on anybody's suggestions about the wording.
• Keep the page free of JavaScript. The CV must be entirely readable with scripting off — an e2e test enforces it. src/main.ts may grow only for things that are genuinely decoration, and for the one thing below.
• The visit counter is the exception, and it is mine. src/hit.ts sends a single same-origin request to /api/hit — the Worker that already serves slavshik.me, on the same zone, so there is no third party, no cookie and no identity that survives midnight. It is silent anywhere but the live host, on ?aqa=1, and for a visitor sending Do-Not-Track or Global Privacy Control, which is why neither the screenshots nor the PDF ever count as a reader. /cv/jobs/ is deliberately not counted: that page carries no module script at all and is not for anybody but me. The rows are read with make stats in the slavshik.github.io repository, where the Worker and its D1 live.
• make size holds the whole page under 12 kB gzip. It is a text document; there is no reason for it to grow.

Print is not a second document

@media print in src/styles.css is what scripts/pdf.mjs renders, so the downloadable PDF and Cmd+P agree by construction. Three things there are load bearing and look wrong out of context:

• a dated row is a block with the date floated right, on both surfaces. It reads as a layout choice and it is also what makes the entry paginate: Chromium will not fragment a grid container across pages, so while these rows were grids a long entry was pushed whole to the next sheet and left a quarter of a page white. Print used to undo the grid by hand for exactly this reason; it does not have to any more, and the rule that did it is gone. Turn these rows back into grids and the white pages come back.
• only h3 carries break-after: avoid. Chaining it onto the company and date lines as well makes Chromium treat the run as one unbreakable lump, with the same result.
• .facts rows stay grids, in print as on screen. They are one line each and carry break-inside: avoid, so there is nothing for them to fragment.