Changelog

New

• An ATS resume checker. A new /resume page: drop in a resume as PDF, Markdown or plain text and get the breakdown a screening pass would produce. The file is parsed in the browser and never uploaded — only the extracted text and a handful of measured layout facts reach the server. Deterministic checks run first — contacts, dates in your experience, pictures standing in for text — then the model reads the text. The result is an index from 0 to 100 with a verdict ("Strong resume", "Good, with room to improve", "Weak resume", "Will not clear screening") and four scored dimensions: "Target Job Fit", "Experience and Results", "Format, Language and ATS" and "Personal and Professional Data". Each dimension shows the band the model rated it at, and the score is derived from that band — so you can see exactly where the points went. Every finding is anchored to a line of the resume: a severity tick sits in the margin, and the card with the explanation and a ready rewrite opens right under that line (on a phone it slides up as a bottom sheet). Alongside are the measured facts about the document, each against its norm: for a PDF that is pages, columns, headings, length, contacts and images; for Markdown or pasted text, just the length. An open critical finding stops a resume from being called strong, whatever the other numbers say.
• A resume against a specific vacancy. Name the target role — as a title, or just a link to the posting (hh.ru, Habr Career, a Telegram channel post): the title and description are pulled in for you, and if the page will not give them up you can paste the description into the field beside it. That switches on the "Target Job Fit" dimension, and the model weighs your experience against that vacancy. Without a target, a resume is scored on the other three dimensions rather than handed a hundred for something you never named.
• Scan history and notifications. A breakdown is kept for 90 days: reopen it from your scan history, recompute it if we read your field or seniority wrong, and delete it at any time. When a breakdown is ready, a message arrives in Telegram or by email — each channel can be switched on or off in the new Notifications panel in your account, and every email carries a one-click unsubscribe link. Allowance: one scan free, for good; then 5, 10 and 20 per 30 days on Basic, Pro and Premium.
• My Progress — one table for everything you've practised (beta). A new /progress page joins tests, AI interviews and the docs by theme for the stack you pick. Each theme gets a bar for tests and one for AI interviews, a weak spot by concept and how much of it you have closed; the weakest theme — with its readiness ring — and the stack totals sit on top. Open a theme and you get its concept breakdown and three doors: read the docs, drill the questions, or take an AI interview on exactly this theme — the builder opens with the themes and the stack already selected. Beside the tabs runs a reel of your recent interviews with the full history by month. Test-mode answers are now recorded per question too — before, only AI interviews reached your stats. Readiness counts coverage as well as quality: one answer graded 7 out of 10 doesn't make a theme "70% ready" while the other questions are untouched. Signed out, the page shows the ladder of themes at zero and offers to sign you in.
• Skills for your LLM agent. A new /skills page with ready-made instruction sets for Claude Code and Codex: C++ code review, an undefined-behaviour hunt, commit messages (PR descriptions — coming soon). A skill downloads as a zip that matches its source byte for byte, installs in three steps in about forty seconds, and the agent picks it up on its own when the task matches. Each cell carries a download counter and an install card with the commands; the table filters by agent.
• A Tools menu in the navigation. The top bar gains a Tools menu — the resume checker and agent skills live there — and Interviews becomes a dropdown: start an interview or open your progress. On a phone the tools sit in their own section of the burger menu and in the Interviews flyout of the bottom bar, and Blog and Pricing move under More so they don't read as tools. To fit every item, the bar folds into the burger on screens narrower than 1240 px (it used to be 900 px).
• Reference cards for articles. A blog post's address standing alone on a line becomes a card: the title, a short description, the category, the reading time and — for a locked post — the plan it opens from. In Telegram Instant View the card becomes a link carrying the article's title. The first two are already in place: the two parts of the DirectX 12 series point at each other.
• A locked-post badge in the blog feed. Cards and the pinned post carry a pill in the colour of the plan that unlocks the post. The level resolves after the page loads, so a subscriber never sees a lock on a post they can read.

Improved

• The Russian tier names. Basic, Pro and Premium are now «Фиксик», «Крококодер» and «Искусственный интеллект» in the Russian version — on pricing, in the account, in the menu, on locked-post badges and in the simulator's upgrade prompt. The English names are unchanged. On the pricing card the name moves into a full-width band in the tier's colour, so a long name no longer pushes its price below its neighbours', and the recommended plan is marked by a filled band instead of a ribbon.
• The voiced interview sounds like speech, not spelling. Clips are synthesised at the chosen speed (0.75×–1.5×) instead of being sped up in the browser, so the voice stays natural at any rate. In Russian, Latin identifiers and acronyms are read the way engineers say them: API and JOIN as words, letter by letter only when no word can be made of it (CQRS, IOCP), and O(n log n) as "O of n log n". The question text on screen is unchanged.
• The landing page shows the real product. The question-graph section swaps its decorative constellation for a live fragment of the actual graph (Go → Channels → concepts → questions): the nodes are clickable, the panel beside them shows a question list or a question card, and an unanswered node stays grey. A Progress section sits between the two modes and the knowledge base, with a sample weakest theme and stack totals. Both calls to action sit under the blocks they refer to and share one style.
• The mobile header. The mark keeps the left corner, and the wordmark is now centred on the screen at any width from 320 px.

Fixed

• A mid-sentence spoiler in Instant View. A spoiler inside a sentence split its paragraph in Telegram, stranding the rest of the sentence as a separate block. Only a spoiler that owns its whole paragraph becomes an expandable block now; a spoiler inside a sentence shows as text in Instant View (on the site it stays hidden).
• The graph filter sheet keeps clear of the top bar. On iPhone the bottom sheet of the question-graph filters was drawn taller than the visible area and covered the logo and menu.
• The callout button in the blog editor no longer eats the selection. Clicking it with text selected deleted that text and left an empty callout; the selection is now wrapped into a callout, and inside an existing quote the click changes its type.