yar

Drive Google Chrome to do a real task on the web — navigate, fill forms, click through flows, extract content, check network/console — via the official Chrome DevTools MCP with named persistent login profiles: isolated-by-default `chrome`, plus persistent per-identity profile servers (chrome-personal, chrome-work, one per company account) whose signed-in sessions persist across runs, so each task runs under the right identity. Acts on DOM/accessibility uids — far more reliable and faster than screenshot+coordinate extensions. Prefer THIS over the Claude-in-Chrome extension for driving websites. Use when the user says "go into Chrome and do X", «برو تو کروم فلان کارو بکن», "fill this form on the site", "log into the portal and ...", "scrape/read this page", "automate the browser", «با پروفایل کاریم», «با پروفایل شخصیم», or any task that means operating a website on the user's behalf. NOT for local PDFs (use md-to-pdf / pdf) and NEVER for typing passwords, 2FA, or card numbers — the user always does those.

Project-agnostic engine for capturing a person's end-of-day retro log — what got done, decisions made, energy/focus (1-10), concerns, and tomorrow's single top priority — one dated Markdown file per day, structured for later pattern analysis (energy trends, recurring concerns). Conversational, not a form; adapts to a tired user with a 3-line short log. Optionally opens with a read-only evening dashboard (tasks completed today, overdue deadlines, tomorrow's calendar) when the caller wires in sources. The caller supplies the context layer: log directory, dashboard sources, decision store, concern routing, and conversation language. Personal retro only — end-of-day *repo/plugin* sweeping is `yar:repo-sweep`, not this. Triggers: "log my day", "daily log", "end-of-day log", "evening retro", "journal my day", "how was my day wrap-up".

FREE, multi-source web research on any topic via the Google **Gemini CLI** (free tier / Google login — no billed key), producing a cited Markdown report. Gemini-powered — distinct from Anthropic's built-in `/deep-research` (which fans out Claude's own WebSearch). A paid Gemini API skill, if you have one, is the sharper and deeper counterpart. Use whenever the user says "go research X", "deep research on …", "free gemini research", Persian "تحقیق رایگان با جمینای", or asks for a researched answer/briefing on a market, vendor, regulation, technology, best practice, country/region facts, pricing, etc. Saves the report as a .md file (by default under reports/). Output is research data — for domain-specific interpretation (legal, medical, financial) treat it as raw facts to be verified by a qualified expert, not authoritative advice.

A git branch-and-merge workflow for small teams coding with Claude Code — short-lived feature branches off main, conventional commits, a pull request the author squash-merges (no required approval), and one git worktree per parallel session so concurrent sessions never collide. Runs only when you invoke it; it does not auto-branch, auto-push, or auto-merge on its own. Use when you ask to start new work, create a branch, commit changes, sync or rebase on main, open or merge a pull request, resolve a merge conflict, run parallel Claude sessions, or clean up branches and worktrees. Triggers include "create a branch", "start a new task", "open a PR", "merge this", "ship/land this", "sync with main", "resolve conflict", "set up a worktree". "Ship" is a single command for the whole chain — commit if needed, push, open the PR, and squash-merge it. It is NOT finished at `gh pr create`: shipping means the PR is merged and main has the change (the plugin's ship-guard hooks enforce this).

Install the plugin's git-level pre-commit guard (blocks committing binaries/secrets) into the current repository. Run once per repo or clone. Invoked manually as /yar:install-guards; it does not auto-run.

Merge yar's recommended permission policy (an allow-list plus a destructive-command deny-list) into the current repository's .claude/settings.local.json — the personal, git-ignored settings file, so the policy never gets committed or pushed. Run once per repo. Invoked manually as /yar:install-perms; it does not auto-run.

Install yar's Persian/RTL chat-rendering rule into the machine's global ~/.claude/CLAUDE.md (idempotent managed marker block) so every project renders mixed Persian-English replies correctly — the whole reply as one self-contained RTL widget card (an always-on base style plus pay-per-use component snippets) when a widget tool exists, English-only plain chat text, atomic LTR isolation for paths/URLs, and a BiDi-safe fallback for plain CLI. Run once per machine; re-run to update. Invoked manually as /yar:install-rtl; it does not auto-run.

Turn a meeting recording or transcript into a clean Markdown summary — narrative, decisions, and action items (owner, deadline, priority) — then optionally route the outputs into the project's own tools: a task per action item, calendar events, formal decisions to a decisions record, per-person/project notes, status flags, and a follow-up email. Transcribes audio/video locally with ElevenLabs Scribe v2 (speaker diarization, any language) and keeps a verbatim transcript in the spoken language. Tracks a per-meeting processing ledger in the summary frontmatter — every stage from transcript through follow-up email — so "review the meetings" shows what is done vs. still pending per meeting. Deletes the source recording after the transcript and summary are saved (MEETING_DELETE_SOURCE). Use when asked to "process this meeting", "summarize this recording", "extract action items", "review the meetings", or after the meeting-recorder skill produces a file. Transcribing needs ELEVENLABS_API_KEY or GEMINI_API_KEY.

Record a meeting (Google Meet, Zoom, or any call) as a small audio-only .m4a on macOS using a self-built CoreAudio process-tap recorder — no third-party app. Captures both the remote participants' audio (system output) and your own microphone, regardless of output device (AirPods or speakers), names the file by date+topic, and hands off to the meeting-processor skill for transcription + summary. Use when asked to "record this meeting", "start/stop recording", "record the call", or to capture a conversation for later transcription. macOS 14.4+ only.

Project-agnostic engine for organizing, sorting, deduplicating, and renaming files in any folder or repo, and for keeping each destination folder's README in sync. Use whenever you sort or clean up files, classify new uploads or downloads, notice possible duplicates, or add/move/create a file into a folder. Enforces meaningful naming (`YYYY-MM-DD-descriptive-name.ext`), detects exact and near-duplicate files (name, checksum, size, text, image resolution), picks the best version (preferring signed, then final, newer, higher-quality, more-complete, else asks), deletes safely (trash, never `rm`; confirm or ledger), and updates the destination folder's README with a one-line content summary on every add. Reads file contents via the `yar:file-inspector` agent; the caller supplies any project-specific destination map, storage split, or off-limits folders. Triggers include "organize", "sort files", "clean up", "tidy", "where does this file go", "duplicate", or "organize inbox".

End-of-day sweep across every local git repo, worktree, and installed Claude Code plugin at once. Reconciles each repo with its remote (pushes committed work, fast-forward pulls, surfaces divergence and dirty trees, never discarding them), prunes stale state — local branches whose content already landed on the default branch (squash-merge aware, merged pull requests as proof), merged or contained worktrees, orphaned worktree directories from repo renames, primaries parked on a merged branch, zero-unique remote branches — and updates every installed plugin to its latest published version, flagging manifest/install version drift. Read-only scan, then a classified plan (safe / needs-you / skipped); mutations only on confirmation. Use on "sweep the repos", "wrap up the day", "end of day", "good night", "clean up branches and worktrees", "any stale branches?", "update all plugins", or Persian "شب بخیر", "جمع‌وجور کن", "برنچ‌ها و ورک‌تری‌های الکی رو پاک کن", "پلاگین‌ها رو آپدیت کن".

Author, edit, and improve Claude Agent Skills following the current Agent Skills standard (SKILL.md frontmatter limits, progressive disclosure, description-based triggering). Bundles the full draft, test, evaluate, iterate, validate, and package loop with a ready template and Python scripts. Use whenever someone wants to create a new skill, says "turn this into a skill", "make/create/build a skill", or asks to edit, fix, improve, validate, package, or optimize the triggering/description of an existing skill — even when they describe a repeatable workflow worth capturing without saying the word skill.

Verify and source the factual, market, or statistical claims in a draft (email, post, memo, proposal) so each one carries a hyperlink to the ORIGINAL independent source — not a vendor's own marketing page — backed by a concrete number. Searches the web for the real URL (never guesses or fabricates links), prefers primary reports / independent press over self-interested sources, and flags any claim it cannot back so the author can cut it or mark it an estimate. Use when asked to "source this", "back these claims with links", "add citations", "fact-check and link the claims", "هر ادعا رو با منبع و لینک پشتیبانی کن", "منبعِ اصلی برای این ادعاها", or before sending a claim-heavy email / publishing a data-backed post.

Design and typeset documents from Markdown in the host project's own house style, with PDF as the output (formerly md-to-pdf). A per-project theme (identity, logo, one accent, type pairing, mood, weights, locale) gives every project a consistent identity; hierarchy comes from font weight and whitespace, never bold-and-bigger. Reports get a cover, letters and memos a letterhead, decks 16:9 slides. Tables are set in their own register (smaller type, right-aligned tabular figures, rules only under the head and above totals, bold label rows read as totals) and financial statements (sections, groups, indented items, double-ruled final line) render from Python rows or a QuickBooks export. RTL/Persian is first-class (mirrored layout, x-height-matched pairs, Persian digits, Jalali dates). Paper is Letter; A4 on request. Fonts open-licence, pinned, cached. Use when asked to "make a PDF", "typeset this", "design a proper document", "build a deck", "lay out these statements", "پی‌دی‌اف بساز". Needs Chrome/Chromium.