Razbakov Skills

Use when reviewing a website's UI copy, content, and meta for brand voice consistency. Triggers include launching a new site, rebranding, noticing inconsistent tone across pages, or preparing content for SEO. Checks nav, CTAs, author bios, meta descriptions, blog posts, project descriptions, config files, and locale files against brand voice guidelines.

Update a YouTube live stream's title, description, and tags in real-time based on session agenda from Notion. Use when the user is streaming and wants to change the stream title, update the live stream description, or make the stream more discoverable. Triggers: 'update stream title', 'change live stream title', 'update my stream', 'make stream title better', or when user mentions they're streaming and wants metadata updated. This is the fast path — no transcription needed since the stream is live and the agenda is already known.

Onboard a new Claude Code user — set up their ~/.claude/CLAUDE.md, install relevant skills, and start a structured learning program covering best practices. Use when setting up Claude Code for the first time, onboarding someone to Claude Code, or continuing a learning session.

Recover lost registrations after a signup/lead form has been silently broken — diagnose via the Resend email log, classify real users vs tests, re-submit missing contacts to HubSpot via the public Forms API (no auth token needed), and produce a personal follow-up list. Use when the user says "backfill dropped signups", "the registration form was broken", "we lost signups", "check what registrations came through", "recover lost leads", or any variation where a form-to-CRM pipeline has been failing and needs reconciliation. Also use proactively after fixing any registration form bug to check whether real users were lost before the fix.

Turn a markdown operational document (runbook, checklist, SOP, playbook, host script) into a styled print-ready PDF and upload it to a specific Google Drive folder, idempotently — so you can edit the source and regenerate without breaking the Drive link. Use when a user wants a markdown doc rendered as a PDF in Drive, when they ask to "make this printable", "put this in the team folder", "update the PDF in Drive", or whenever an ops doc needs a physical/printable artifact with a stable shareable link. Composes the `latex-pdf` skill (LaTeX template) and the `google-drive` skill (gog CLI) — use this one instead of calling those directly, because it handles the storage convention, the Unicode gotchas, and the re-upload-without-breaking-the-link pattern that you will get wrong by default.

Seed a new or empty Instagram account with a 9-post grid (3×3) so the profile looks established the moment a new visitor lands. Designed for festivals, new businesses, product launches, conferences, communities — any time an empty IG profile would hurt conversion from external traffic (QR scans, flyer drops, cross-promo). Generates assets via /image-from-gemini (per content-publishing rules — never HTML), writes captions with hashtag sets, and outputs a posting order + cadence plan. Trigger generously: phrases like '9 posts for instagram', 'fill my IG', 'starter grid', 'launch grid', 'instagram seed', '9-post grid', 'IG account not to look empty', 'first instagram posts', 'feed bootstrap', '3x3 grid', 'instagram launch content'. Even if the user mentions only one piece (just the images, just the captions, just the order), use this skill — the grid only works as an integrated bundle.

Runs dependency vulnerability scans and produces a complete report with exact installed versions, reason for risk, and remediation priority for each finding. Use when the user asks for dependency scan, npm audit, bun audit, package vulnerabilities, CVE review, or security status of dependencies.

End-to-end video/podcast post-production pipeline: split recordings into topic-based clips, remove dead time, generate thumbnails (landscape 16:9 and vertical 9:16), create YouTube metadata (titles, descriptions, tags), and cut Shorts. Use when the user has a video or audio recording (meeting, podcast, interview, webinar, stream) and wants to process it for publishing on YouTube, social media, or any platform. Triggers: 'cut this recording', 'make YouTube clips', 'create thumbnails', 'process this meeting recording', 'split into episodes', 'make shorts from this video', 'publish this recording'.

Turn a product's landing/marketing pages into a prioritized, gap-analyzed backlog. Use this whenever the user wants to audit what a site promises vs. what is actually built, convert marketing/landing copy into user stories, map work to Critical User Journeys (CUJs) and Jobs-To-Be-Done (JTBDs), score a backlog with WSJF, or populate an issue tracker (Linear, Jira, GitHub Issues) plus a tracking spreadsheet from a product's promises — even if they don't say the word "backlog" or "skill". Reach for it on requests like "audit our landing pages", "promise vs reality / what do we promise that we haven't built", "turn the marketing into a backlog", "make user stories from the site", "build a CUJ/JTBD map", "prioritize this with WSJF", or "get our promises into Linear/Jira". Also use it to maintain such a backlog after it exists.

Verify a Netlify-hosted form is registered and accepting submissions BEFORE distributing any offline asset (flyer, QR code, business card, print ad, physical signage) that points at the form. Catches the silent failure mode where Netlify only indexes forms seen during a build with form-detection enabled — submissions made before the form is registered hit the static 404 and are irrecoverable. Use this skill whenever the user is about to print/distribute QR codes, flyers, business cards, or any offline material that drives leads to a Netlify-hosted form; whenever a form "works in prod but no submissions are showing up"; whenever the user says "I enabled forms but can't see anything"; or proactively after merging any PR that ships a new Netlify-forms-backed lead capture page. Also covers verifying email notifications are wired so leads are not silently sitting in a dashboard nobody checks.

Process a meeting or call transcript into structured project artifacts — session notes, contact profiles, action items routed to the right destination (markdown / Google Calendar / Google Doc / GitHub issue), and doc updates. Fetches from Jamie (meetjamie.ai) automatically when the user references a recent call they recorded. Use when the user pastes a call transcript, says "process the last Jamie call", "process my call", "extract from meeting", or when a transcript with speaker labels appears in the conversation.

Download a YouTube video, transcribe it, generate optimized title/description/chapters/tags, create a thumbnail, and update the video via YouTube Data API. Use when the user shares a YouTube URL and wants title, description, chapters, tags, or thumbnail updated. Triggers: 'update youtube', 'need title and description', 'add chapters', 'fix youtube metadata', 'generate thumbnail for youtube', or when user shares a youtube.com/watch URL and asks for metadata.

Recommend installed skills by mining Cursor and Codex transcript history and matching recurring needs to skill metadata. Supports both standard recommendations and "new/unused skill" discovery with evidence, trigger phrasing, and SKILL.md paths.

Analyze all projects in a portfolio to identify skill gaps, then recommend roles/teams to hire and produce job descriptions. Use when the user asks to figure out hiring needs, what roles to hire, team structure, or skill gaps across their projects.

Take a long stream-of-thought transcript (voice memo, video reflection, text dump) and turn it into a cluster of 5-20 publishable blog essays. Clusters ideas by theme, drafts each essay via parallel sub-agents, generates Gemini hero images, batch-commits, and (optionally) sends a grouped link list via the user's preferred messaging surface. Use when the user shares a transcript and says 'turn this into essays', 'write a series', 'split into posts', 'draft all of them', or pastes a long stream of ideas they want shipped.

Transcribe video or audio recordings locally using faster-whisper via Docker. Sets up the transcription service, extracts audio, and produces timestamped transcription (JSON + plain text). Use when the user asks to transcribe a video, audio file, meeting recording, podcast episode, or any media file. Triggers: 'transcribe this video', 'transcribe this audio', 'get transcript', 'transcribe recording', 'speech to text'.

Re-ground a session/meeting summary (and every artifact built on it) against the raw transcript, when the summary was written first and the transcript became available later. Use whenever a raw transcript becomes accessible after a summary-based artifact exists, whenever the user pushes back on a fact in a summary, whenever you notice a downstream contradiction, or whenever you're about to edit a contact/reflection/issue that was built on a pre-transcript summary. Do not skip this even if the summary "seems fine" — summaries written from memory or from an AI auto-summary routinely contain speaker mis-attributions, framing drift, missed threads, and inverted consents, and every downstream artifact silently inherits those errors until somebody pulls the tape. This applies to any transcription source (Jamie, Fireflies, Otter, Granola, Rev, manual recording) and any artifact system (markdown sessions, CRM contacts, coaching journals, issue trackers).

Use when a software project under ~/Projects should become a separate organization under ~/Orgs, or when governance files are mixed into a code repo and need to be split into a dedicated org workspace.