vdk888

Cron-fire post-flight scaffolding. Audit-log INSERT, git commit + push, notification send. Replaces inline wrap-up boilerplate across scheduled tasks.

Build or refresh a Research Capsule for a sales prospect. Pulls background, workflow pain points, recent signals, and three outreach angles (V1/V2/V3). Writes a structured Markdown file to your local vault. Read-only on social sources — never sends or posts anything.

Switch a Claude Code agent (Mac CLI, VPS systemd, or launchd) from Anthropic to DeepSeek as the LLM backend. Use when you want to run Claude Code with DeepSeek while keeping the full harness (tools, files, skills, channels). Covers the reasoning-effort proxy, model routing, channels fix, and rollback.

Source-tier classification + independent-verification rubric for any external signal (news, Substack, paywalled wires, sentiment trackers). Load whenever you're about to act on a third-party claim or write "<outlet> reports X" in a memo.

Canonical 7-part investment thesis structure. Use whenever you write a position note, proposal memo, decision-reasoning row, idea-scout idea, theme note, or chat brief naming a specific bet. Every new thesis MUST follow this skeleton.

Produce 1-5 min landscape (1920x1080) video essays from an article, using a fully local Remotion project — React scene components, local TTS voice cloning, Pexels b-roll, kinetic typography, burned-in subtitles, chapter transitions, ambient music. Use when asked to turn an article into a narrated explainer video.

Publish a short-form Substack Note (300-600 words) with an optional hero image, using the Claude-in-Chrome extension (Playwright gets 403 on Substack). Use when you have approved note content ready to post.

Publish a long-form Substack Post (1500-2500 words) with inline images, hero image, and a subscribe button, via a logged-in Playwright browser profile. Supports scheduling. Use when you have an approved essay ready to publish.

Help a non-technical CGP / financial advisor understand, set up, and use Bubble Shield — protecting a client folder, the before/after, the masquer/conserver settings, the optional accuracy pack (better DETECTION, not everywhere-masking). Triggers when the user asks how it works, how to set up / configure it, which folders are protected, how to anonymise a dossier, see the before/after, the masquer/conserver table, protect data everywhere or catch PII in emails, turn on the smart detection — or seems unsure how to operate it, even without naming it. Also 'démarrer', 'onboarding', 'montre-moi', 'première fois' launch the guided first-run demo. Plain language, no jargon. CRITICAL HONESTY: do NOT claim it anonymises everywhere or that e-mail is auto-protected in Cowork (PostToolUse does not fire on built-in Read/connectors). Reliable protection = the marked FOLDER + bubble_shield_read; for e-mail, save the message into the folder first.

Triage a Gmail inbox 3×/day (or on a morning scheduled task) WITHOUT the model ever seeing real client PII — reading each mail anonymised through bubble_shield_mail_read, classifying it into a 5-tier taxonomy (Clients / Important / Newsletters / Structurés / CV / Transition), applying Gmail labels + archiving via bubble_shield_mail_apply, and preparing reply/transfer DRAFTS in the user's voice (never sending). Triggers when the user asks to sort/triage/classer their emails, 'traite mes mails', 'range ma boîte', 'prépare mes brouillons', set up a morning mail routine, or an email-triage scheduled task. GOLDEN RULES: doubt → ARCHIVE, never delete; NEVER auto-send (drafts only, human sends). Runs entirely host-side: raw mail never enters the model context.

Manage your personal boycott list. Add brands to avoid, track reasons, sync with Chrome extension. Use when the user complains about a brand, says "never buying from X again", wants to block a company, or mentions the boycott filter.

Generate speech from text locally using VoxCPM2 (Apache-2.0, 30 languages, voice design, voice cloning). Use when the user asks to "say" or "speak" something, wants a voiceover or narration, wants to clone a voice, or wants to generate audio from text. Free, runs locally, no API calls. Output is a 48 kHz wav file.