Manual escalation lever above the default `prompt-refinement.md` rule. When the user invokes `/align N` they are explicitly summoning a clarification gate with exactly N numbered + lettered questions, each carrying a single committed recommendation. Single-shot per invocation - after Phase 3 approva

Global, geopolitical, and cross-domain intelligence brief for the CEO - world events, sovereign-tech shifts, macro and threat signals relevant to 31C, rendered as a dated HTML/PDF brief. Use for world-scale situational awareness, not a single entity. Trigger when the user says "world intel", "geopolitical brief", "what's happening globally", "global threats", or "CEO intelligence brief". Do NOT use for a specific company or person (use /osint) or an externally published newsletter (use /intel-briefing-newsletter).

Drain overdue CRM "red debt" into the Action Queue as routed, voice-drafted nudges for one-click CEO approval. Use when the CEO says "cold sweep", "drain the cold contacts", "sweep the overdue CRM", or wants the overdue-contact backlog turned into ready-to-send drafts. NOT for sending email (that is the human-approved executor) - this skill only drafts into the queue. NOT for a single follow-up (use /follow-up). NOT for pipeline review (use /crm). CEO-only (not synced to executives) during the prove-out.

Produce a branded 31C external letter (PDF + DOCX) using the locked corporate template. Use for formal correspondence to customers, partners, regulators, investors, counsel, and institutional contacts. Applies the locked letterhead, signature block, footer, and confidentiality notice automatically. Trigger when the user says "write a letter to", "letter of introduction", "letter of interest", "formal letter", "external letter", "send a letter to [named recipient]", or when the user requests formal written correspondence that is not a proposal, MOU, or official document.

Prepare investor data room documents, due diligence responses, and fundraising materials. Use when preparing for investor conversations, responding to due diligence questions, or building the formal data room. Trigger when the user says "data room", "due diligence", "DD response", "investor materials", or asks to prepare documents for investors.

Advanced one-shot deep research. Runs token-heavy web acquisition (Perplexity) and reasoning/verification (Kimi) headless in a script, then Claude audits the findings, writes a cited report, and optionally distills to Odin. Use for a deep, multi-source, fact-checked report on a PUBLIC-web topic. Trigger on "deep-research-advance", "advanced deep research", "deep research on [topic] with verification". Do NOT use for private/internal questions (the topic and corpus go to third-party clouds), quick lookups (use /osint or WebSearch), or Odin-brain-only recall (use /odin). BEFORE running, if the question is underspecified, ask 2-3 clarifying questions to narrow scope.

Force N contrarian critique points against a target position. Each point carries a severity tag (BLOCKER / HIGH / MEDIUM / LOW) and a committed paragraph attacking from a distinct angle. Target is either inline (`/devil 5: <claim>`) or walked back to the latest substantive decision/claim in conversation. Manual sycophancy-breaker. Cheap alternative to /council (external) and /scrutinize (multi-phase). Single-shot per invocation. Default N=5, valid range 1-10. Honesty floor: stops early rather than fabricate weak points to hit N.

Parse documents (PDF, DOCX, PPTX, XLSX) with spatial bounding boxes via LiteParse. Answer questions with visual citations showing exact source locations on page images. Generate 31C-branded HTML/PDF reports. Maintains a parse cache for efficiency. Use when: "parse this document", "docparse", "visual citations", "show me where it says", "document analysis with citations", "parse with bounding boxes", "extract from this PDF". NOT for: plain text extraction without spatial data (use datastore-extract.py), email analysis (use /email-intel), web scraping (use /playwright or /firecrawl).

Document-level STRUCTURAL editor for long deliverables. Reviews a draft's argument architecture -- section ordering, claim-to-evidence linkage, hierarchy, redundancy, buried lede, missing sections -- and proposes CUT / MERGE / MOVE / CONDENSE / SPLIT / ADD / PROMOTE / PRESERVE operations with word-savings estimates, BEFORE any sentence is touched. Content is sacrosanct: it reorganizes, it never argues with the ideas. Phase 2 hands all sentence-level work (rhythm, specificity, vocabulary) to the always-on humanization.md prose rule -- it never duplicates that. Use when the user says "editorial pass", "structural review", "review the structure of this", "tighten this document", "restructure this draft", or wants a long proposal / brief / report's argument arc checked. Do NOT use for: sentence-level prose / "make this sound human" (that is humanization.md), typo or grammar fixes (sanitize-text + humanization-check), fact-checking (/validate), artifact grading (/evaluate), short chat replies, or atomic /zk notes.

Draft a new outbound email in Misha's voice - Voss-calibrated, concise, with a clear ask. Use for composing a fresh email to a named recipient. Trigger when the user says "draft email to", "write email to", or "email [person] about". Do NOT use for replying to an existing email (use /email-respond) or peer CEO correspondence (use /ceo-to-ceo). NEVER auto-sends - produces a draft for review.

Daily email intelligence processor. Scans incoming + outgoing 31C Exchange emails ([email protected] via exchangelib/EWS), groups by conversation thread, categorizes for CRM actions, tasks, pipeline updates, knowledge capture, and relationship signals. Presents digest for approval before executing. Integrates with /prime for morning briefing. Trigger: '/email-intel', 'process emails', 'email digest', 'check my email'. Do NOT trigger for: sending email (use send-email.py), reading specific emails, email search queries, or Gmail operations.

Generate images using AI models via Replicate API. Supports Nano Banana 2 (Google DeepMind, default) and FLUX.2 max (Black Forest Labs). Use when the user asks to generate, create, or produce an actual image from a prompt. This skill creates real image files on disk - not just text descriptions. Trigger when the user says "generate an image", "create an image", "flux-image", or asks to turn a prompt into an actual picture. Works standalone or as a follow-up to the image-prompt skill which produces text prompts.

Search, add, edit, and manage contacts in Google Contacts via the People API. Use for any task involving looking up contact details, adding new contacts to Google, updating phone numbers or emails, or browsing the contact list. Trigger when the user says "google contacts", "look up contact", "add to contacts", "add to google contacts", "find their number", "update their email", "search contacts for", or any reference to managing contacts in Google. This is a STANDALONE connector -- completely separate from the workspace CRM system. Do NOT trigger for CRM operations (use /crm for that). Do NOT trigger for Exchange/Outlook contacts.

Produce an external-facing intelligence briefing newsletter - curated, publishable intel framed for an audience beyond the CEO. Use when the output is meant to be shared/published, not a private desk brief. Trigger when the user says "newsletter", "intel briefing", "publish intelligence brief", or "external intel brief". Do NOT use for an internal-only situational brief (use /ceo-intel).

Produce full investor pitch deck content with narrative arc, slide-by-slide structure, and speaker notes. Supports strategic, venture, sovereign fund, angel, and corporate VC audiences. Use when preparing for investor meetings, fundraising presentations, or pitch deck creation. Trigger when the user says "investor pitch", "pitch deck", "fundraising deck", "investor presentation", or asks to prepare materials for an investor meeting.

Produce a presentation deck for an event, VIP meeting, or government engagement. Creates slide-by-slide structure with titles, key messages, content, and speaker notes. Supports MWC, regional expos, Black Hat, government meetings, and investor presentations. Trigger when the user says "keynote deck", "event presentation", "conference slides", "speaking deck", or asks to prepare a presentation for an event or high-level meeting.

Move the latest published LinkedIn content (article/post/comment) plus any attached images from outputs/content/linkedin into the typed datastore archive. Auto-fires on English or Russian "I published it on LinkedIn" phrasing. Asks for the content type if it cannot be inferred. Confirms before mutating disk.

LinkedIn Content Series

Benchmark Mullvad VPN relays by real network latency and connect on selection. Default invocation (/mullvad with no args) detects current connection state, disconnects if needed for an unbiased measurement, runs the parallel ICMP benchmark via scripts/mullvad-fastest.py, and presents the 3 fastest servers as a numbered menu. When the user replies with "connect 1", "connect 2", or "connect 3" (or an equivalent natural-language selection from the 3 shown), the skill executes mullvad relay set location HOSTNAME and mullvad connect and verifies the tunnel is up. Trigger on /mullvad, fastest mullvad server, switch mullvad server, check mullvad speed. Do NOT trigger on generic VPN questions, Mullvad account billing, other VPN providers, reading Mullvad help pages (use WebFetch), or when the user says the literal phrase "Mullvad" only as a topic of conversation without a clear operational intent.

EXPLICIT INVOCATION ONLY -- never auto-trigger. This skill activates ONLY when the user literally types '/osint-advanced' or says 'run osint-advanced'. DO NOT trigger from 'investigate', 'research', 'dig into', 'OSINT', 'intelligence on', 'background check', 'due diligence', 'profile', 'recon', 'dossier', 'deep dive', or ANY general intelligence request -- those belong to /osint. This skill queries specialized OSINT databases (sanctions lists, corporate registries, breach databases, threat actor platforms, infrastructure scanners) that the base /osint does not cover. Requires explicit user approval before each execution.