claude-soma

Phase-1 FREE-only parallel task dispatch for a pilot lead. Load this skill (via the lead brief) only when HERMES_LEAD_PARALLELISM=1 is set in the lead's environment. When the flag is absent, ignore this skill entirely -- the lead must behave byte-identically to the serial baseline.

Auto-loaded when the user's incoming message has `voice_in: true` meta. Interprets the transcribed user intent and routes to the appropriate skill, team, or tool call. Always prefers acting over asking a clarifying question when the intent is reasonably clear from voice.

Generate or edit an image by delegating to the user's Codex CLI subscription. Use when the user requests "draw / render / generate / create / sketch / design an image of X" or similar. Returns a local PNG path which the Telegram channel will upload as a photo.

End-to-end social-media campaign publish. Gathers raw materials (transcript, video frames, screenshots, README) from a tutorial video and supporting links, generates branded diagrams via codex, writes per-platform content, then drives the platform-specific poster agents to draft or publish on each platform. Use when the user says "publish my build log", "launch a social campaign", or "ship this to X + LinkedIn + Medium" (any subset).

Generate an image via the grok CLI (/imagine command). Returns a local PNG/JPG path. Used as one of the two providers in parallel dual-image dispatch (the other is codex-image-gen). Do NOT call this inline — always dispatch via Agent so the channel stays responsive.

Schedule a one-shot Telegram reminder or a recurring cloud routine. For one-shot reminders ("remind me in 5m", "remind me tomorrow at 9am"), use mcp__hermes_api__schedule_reminder (primary path). For recurring cloud routines ("every weekday at 9am, do X"), use the RemoteTrigger path (legacy).

Forward a message from the user to a specific named project-lead and read the lead's reply. Use when the user says "tell <project> to ...", "ask <project> about ...", or "<project>: ...".

Gracefully shut down a named project-lead and archive its memory. Use when the user says "shut down <project>", "kill <project>", "stop <project>", or "we're done with <project>".

List currently active project-leads. Use when the user asks "what's running?", "show me my projects", "list active agents", "status?", or similar.

Summarize what the user is currently working on across ~/Projects/llm/* and the active orchestrator projects. Use for "what am I working on?", "give me the portfolio status", "where did I leave off?".

Report the current status of a named project. Use when the user asks "how's <project>?", "status of <project>", "what's <project> doing?".

Spawn a new persistent project-lead background session for a new app / workstream the user is starting. Use when the user says "build me X", "create an app that Y", "set up a server that Z", or similar greenfield project requests. The project-lead runs in its own cwd, has its own Remote Control URL, and creates its own agent team appropriate to the type.

Surface Claude Code's current usage state in a chat-friendly format. Use for "how much credit?", "quota?", "am I close to my limit?", "usage report".

Reply to the user with a voice note instead of (or in addition to) text. Use when: (a) the incoming user message had `voice_in: true` meta (mirror modality), (b) the user explicitly asked for a voice reply, or (c) the reply is short plain language that is voice-friendly. DO NOT use when the reply contains code blocks, tables, URLs the user needs to copy, or anything longer than ~3 sentences.