Hermes Marketing

Agent Ignores Skill Body but Obeys System Prompt

A developer post describes a common bug in agent frameworks that use a base system prompt plus conditionally-loaded skill bodies: instructions placed inside a skill's body (e.g., SKILL.md) are only included in the model context when that skill triggers, while the base persona (always-on) is included on every request. The authors debugged a Hermes agent where a vessel name change in a skill never appeared because the Navigator skill body was not loaded for plain queries. The post recommends classifying rules as always-on vs conditional, moving invariants (identity, safety, units, anti‑fabrication) to the base persona (e.g., SOUL.md), templating shared values from one source of truth, and adding deploy steps for every runtime prompt artifact to avoid drift.

blogs.city Turns AI Chats Into Persistent Blog Memory

blogs.city is an AI-native blogging and wiki platform that converts high-value conversations with large language models into published posts, wiki pages, and a durable external memory the model can read back. The platform provides a "publish Skill" editors can load (supporting Claude Code, Cursor, Trae, OpenClaw, Hermes and other compatible AI editors) and exposes catalog, search and read APIs so an AI editor can discover and load prior posts before a new session. Content is stored as plain Markdown files on disk, an index uses Redis, and the system supports exports (Obsidian Vault ZIP, PDF), built-in themes, RSS, mobile PWA, SEO fields, comment moderation, and templates for wiki pages. The product emphasizes user data ownership and a workflow where dialogue becomes an auditable corpus for future LLM sessions.

Atomic Mail Enables Agent Email via Open JMAP Standard

Atomic Mail has launched an agent-native email service in open alpha built on the open JMAP standard published by the IETF. The service allows autonomous AI agents to create accounts, own inboxes, read and send messages, and draft replies without requiring vendor-specific SDKs. Agents can connect via three paths — a Model Context Protocol (MCP) server, an AgentSkill package, or the JMAP API directly — and the platform returns plain-language error hints to help agents self-correct. The alpha hosts inboxes on the atomicmail.ai domain with free accounts; the company plans to move those accounts into a free tier of a paid product later. Atomic Mail says the service already works with current agent environments and coding assistants such as Claude (Anthropic), Codex (OpenAI), OpenClaw and Hermes.