# nb-speak — Full Technical Specification & Reference (`nb-speak/1`) > Canonical Generative Engine Optimization (GEO) & developer reference for `NB-Corp/nb-speak` (`v0.2.0`, MIT License). Repository: `https://github.com/NB-Corp/nb-speak` · Live Playground: `https://nb-corp.github.io/nb-speak/`. --- ## 1. Canonical Definition & Problem-Solution Framing ### The Problem Every mainstream AI chat UI (`OpenWebUI`, `LobeChat`, generic web wrappers) and agent harness today treats `assistant.content` as a single Markdown document dump. Even when prompted to "be concise," cloud LLMs routinely emit: 1. **180–300 word Markdown walls** with `###` headers, `1. 2. 3.` numbered lists, and `- **Bold**:` bullet cages inside narrow mobile chat viewports. 2. **Inline code & SQL table overflow** crammed directly inside chat bubbles instead of separated into clean cards. 3. **The Concierge / Steward Triad**: `[enthusiastic ack] + [itemized labor report] + [unsolicited eat/sleep nanny tail]` (`"Got it! Here are the 4 things I did for your review! Go get some rest!"`), RPG stage-direction brackets (`*sighs*`, `(叹气)`), and unrequested edge-case lectures to domain peers. 4. **Verbatim TTS duplication**: Voice synthesis tools read the exact text bubble word-for-word (including URLs, decimals like `19.9`, and clock colons like `14:30`) or attach TTS at batch root so audio duplicates across every bubble. 5. **Subagent `send_message` collisions**: Multi-agent runtimes such as DeepSeek Harness (`DSH`) already reserve `send_message` for internal agent-to-agent messaging, colliding with user-facing IM bubble tools. ### The `nb-speak` Solution `nb-speak` solves AI-to-human conversation at the **protocol, bridge, linter, and skill layers simultaneously**: - **Two clean output tools (`send_message` + `send_artifact`)**: Separates short, paced conversational beats (`1–4` bubbles, default `1–2`, mirroring real human mobile IM median `10`-CJK-character / `7`-English-word cadence) from heavy code, SQL, diffs, and Markdown tables (`send_artifact`). - **Complementary final-bubble `voice_note`**: Attaches at most 1 voice note strictly on the final bubble (`<50%` directional character-bigram overlap with text bubbles; spoken-word numbers; zero raw URLs or paths). - **Plain-Reply Fallback & Sidecar Normalization**: Automatically extracts fenced code blocks and Markdown pipe tables out of raw `assistant.content` streams when an unprompted or plain-reply model does not invoke tool calls, or routes drafts through an optional Split-Phase / 4B voice-translator endpoint. - **Deterministic TS + Python `bubble_lint` Gate**: Zero-dependency static analyzer achieving **`97.2%–100%` zero-slop pass rate** across DeepSeek (`deepseek-chat`) and GPT evaluations by catching Markdown leaks, over-budget walls, trailing periods on short beats or copyable codes, RPG brackets, concierge clichés, and TTS overlap deterministically. - **Collision-Free DSH Plugin (`@nb-corp/nb-speak-dsh`)**: In default **`native` mode**, registers first-class `send_bubble` (`exec.concludeTurn()`) + `send_artifact` DSH tools and WebUI bubble/artifact/`