What problem does it solve? Claude Code CLI routed through the pandai relay produces DSML marker interference, thinking blocks that swallow the generated prose, unstable word counts, and high per-chapter cost when writing Chinese web novel chapters. This Skill replaces that pipeline with Hermes native delegate_task subagents (glm-5.2-thinking) for the DRAFT, EXPAND, and POLISH steps, cutting per-chapter cost from $2.36 to about $0.05 and time from 712 to 238 seconds. ## Core Features & Use Cases - DRAFT generation: The main agent reads the chapter skeleton, recent timeline, and book configuration, then dispatches a leaf subagent with a goal containing the full skeleton and three hard rules (no real Chinese place names, system panels verbatim, word count and dialogue ratio targets). - EXPAND expansion: When a draft falls below 2000 characters, a second subagent expands it by adding dialogue exchanges, bystander reactions, and opponent responses—never padding with scenery or mechanical actions. - POLISH and validation: The main agent runs review_scan.py or chapter_fast_gate.py to verify word count (2000-2800), dialogue ratio (>=38%), and mechanical issues; truncation is explicitly forbidden. - Use Case: Writing chapter 66 of a Tomato Novel platform serial: the main agent dispatches DRAFT, detects the 1760-character draft is short, dispatches EXPAND, and receives a compliant 2738-character chapter in under 4 minutes. ## Quick Start Ask the agent to draft the current chapter using the delegate_task workflow defined in this skill, reading the chapter skeleton and book configuration first.
