Long Report Writing Agent

What problem it solves

Very long feasibility reports often have many chapters, strict terminology requirements, and cross-references that can drift during repeated revisions. A single-agent sequential workflow tends to either rewrite unchanged sections repeatedly or introduce inconsistencies in terms, metrics, and citations. long-report-agent structures long-report writing into planning, batched drafting, consistency review, and final integration, using multi-agent collaboration to produce the content and render the final output as docx.

How it works

The skill follows a staged pipeline: Phase 0 requirement confirmation → Phase 1 planner → Phase 2 parallel sub-agents → Phase 2.5 cross-chapter consistency review → Phase 3 final integration. Phase 0 confirms the topic, background, reference materials, and outline. References can come from local files, Feishu document RAG retrieval, or pasted content. Phase 1 generates plan.json and plan_outline_snapshot.md, which become the basis for batched writing.

Phase 2 supports up to five concurrent sub-agents and uses parallel_tracker.py to monitor batch state. Incremental updating can skip chapters whose content has not changed, avoiding unnecessary full rewrites. The skill also generates cover pages, table-style contents, colored chapter headings, highlight boxes, and styled tables, with six cover styles for different reporting contexts. v4.0 separates configuration, core logic, and the CLI entry point into src/config.py, src/engine.py, and src/cli.py, while keeping integrate_report.py as a backward-compatible facade. It also makes Mermaid CLI lazy-loaded to avoid subprocess side effects at import time.

Scope and cautions

This skill is best suited for feasibility reports, project proposals, and industry analyses with clear structure and substantial reference material. If the input is sparse, the output may still diverge from business specifics, so Phase 0 reference quality matters. After modifying underlying code, clean __pycache__ artifacts. If incremental mode prevents expected rewrites, delete content_hashes.json and regenerate. The default notification channel is log, but it can be configured to feishu or openclaw-weixin.