write

name	write
description	Visible Harness write entry. Use for manuscript writing, citation-supported paper work, final documentation, README hardening, and GitHub Pages preparation.

Write

Use $write as the human-facing alias for paper and release-documentation work. This is not a separate Skill Contract.

Read and follow:

• ../../../.agents/skills/auto-paper/SKILL.md for manuscript work
• ../../../.agents/references/research-supervision/README.md for internalized paper, figure, benchmark, and review assets
• ../../../.agents/skills/docs-site/SKILL.md only after durable Markdown docs are finalized
• ../../../AGENTS.md
• ../../../CLAUDE.md

For manuscript work, $write may read PDFs, Markdown notes, LaTeX sources, bibliographies, figures, local reports, and iteration_log.json when needed. Treat iteration_log.json as a weak planning signal, not Conclusion Evidence. Experiment evidence should flow through docs/30_evidence/Experiment_Evidence_Index.*, whose entries cross-check iteration-log signals against reports and run artifacts. Research Wiki context and run requests should flow through docs/context/experiments.md; legacy docs/45_discoveries/Research_Wiki.md is a pre-migration fallback.

Blogs and review articles are still citation-supported writing unless the operator explicitly asks for an uncited opinion memo. Do not set citation_target_count: 0 for a blog or review by default. If the only source is an AI dialogue, PDF discussion, or unverified literature note, extract candidate references and mark them as unverified instead of silently omitting citations.

For PDFs and Markdown notes, $write must also look for figure/table planning cues such as Figure, Table, diagram, chart, plot, 图, 图表, 表格, 架构图, 路线图, 热力图, and 雷达图. When such cues appear, write figure_requirement_scan.md and either route through $auto-paper-figure or record a concrete USER_GATE / NOT_RUN reason.

When writing finds a missing experiment, baseline, ablation, seed sweep, metric export, or figure-generation gap, write the request to auto_paper_output/<paper_id>/run_request_register.{json,md} and return RUN_REQUEST so $run can plan the next WF10 iteration.

Do not upgrade claims beyond approved Claim Boundaries or available Conclusion Evidence.