Skip to main content

storm-write

Run STORM phase 3 — per-section article writing. This skill should be used when the user asks to "write the article sections", "draft the storm article", or invokes /storm:write. Writes each outline section in parallel with inline citations grounded in the research sources.

Informations de source

Dépôt
FradSer/dotclaude
Dernière activité de la source
22 juin 2026 à 11:11
Langue détectée de SKILL.md
anglais
Étoiles
594
Forks
50

Options d'installation

Le prompt qui vérifie d'abord la source est sélectionné par défaut. Vous pouvez passer à une commande directe ou télécharger une copie locale.

Vérifiez les fichiers source

Lisez SKILL.md et les fichiers associés affichés par SkillsMP avant de décider de l'installer.

Affichage de SKILL.md

SKILL.md
Instructions source · Aperçu en lecture seule
name
storm-write
description
Run STORM phase 3 — per-section article writing. This skill should be used when the user asks to "write the article sections", "draft the storm article", or invokes /storm:write. Writes each outline section in parallel with inline citations grounded in the research sources.
user-invocable
true
argument-hint
<topic> [--retrieve-top-k N] [--output-dir PATH | --save] [--force]
allowed-tools
["Read","Write","Edit","Glob","Grep","Bash(mktemp:*)","Bash(mkdir:*)","Bash(date:*)","Task","Skill"]
# /storm:write Phase 3 of the STORM pipeline. Writes each outline section in parallel (one Task subagent per section), each grounded in top-k relevant sources from the research information table, with inline `[n]` citations. ## CRITICAL: Prerequisites - Load `storm-engine` via the Skill tool. - `outline.md` MUST exist (phase 2 complete). If absent, stop and instruct the user to run `/storm:outline` first. ## Completion Contract This phase is complete iff `article.md` exists and every outline section (except Introduction/Conclusion/Summary placeholders) has body text. If `--force` is not set and it exists, skip and exit early. ## Procedure 1. Resolve output dir. 2. Read `outline.md` and `research/sources.json`. Index sources for retrieval (simple: rank by keyword/heading overlap with the section title). 3. **Identify sections** to write — skip any whose heading is exactly "Introduction", "Conclusion", or "Summary" (these are filled in `polish`). 4. **Parallel writing** — for each section, launch ONE Task subagent (single message, parallel). Each subagent: - Receives: topic, section heading, top-k relevant sources (default `--retrieve-top-k` 3), the full outline for context. - Writes the section body with inline `[n]` citations mapping to `sources.json` ids. - If no relevant sources, writes uncited and marks `<!-- TODO: no source -->`. - Returns the section markdown. 5. **Assemble** — concatenate sections in outline order under their headings into `article.md`. Preserve the heading hierarchy from `outline.md`. 6. Verify every non-placeholder section has body. Update `run-config.json`: `phases.write = "completed"`, section count, any TODO-flagged sections. ## Citation Rules - Only cite `[n]` where `n` is an existing `id` in `sources.json`. Never mint a new id here. - Place `[n]` immediately after the claim it supports. Multi-source claims use `[1][2]`. - Do not cite in placeholder sections. ## Output Report: number of sections written, number of citations used, list of sections flagged as TODO (no source), and the path to `article.md`.
Voir sur GitHub