AGENT SKILL

paper-explainer-generate-site

Internal paper site-production phase that builds a local reading site from an exact run request, Content brief, Run Manifest, source structure/evidence, and report set.

License 许可
MIT
最近更新
2026-10-27
Stars
★ 1
Forks
0

01 前言 Front Matter

name
paper-explainer-generate-site
description
Internal paper site-production phase that builds a local reading site from an exact run request, Content brief, Run Manifest, source structure/evidence, and report set.
user-invocable
false

03 SKILL.md 正文

Generate Site — paper-explainer reports as an authored website

Apply [[explainer-content-workflow-base]], read its Artifact contract, and require the paper profile. Own only site authoring and local build; do not perform workflow discovery, user consultation, Content modeling, or Planning.

Require exact paths/digests for the run request, selected Content brief, Run Manifest, source-structure.md, spine.md, and every planned report. The unique site entry must be create or replace; otherwise return a structured missing/incompatible-input result. Never infer latest planning Artifacts.

Planning Artifacts supplement rather than replace reports and paper evidence. Consult the narrow source/OCR locus whenever fidelity requires it.

Turn a paper-explainer work dir (<dir>/<slug>/, at least one report under reports/ — typically overview.md plus the perspective reports) into a reading-guide website under <WORK_DIR>/site/: a landing page with report cards and one authored page per report, the overview's audio guide playable in-page.

The whole build pipeline — scaffold, parallel semantic-Markdown authoring, the single generator build, the semantic landing page, the nav manifest, the gotchas, and the success criteria — is the shared reading-site pipeline owned by [[explainer-reading-site-generate-base]]. Delegate the build to that skill. This skill adds only what is paper-specific: the canonical source-structure path, page-ordering profile, and landing vocabulary below. Do not re-derive the pipeline here; follow the base skill for every phase.

paper-explainer inputs

Source locator kind

Pass pdf-page to the shared page-authoring and fan-in sweep contracts. Paper prose continues to use [pNN]{.p} / [pNN–pMM]{.p}; no EPUB locator map applies.

Canonical source structure

Pass the absolute path to <WORK_DIR>/source-structure.md to the shared pipeline. It records the paper's source-authored headings separately from spine.md, which remains the authority for confirmed facts. Perspective-report headings are editorial artifacts and are never a substitute for this source structure.

Ordered page list — [{ slug, kicker }]

A paper's reports are not chapters; they are fixed perspectives. Resolve reports/*.md into the base skill's ordered list with this canonical table, keeping only the reports that exist and preserving this order:

report slug (reports/<slug>.md) kicker
overview 全体レポート
background 背景
method 手法
experiments 実験
discussion 議論
related-work 関連研究
  • overview is always first and is the landing hero CTA / home. Do not renumber when a subset is present — keep the canonical order for whichever exist.
  • Any report whose slug is not in the table (a hand-added report) is appended last in natural sort, with a kicker derived from a readable title-case of its slug. Do not drop it.

This satisfies the base contract (overview first; every existing report once; deterministic order; non-empty kicker). Reports are not color-coded — which report you are on is answered by the nav, the title, and the index.

Landing document-type vocabulary

  • guide kicker (hero eyebrow): 論文ガイド
  • cards-section heading: レポート
  • count-chip unit: レポート (hero chip reads 全Nレポート)

Site title = the paper title (from reports/overview.md's <h1>).

Build

Hand the profile and canonical source-structure path above to [[explainer-reading-site-generate-base]] and run its pipeline (Phase 1 scaffold → Phase 2 author reports + landing → Phase 3 build + nav manifest → hand off to [[explainer-reading-site-deploy]]). The base skill's Success criteria are the acceptance for this skill. Return the exact output inventory/digest and AI checks to the coordinator. Human interaction belongs to the coordinator. If site is a selected human gate, return the checked site as its checkpoint subject. Never treat approval as permission to deploy.

04 同仓库更多 Skills