Stats
Actions
Tags
From quasi
Use when the user wants to build an author profile from a scholar's representative books and papers.
How this skill is triggered — by the user, by Claude, or both
Slash command
/quasi:process-authorThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
搜索、下载、分析并综合用户指定作者的代表作。
搜索、下载、分析并综合用户指定作者的代表作。
从用户请求中归一化出:
author_name:kebab-case,如 donna-harawayfull_name:可由用户显式给出,否则由 author_name 反推后交给 search-agent 校正topic:可选范围提示.quasi/authors/{author_name}/manifest.json。status 控制 downstream phase;warning 字段不能替代可消费状态。process-book 的 chapter manifest / output 命名一致。主进程 (dispatcher)
├─ Phase 0: local author/work recall + manifest/cache resume
├─ Phase 1: search-agent (opus, 前台, only missing discovery) → manifest
├─ Phase 2: local representative-work reconcile → download-agent (sonnet, 前台) × 2
├─ Phase 3: 书籍处理
│ ├─ 逐本: extract-agent (sonnet, 前台)
│ ├─ 主进程: 读 manifest → 全部章节
│ ├─ analyse-agent ×N (opus, 后台, 跨书并行)
│ ├─ Glob 轮询
│ └─ 逐本: synthesis-agent(mode=book) (opus, 前台)
├─ Phase 4: analyse-agent ×M (opus, 后台) → 论文
├─ Phase 5: synthesis-agent(mode=author) (opus, 前台) → profile.md
├─ Phase 6: audit-agent (sonnet, 前台) → 校验 + 修复所有生成文件
└─ Phase 7: search-agent + quasi-helpers localise → 中译本 metadata 回填
author_name, full_name = parse_args()
profile_path = f"vault/authors/{author_name}.md"
manifest_path = f".quasi/authors/{author_name}/manifest.json"
books_path = f".quasi/authors/{author_name}/books.json"
papers_path = f".quasi/authors/{author_name}/papers.json"
# 0. LOCAL AUTHOR/WORK RECALL — before search-agent/download-agent.
# 先看 author profile / manifest / discovery caches;exact miss 后在 authors/books/papers/sources
# 做一把 rg fuzzy recall。高置信 profile 直接跳过;高置信 manifest/cache 则续跑。
if exists(profile_path):
report(f"已有作者 profile,无需重复处理: {profile_path}"); return
if not exists(manifest_path):
inspected_author = inspect_author_candidates(rg_fuzzy_recall(
tokens=[full_name, author_name, author_surname(full_name), topic],
paths=["vault/authors/*.md", "vault/books/*/00-overview.md", "vault/papers/*.md",
"sources/*", ".quasi/authors/*/manifest.json"],
))
if inspected_author.high_confidence_profile:
report(f"已有作者 profile,无需重复处理: {inspected_author.profile_path}"); return
if inspected_author.high_confidence_manifest:
manifest_path = inspected_author.manifest_path
elif inspected_author.candidates:
report_candidate_list(inspected_author.candidates, note="local author recall only; do not blindly skip")
# 1. DISCOVER — two structured search-agent calls (kind=book + kind=paper), only for missing caches.
# search-agent only returns curated JSON; it never writes files. The skill main
# process may cache those JSON payloads, then merges them into the manifest that
# Phase 2+ expects.
if not exists(manifest_path):
if not exists(books_path):
books_search = Agent("quasi:search-agent", foreground=True, prompt=f"""\
task: find top representative books by {full_name} on topic {topic}, sorted by citations
context:
kind: book
author: {full_name}
topic: {topic}
constraints:
count: 5
sort: citations
""")
write_json(books_path, books_search)
if not exists(papers_path):
papers_search = Agent("quasi:search-agent", foreground=True, prompt=f"""\
task: find top representative papers by {full_name} on topic {topic}, sorted by citations
context:
kind: paper
author: {full_name}
topic: {topic}
constraints:
count: 10
sort: citations
""")
write_json(papers_path, papers_search)
# Merge into the manifest shape Phase 2+ expects.
# (Pseudocode — at runtime: Read books_path + papers_path with Read tool,
# parse JSON, build manifest dict, Write to manifest_path.)
books_raw = read_json(books_path)
papers_raw = read_json(papers_path)
manifest = {
"author": full_name,
"slug": author_name,
"discovered": today_iso(),
"books": [
{**b, "status": "discovered", "md5": None}
for b in (books_raw if isinstance(books_raw, list) else books_raw.get("candidates") or books_raw.get("results", []))
],
"papers": [
{**p, "status": "discovered", "oa_url": None}
for p in (papers_raw if isinstance(papers_raw, list) else papers_raw.get("candidates") or papers_raw.get("results", []))
],
}
write_json(manifest_path, manifest)
# 2. ACQUIRE — 先 reconcile representative works 的本地 final/source/partial artifacts,
# 再 download 仍是 discovered 且没有本地 output/source 的 item。completed/partial 从 artifact 推断,
# 不新增 manifest status。
manifest = reconcile_representative_works_with_local_artifacts(manifest)
# skill merges per-item status + year_evidence back into the manifest.
# Batch policy: year_mismatch / year_ambiguous books DO NOT pause; skill
# accepts tmp_path → sources/{slug}.{ext} (slug authoritative), keeps
# status="acquired" so Phase 3 consumes the book, and records year_review /
# year_evidence / year_warning for end-of-run report.
write_json(manifest_path, manifest)
manifest = read_json(manifest_path)
# 2a. Books
discovered_books = [b for b in manifest["books"] if b["status"] == "discovered"]
if discovered_books:
book_result = Agent("quasi:download-agent", foreground=True, prompt=f"""\
kind: book
items:
{format_yaml_list([
{"slug": b["slug"],
"expected_author": full_name,
"expected_title": b["title"]}
for b in discovered_books
])}
output_dir: sources/
""")
# Merge per_item back into manifest.books. Agent status → manifest status:
# ok/year_* accepted into sources → acquired; download_failed → failed.
for item in book_result.per_item:
i = index_of(manifest["books"], slug=item["slug"])
if item["status"] == "ok":
manifest["books"][i]["status"] = "acquired"
elif item["status"] in ("year_mismatch", "year_ambiguous"):
# Override agent's "keep as tmp" — author batch accepts into sources
# anyway, but records the year evidence for offline review.
Bash(f"quasi-download accept --path {item['tmp_path']} --slug {item['slug']} --kind book --json")
manifest["books"][i]["status"] = "acquired"
manifest["books"][i]["year_review"] = item["status"]
manifest["books"][i]["year_evidence"] = item["year_evidence"]
manifest["books"][i]["year_warning"] = (
f"slug_year={item['year_evidence']['slug_year']} but "
f"recommended_year={item['year_evidence']['recommended_year']} "
f"({item['year_evidence']['recommendation_reason']}); "
f"file finalised under slug — re-run /quasi:process-book {item['slug']} "
f"to override if you want recommended_year"
)
else: # download_failed
manifest["books"][i]["status"] = "failed"
manifest["books"][i]["failure_note"] = item.get("verdict_note", "download_failed")
write_json(manifest_path, manifest)
# 2b. Papers
discovered_papers = [p for p in manifest["papers"] if p["status"] == "discovered"]
if discovered_papers:
paper_result = Agent("quasi:download-agent", foreground=True, prompt=f"""\
kind: paper
items:
{format_yaml_list([
{"slug": p["slug"],
"expected_author": full_name,
"expected_title": p["title"],
"identifiers": {"doi": p["doi"]}}
for p in discovered_papers
])}
output_dir: sources/
""")
# Papers: no year_evidence; status is just ok | download_failed.
# Fail-fast (download-agent paper flow has no candidate retry — single DOI).
for item in paper_result.per_item:
i = index_of(manifest["papers"], slug=item["slug"])
if item["status"] == "ok":
manifest["papers"][i]["status"] = "acquired"
manifest["papers"][i]["local_path"] = item["path"]
else:
manifest["papers"][i]["status"] = "failed"
manifest["papers"][i]["failure_note"] = item.get("verdict_note", "download_failed")
write_json(manifest_path, manifest)
# End-of-acquire summary (printed by skill main process for visibility):
n_year_warned = sum(1 for b in manifest["books"] if b.get("year_review"))
n_paper_failed = sum(1 for p in manifest["papers"] if p["status"] == "failed")
if n_year_warned or n_paper_failed:
report(f"Acquire summary: {n_year_warned} book year warnings, "
f"{n_paper_failed} paper download failures — review {manifest_path}")
# DOI liveness check is opt-in caller responsibility (bin no longer ships
# batch validate verb). If you want it, loop manifest.papers and
# `curl -sI --max-time 10 https://doi.org/{doi}` per entry.
# 3. PROCESS BOOKS
# Phase 2 结束时 download-agent 已对每本书调用 `quasi-download accept`，manifest 中
# books[*].slug 与 sources/{slug}.{ext} 均已是 canonical。Phase 3 起所有路径
# 直接复用 book.slug，不再重新派生。
manifest = read_json(manifest_path)
# 3a. 逐本提取（extract-agent 自含验证+修复）
for book in manifest.books where status == "acquired":
chapters_dir = f"processing/chapters/{book.slug}/"
if not exists(f"{chapters_dir}/manifest.json"):
Agent("quasi:extract-agent", foreground=True,
prompt=f"source_file: sources/{book.slug}.*, chapters_dir: {chapters_dir}")
# 3b. 跨书并行分析（每本书全部章节，不筛选）
all_chapters = []
for book in acquired_books:
if exists(f"vault/books/{book.slug}/00-overview.md"):
continue
book_manifest = Read(f"processing/chapters/{book.slug}/manifest.json")
all_chapters.extend(book_manifest.chapters)
for ch in all_chapters:
# slot: "01".."99" 真章 / "00a" 前言 / "99a" 后记 / "01b" 章间插曲
# chapter_label: 由 slot + title 推导的人类可读标签（"第3章" / "前言" / "后记" / "第2章（附）"）
if not exists(ch.output_path):
Agent("quasi:analyse-agent", background=True,
prompt=f"""\
type: A
book_slug: {book.slug}
book_title: {book.title}
slot: {ch.slot}
chapter_label: {ch.chapter_label}
chapter_title: {ch.title}
year: {book.year}
chapter_authors: {ch.authors or book.authors}
input: processing/chapters/{book.slug}/{ch.filename}
output: {ch.output_path}
topic: {topic}
""")
while not all_done:
sleep(30)
# 3c. 逐本概览(走大一统 synthesis-agent, mode=book)
for book in acquired_books:
if not exists(f"vault/books/{book.slug}/00-overview.md"):
Agent("quasi:synthesis-agent", foreground=True,
prompt=f"mode: book\noutput_dir: vault/books/{book.slug}/\nbook_title: ...\ntopic: ...")
# 4. PROCESS PAPERS
# paper.slug 必须形如 {author-surname}-{title}-{year}（全局唯一）
# 所有论文产出均落到扁平 vault/papers/，与全库共享 slug 命名空间
for paper in manifest.papers where status == "acquired":
output_path = f"vault/papers/{paper.slug}.md"
if not exists(output_path):
Agent("quasi:analyse-agent", background=True,
prompt=f"""\
type: B
title: {paper.title}
authors: {paper.authors}
year: {paper.year}
journal: {paper.journal}
doi: {paper.doi}
input: {paper.local_path}
output: {output_path}
topic: {topic}
""")
while not all_papers_done:
sleep(30)
# 5. SYNTHESIS (mode=author 走大一统 synthesis-agent)
# 如果作者 profile 不存在但代表作已有 final outputs,直接复用这些 outputs 做 synthesis,
# 不因为旧 slug/本次 slug 微差而重新下载或分析。vault/authors/{author_name}.md 才表示作者完成。
profile_path = f"vault/authors/{author_name}.md"
if not exists(profile_path):
paper_paths = [f"vault/papers/{p.slug}.md" for p in manifest.papers if status == "acquired"]
book_overview_paths = [f"vault/books/{b.slug}/00-overview.md" for b in acquired_books]
Agent("quasi:synthesis-agent", foreground=True,
prompt=f"mode: author\nauthor_name: {author_name}\nfull_name: {full_name}\ntopic: ...\n"
f"output_path: {profile_path}\n"
f"book_overview_paths: {book_overview_paths}\npaper_paths: {paper_paths}")
# 6. AUDIT
# 对本次生成文件运行 audit-agent。
# 每个 path 一个独立 Agent() 调用,顺序前台跑(audit-agent 单文件 ~30-60s)。
audit_targets = [profile_path]
for b in acquired_books:
audit_targets.append(f"vault/books/{b.slug}/") # overview + 所有章节
for p in manifest.papers where status == "acquired":
audit_targets.append(f"vault/papers/{p.slug}.md")
for path in audit_targets:
audit = Agent("quasi:audit-agent", foreground=True,
prompt=f"path: {path}")
# audit-agent only performs local minimal repairs. Escalated items mean the
# owning generation step must redo the corresponding file/subtree.
if audit.audit_result.escalated:
for item in audit.audit_result.escalated:
p = item.path
if p == profile_path:
Agent("quasi:synthesis-agent", foreground=True,
prompt=f"mode: author\nauthor_name: {author_name}\nfull_name: {full_name}\ntopic: ...\n"
f"output_path: {profile_path}\n"
f"book_overview_paths: {book_overview_paths}\npaper_paths: {paper_paths}\n"
f"overwrite: true\nreason: audit escalated {item.kind}: {item.reason}")
elif "/vault/books/" in p and p.endswith("/00-overview.md"):
b = find_book_for_overview(acquired_books, p)
Agent("quasi:synthesis-agent", foreground=True,
prompt=f"mode: book\noutput_dir: vault/books/{b.slug}\nbook_title: {b.title}\ntopic: ...\n"
f"overwrite: true\nreason: audit escalated {item.kind}: {item.reason}")
elif "/vault/books/" in p and "/ch" in basename(p):
b, ch = find_book_chapter_for_output(acquired_books, p)
Agent("quasi:analyse-agent", foreground=True,
prompt=f"""\
type: A
book_slug: {b.slug}
book_title: {b.title}
slot: {ch.slot}
chapter_label: {ch.chapter_label}
chapter_title: {ch.title}
year: {b.year}
chapter_authors: {ch.authors or b.authors}
input: processing/chapters/{b.slug}/{ch.filename}
output: {p}
topic: {topic}
overwrite: true
reason: audit escalated {item.kind}: {item.reason}
""")
elif "/vault/papers/" in p:
paper = find_manifest_paper_for_output(manifest, p)
Agent("quasi:analyse-agent", foreground=True,
prompt=f"""\
type: B
title: {paper.title}
authors: {paper.authors}
year: {paper.year}
journal: {paper.journal}
doi: {paper.doi}
input: {paper.local_path}
output: {p}
topic: {topic}
overwrite: true
reason: audit escalated {item.kind}: {item.reason}
""")
else:
report(f"audit escalated unknown author-processing path: {p}")
audit = Agent("quasi:audit-agent", foreground=True,
prompt=f"path: {path}")
if audit.audit_result.escalated:
report(f"audit still escalated for {path} after one regeneration pass")
return
# 7. LOCALISE
# 只对本次获得的 book overview 回填中译本 / 中文版本 metadata。
# search-agent 返回核验过的 localisations.zh.candidates;顶层调用 helper
# 写入 .quasi/localise/cndouban.json。helper 按原书 ISBN 幂等。
for b in acquired_books:
scan = Bash(f"quasi-helpers localise scan --path vault/books/{b.slug}/ --json")
if scan.pending == 0:
continue
search = Agent("quasi:search-agent", foreground=True,
prompt=f"kind: book\ncontext: read vault/books/{b.slug}/00-overview.md and search metadata/localisations")
candidates_file = write_temp_json(search.localisations.zh.candidates)
Bash("quasi-helpers localise write "
f"--book-path vault/books/{b.slug}/00-overview.md "
f"--candidates-file {candidates_file}")
# 8. OPEN IN MARPLE (best-effort UX)
# Open the final author profile. This must never fail the workflow; on failure,
# print the manual command and continue.
Bash(f"/opt/homebrew/bin/marple-cli open '{profile_path}' || marple-cli open '{profile_path}' || echo 'Marple open skipped; run: marple-cli open {profile_path}'")
.quasi/authors/{author_name}/manifest.json 由本 skill 主进程维护。
completed / partial 不写成新的 manifest status; completed/partial is inferred from artifact:
vault/authors/{author_name}.md、vault/books/{book.slug}/00-overview.md、vault/papers/{paper.slug}.md
是 final outputs;sources/*、processing/chapters/*/manifest.json、ch*.md 是可续跑 state。
Book status:
discovered — 已由 search-agent 选入代表作清单,尚未获取文件。acquired — sources/{slug}.{ext} 已可供 Phase 3 使用。failed — 获取失败,带 failure_note。Book year ambiguity 不改变 status:已 accept 入库的书仍是 acquired,
另写 year_review ∈ {year_mismatch, year_ambiguous}、year_evidence、
year_warning 供收尾报告和人工复查。
Paper status:
discovered — 已入清单,尚未获取 PDF。acquired — local_path 指向稳定 PDF。failed — 获取失败,带 failure_note。| 阶段 | 检查 | 跳过条件 |
|---|---|---|
| Phase 0/1 | local recall: vault/authors/{author_name}.md / manifest / discovery caches | profile 已有则跳过;manifest/cache 已有则续跑或只 search 缺的一侧;候选低置信度只列证据 |
| Phase 2 | representative works local reconcile | 已有 book/paper final output、source、chapter manifest 或 paper cache 时先复用/续跑;只有仍缺 source/output 的 discovered item 才 download |
| Phase 3a | {chapters_dir}/manifest.json | 存在则跳过提取 |
| Phase 3b | ch{NN}-*.md | 存在则跳过该章 |
| Phase 3c | 00-overview.md | 存在则跳过该书 |
| Phase 4 | vault/papers/{paper.slug}.md | 存在则跳过 |
| Phase 5 | vault/authors/{author-name}.md | 存在则跳过；若 profile 不存在但代表作 outputs 已存在,直接复用 outputs synthesis |
| Phase 6 | 无 —— 幂等,可重复跑 | 上次 audit clean 时几乎无成本 |
| Phase 7 | .quasi/localise/cndouban.json#by_isbn[isbn] | 已存在 entry(status found/none)则 helper scan 跳过 |
.quasi/authors/{author-name}/
└── manifest.json ← 采集状态机 + curation reason
vault/authors/{author-name}.md ← 单文件 profile（扁平）
vault/papers/{paper-slug}.md ← 该作者所有论文分析（与全库扁平共享）
vault/books/{book-slug}/ ← 该作者所有书的逐章分析
├── 00-overview.md
└── ch{NN}-{title}.md
processing/chapters/{book-slug}/
├── manifest.json
└── *.txt
sources/{book-slug}.*
paper.slug 与 book.slug 均为 {author-surname}-{title}-{year}，全库唯一。
npx claudepluginhub giraphant/quasi --plugin quasiCreates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.