科研 Skill

古文书学手稿释读

paleography-manuscript-transcription

Supports transcription, abbreviation expansion, and script-based dating of handwritten manuscripts using HTR tools. Use when a source is handwritten, not printed.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Paleography and Manuscript Transcription

Supports reading, transcribing, and dating handwritten historical documents — diplomatic transcription of letter forms and abbreviations, script classification, and the codicological description needed to argue a manuscript's date and origin. Complements historical-ocr (which targets printed and lightly-degraded text with Tesseract/Kraken) by covering the specifically handwritten, codicological, and script-dating side of manuscript work that OCR engines do not do.

When to use

  • The source is handwritten (a letter, charter, chronicle manuscript, diary, marginal note) rather than printed, and needs a reliable transcription.
  • A manuscript's date or origin is argued partly from its script (e.g. "is this Carolingian minuscule or a later imitation?", "does the hand match the other known autograph letters of this author?").
  • Abbreviations, ligatures, or non-standard orthography in the source need systematic expansion with the expansion marked, not silently normalised.

NOT for: printed or typed historical documents (use historical-ocr); authorship attribution from prose style once a transcription exists (use stylometry-authorship-attribution); comparing multiple transcribed witnesses of the same text (use textual-criticism-collation, which this skill feeds with diplomatic transcriptions).

Prerequisites

  • Good-quality digital images of the manuscript (minimum ~300 DPI, colour or greyscale; raking light or multispectral images if available for faded or palimpsest text) — this skill does not do image capture.
  • For automated handwritten text recognition (HTR) at scale: an account on Transkribus (https://readcoop.eu/transkribus/, free tier available, browser- and desktop-based) or a local install of Kraken (pip install kraken, open-source, scriptable, better suited to batch/CLI pipelines and to scripts with published community models, e.g. medieval Latin or specific national hands). Both require either a pre-trained model for the script in question or enough manually transcribed ground truth (typically 5,000–15,000 transcribed words minimum) to fine-tune one — HTR accuracy on an untrained hand without a matching model is often poor and must be verified, not trusted.
  • A paleographic reference for the script family and period in question (e.g. Cappelli's dictionary of abbreviations for Latin/Italian manuscripts, or a published paleographic atlas for the relevant script) — this skill provides a workflow, not a substitute for trained paleographic knowledge of a specific script and period.

Step 1 — Describe the manuscript (codicology)

Before transcribing, record the physical facts that support later dating and provenance arguments:

  • support (parchment/paper — paper has datable watermarks; parchment does not), collation (how bifolia are gathered into quires), number of lines per page, ruling pattern.
  • ink and script changes across the document (a single hand throughout, or multiple hands/campaigns of writing — look for changes in letter forms, ink colour, line spacing).
  • any marginalia, corrections, or later additions, and whether they are in the same hand as the main text.

Keep this as codicology.md — it is evidence for any later dating argument and should not be reconstructed from memory after the fact.

Step 2 — Classify the script

Identify the script type using standard paleographic terminology for the tradition in question (e.g., for Latin-script manuscripts: Caroline minuscule, Gothic textualis/cursiva, humanist minuscule; for Greek: majuscule vs. minuscule bookhand; equivalent typologies exist for Hebrew, Arabic, and other scripts — consult the relevant tradition's own paleographic literature). Note diagnostic letter forms (e.g. the shape of a, g, ligatures like &/et, the presence/absence of specific abbreviation marks) that support the classification, not just the type name — a dating argument built on "this looks like Gothic" without naming the diagnostic letter forms is not verifiable by another scholar.

Step 3 — Transcribe diplomatically first

A diplomatic transcription preserves exactly what is on the page — including abbreviations, superscript letters, unusual spelling, and line/page breaks — before any normalisation:

Example line (diplomatic):
  "qͨ dñs noster ihc̄ xp̄s dixit"

Expanded (editorial, abbreviation marked):
  "q[uia] d[omi]n[u]s noster ih[esu]c[hristu]s dixit"

Mark every expansion with brackets so a reader can tell what is on the parchment and what the transcriber supplied. Never silently expand an abbreviation — a wrongly resolved abbreviation is a fabricated reading, not a transcription.

If using Transkribus or Kraken for the bulk pass, treat the model's output as a draft requiring line-by-line verification against the image, not a final transcription — HTR confidence scores (where the tool reports them) flag likely errors but do not replace visual checking, especially for abbreviations and damaged text.

# Kraken CLI, once a matching recognition model is available:
# kraken -i page.png page.txt segment -bl ocr -m model.mlmodel

Step 4 — Expand abbreviations with a logged key

Maintain abbreviation-key.md mapping each abbreviation mark or convention used in the manuscript to its expansion and the authority for that expansion (a specific paleographic reference, or "inferred from parallel passages in this same manuscript" — which is weaker evidence and should be marked as such).

Step 5 — Argue the date (if required)

Paleographic dating is comparative, not absolute: argue by placing the hand alongside securely dated comparanda (dated or datable manuscripts/documents with similar letter forms), not by citing a general "this script was used from X to Y" range alone — broad script-period ranges can span a century or more. State the dating as a range with the comparanda that bound it, and name any contrary ink/support evidence (e.g. a watermark date that conflicts with the script date — flag the conflict rather than resolving it silently in the write-up).

Quality checks

  • Every bracketed expansion in the transcription has a corresponding entry in abbreviation-key.md.
  • Spot-check at least 10% of HTR-generated lines against the manuscript image directly (not against a prior transcription) before treating the output as reliable; report the observed character- or word-error rate on that sample.
  • Any dating claim names the specific comparanda (manuscript, date, holding library/shelfmark) it is argued from — "paleographically datable to the 12th century" with no comparison named is not a checkable claim.
  • Diplomatic and expanded versions are kept as separate, clearly labelled outputs — never overwrite the diplomatic layer with the expanded one.

Limitations

  • This skill does not replace trained paleographic expertise: script classification and dating require familiarity with the specific regional and period conventions, which vary enormously (a workflow tuned to Latin manuscripts does not transfer directly to Arabic, Chinese, or Mayan script traditions — consult that tradition's own paleographic literature and, ideally, a specialist).
  • HTR model availability is uneven: well-resourced scripts and periods (medieval Latin, early modern German/English) have strong public models; many regional hands, non-Latin scripts, and non-European traditions have little or no pre-trained model coverage, and manual transcription or substantial fine-tuning data may be the only realistic path.
  • Watermark- and ink-based dating methods (not detailed here) can corroborate or contradict script dating; this skill covers the script/codicology side only and flags, rather than resolves, conflicts with other dating evidence.
  • Damaged, palimpsest, or heavily abbreviated passages may be genuinely illegible; mark these [illegible] rather than guessing a plausible reading.

示例提示词

帮我转录这份手写古代文献,标注缩写展开并依据字体风格推断年代。

所属插件

人文学科面向历史、哲学、语言学、考古与艺术研究:TEI/OCR 与数字档案、论证分析、语料库与语音学、放射性碳测年、乐谱与音频分析,以及人文论证写作与审稿。

安装该插件会一并获得同领域的 62 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

古文书学手稿释读能做什么?

Supports transcription, abbreviation expansion, and script-based dating of handwritten manuscripts using HTR tools. Use when a source is handwritten, not printed。

如何安装古文书学手稿释读?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「古文书学手稿释读」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

古文书学手稿释读需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

古文书学手稿释读经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。