Transcribe (Auto-Detect)
🤖 AI Drafted (Not reviewed)
Use this when: you don’t know the document’s script type and want automatic routing. Classifies the script type (Typescript / Manuscript / HTR / Paleography) then runs the matching transcription profile — including two-pass review for historical and archaic scripts.
| Folder | /Transcribe |
| Steps | 10 |
| Tags | preset, transcribe, auto-detect, classify, routing |
Steps, in run order
1. Files
Tool: Files — Pass through input files from workflow context
2. Classify Script Type
Tool: Classify Script Type — Detect whether a document is typescript, manuscript, HTR, or paleography
Settings this step uses:
| Option | Value |
|---|---|
save_to_db |
yes |
vision_mode |
llm |
What this step asks the model:
Examine this document image and classify its script/handwriting type.
Choose EXACTLY ONE type from:
- typescript : typewritten, printed, or born-digital text (no handwriting)
- manuscript : modern handwriting, 20th–21st century (cursive or print)
- htr : historical handwriting, 16th–19th century (legible but archaic letterforms)
- paleography : archaic script with non-standard letterforms, heavy abbreviation, or
specialist scribal conventions (pre-18th century or highly specialised)
Return a JSON object with exactly these fields:
{
"script_type": "<one of the four types>",
"confidence": <float 0.0–1.0>,
"notes": "<one sentence explaining the classification>"
}
Be conservative: if the image is ambiguous between htr and paleography, pick the harder
category (paleography) and lower your confidence. Only mark typescript when there is
clearly no handwriting. Do not add any text outside the JSON object.
3. Transcribe (typescript)
Tool: Transcribe — Extract text from images (OCR)
Settings this step uses:
| Option | Value |
|---|---|
language |
auto |
prompt |
Transcribe the text visible on this image. The document is typewritten or printed. Rules: - Output ONLY the transcription. No headings, preamble, commentary, or summary. - Preserve original layout, line breaks, and paragraph structure. - Preserve original spelling, capitalisation, and punctuation exactly as printed. - Include every visible text element — headers, body, footers, page numbers, marginalia, stamps, annotations. - For text you cannot confidently read, write [ILLEGIBLE] inline at that position. - If the image contains no text, output [no text]. |
update_page_content |
yes |
vision_mode |
auto |
What this step asks the model:
Transcribe the text visible on this image.
Language: transcribe in the language of the source. Do not translate, and do not assume the document is in English.
Rules:
- Output ONLY the transcription. No headings, no preamble, no commentary,
no summary, no notes, no explanations, no observations about quality or
legibility, no descriptions of seals or images, no language about the
difficulty of the handwriting.
- Preserve original layout, line breaks, and paragraph structure.
- Preserve original spelling and capitalisation, including ALL CAPS
headers if they appear that way.
- Preserve orthography exactly as written, including all diacritics
and accent marks (e.g., keep "Chocó" as "Chocó", never "Choco";
keep "Ramón" as "Ramón", never "Ramon").
- Do not strip accents, tildes, cedillas, or umlauts. If a mark is
visible, keep it.
- Include every visible text element — headers, body, marginalia, stamps,
signatures (transcribe the signed name as written), printed labels,
handwritten annotations.
- For text you cannot confidently read, use explicit uncertainty markers:
[ilegible] for unreadable text and [uncertain] for plausible-but-low-
confidence readings. Place the marker inline at the uncertain span.
Do not guess. Do not fill in.
- Do NOT invent dates, numbers, names, or words that are not legibly
present. Do not normalise dates ("23/7/1999" stays "23/7/1999", not
"1999-07-23").
- Do NOT repeat any portion of the transcription. Output each visible
passage exactly once.
- If the image contains no legible text, output the single token
[sin texto].
4. Transcribe (manuscript)
Tool: Transcribe — Extract text from images (OCR)
Settings this step uses:
| Option | Value |
|---|---|
language |
auto |
prompt |
Transcribe the handwritten text visible on this image. The document contains modern handwriting (20th–21st century). Rules: - Output ONLY the transcription. No headings, preamble, commentary, or summary. - Preserve original layout, line breaks, and paragraph structure. - Preserve original spelling, capitalisation, and punctuation as written — do not correct errors. - Include every visible text element — headers, body, dates, signatures, annotations, marginalia. - Transcribe signatures as written (the full signed name if legible). - For text you cannot confidently read, write [ILLEGIBLE] inline at that position. Do not guess. - Do NOT invent or normalise content. What is on the page is the source of truth. - If the image contains no legible text, output [no text]. |
update_page_content |
yes |
vision_mode |
auto |
What this step asks the model:
Transcribe the text visible on this image.
Language: transcribe in the language of the source. Do not translate, and do not assume the document is in English.
Rules:
- Output ONLY the transcription. No headings, no preamble, no commentary,
no summary, no notes, no explanations, no observations about quality or
legibility, no descriptions of seals or images, no language about the
difficulty of the handwriting.
- Preserve original layout, line breaks, and paragraph structure.
- Preserve original spelling and capitalisation, including ALL CAPS
headers if they appear that way.
- Preserve orthography exactly as written, including all diacritics
and accent marks (e.g., keep "Chocó" as "Chocó", never "Choco";
keep "Ramón" as "Ramón", never "Ramon").
- Do not strip accents, tildes, cedillas, or umlauts. If a mark is
visible, keep it.
- Include every visible text element — headers, body, marginalia, stamps,
signatures (transcribe the signed name as written), printed labels,
handwritten annotations.
- For text you cannot confidently read, use explicit uncertainty markers:
[ilegible] for unreadable text and [uncertain] for plausible-but-low-
confidence readings. Place the marker inline at the uncertain span.
Do not guess. Do not fill in.
- Do NOT invent dates, numbers, names, or words that are not legibly
present. Do not normalise dates ("23/7/1999" stays "23/7/1999", not
"1999-07-23").
- Do NOT repeat any portion of the transcription. Output each visible
passage exactly once.
- If the image contains no legible text, output the single token
[sin texto].
5. Transcribe — HTR Pass 1 (Draft)
Tool: Transcribe — Extract text from images (OCR)
Settings this step uses:
| Option | Value |
|---|---|
language |
auto |
prompt |
You are performing Handwritten Text Recognition on a historical document in a single careful pass. Take your time and think through hard lines before committing to a reading. Before transcribing, note in a single bracketed line: [Script: |
thinking_mode |
long |
update_page_content |
no |
vision_mode |
auto |
What this step asks the model:
Transcribe the text visible on this image.
Language: transcribe in the language of the source. Do not translate, and do not assume the document is in English.
Rules:
- Output ONLY the transcription. No headings, no preamble, no commentary,
no summary, no notes, no explanations, no observations about quality or
legibility, no descriptions of seals or images, no language about the
difficulty of the handwriting.
- Preserve original layout, line breaks, and paragraph structure.
- Preserve original spelling and capitalisation, including ALL CAPS
headers if they appear that way.
- Preserve orthography exactly as written, including all diacritics
and accent marks (e.g., keep "Chocó" as "Chocó", never "Choco";
keep "Ramón" as "Ramón", never "Ramon").
- Do not strip accents, tildes, cedillas, or umlauts. If a mark is
visible, keep it.
- Include every visible text element — headers, body, marginalia, stamps,
signatures (transcribe the signed name as written), printed labels,
handwritten annotations.
- For text you cannot confidently read, use explicit uncertainty markers:
[ilegible] for unreadable text and [uncertain] for plausible-but-low-
confidence readings. Place the marker inline at the uncertain span.
Do not guess. Do not fill in.
- Do NOT invent dates, numbers, names, or words that are not legibly
present. Do not normalise dates ("23/7/1999" stays "23/7/1999", not
"1999-07-23").
- Do NOT repeat any portion of the transcription. Output each visible
passage exactly once.
- If the image contains no legible text, output the single token
[sin texto].
6. Reference Corpus Search
Tool: Search — Find files matching a search query
Settings this step uses:
| Option | Value |
|---|---|
limit |
5 |
search_type |
hybrid |
status_filter |
completed |
7. Transcribe — HTR Pass 2 (Review)
Tool: Transcribe Review — Second-pass QA of a prior transcription against the image
Settings this step uses:
| Option | Value |
|---|---|
language |
auto |
lens |
abbreviations,orthography,names-and-dates |
thinking_mode |
medium |
update_page_content |
no |
vision_mode |
llm |
What this step asks the model:
You are a paleographer reviewing an EXISTING transcription of the manuscript image you are shown. The prior transcription is provided as context. Take your time: think through every hard line against the strokes before accepting or changing a reading. Change only what the strokes justify.
This review makes no assumption about the language, the genre or the conventions of the documents: it reads the strokes, not what the words are expected to be.
**This pass: abbreviations, contractions and formulary; orthography and consistency; names, dates and the final check.**
- Verify every expansion. Brackets belong ONLY around the letters or signs supplied by expansion; remove bracket noise around signs that are visible on the page.
- Expand an abbreviation the draft left unexpanded only where the expansion is unambiguous; otherwise keep the abbreviated form exactly as written.
- Identify the document's genre and check the draft against the formula that genre usually follows. Where the draft breaks the formula AND the image is damaged or ambiguous there, propose the formula reading as [UNCERTAIN: ...]; never insert formula text as if it were read.
- Sign-shape audit (the letters, characters or syllables this script writes): where the draft's reading implies a stroke the image does not show, correct it. Watch for signs of this script that look alike, and for ligatures or joined strokes read as two signs (or the reverse).
- Division: where this script divides its words, transcribe them as the scribe divided them and check the split the draft chose against the pen lifts and the gaps; where it does not, add no spaces. Keep the order the lines and signs run in on the page.
- Consistency: the same word or name must be spelled the way the scribe spells it elsewhere on the page; where two readings conflict, keep the better-attested one and mark the other [UNCERTAIN: ...].
- Collect every personal name, place name and date on the page (and in any reference material provided). The same person must read the same way throughout; a date must agree with the document's own heading or docket. Mark what cannot be resolved [UNCERTAIN: ...].
- Uncertainty audit: every [UNCERTAIN: ...] either resolves (the strokes decide) or stays with its best single reading; every [ILLEGIBLE] is confirmed truly unreadable at this resolution.
- Layout check: line breaks, marginalia, insertions, deletions, signatures and seals are where the page has them.
Rules for every pass: keep the writing as the scribe wrote it, in its own spelling, script, direction and diacritics; never modernise, translate or transliterate it. Brackets belong only around letters or signs supplied by expansion. Keep the conventions the draft already uses for what cannot be read ([UNCERTAIN: ...], [ILLEGIBLE], [deleted: ...]). Do not invent text that is not visibly present. If a line is already correct, return it unchanged.
Output the COMPLETE corrected transcription (same line breaks), nothing else.
8. Transcribe — Paleography Pass 1 (Draft)
Tool: Transcribe — Extract text from images (OCR)
Settings this step uses:
| Option | Value |
|---|---|
language |
auto |
prompt |
You are a paleographer transcribing a historical manuscript image in a single careful pass. Take your time and think through hard lines before committing to a reading. Before transcribing, note in a single bracketed line: [Script: |
thinking_mode |
long |
update_page_content |
no |
vision_mode |
auto |
What this step asks the model:
Transcribe the text visible on this image.
Language: transcribe in the language of the source. Do not translate, and do not assume the document is in English.
Rules:
- Output ONLY the transcription. No headings, no preamble, no commentary,
no summary, no notes, no explanations, no observations about quality or
legibility, no descriptions of seals or images, no language about the
difficulty of the handwriting.
- Preserve original layout, line breaks, and paragraph structure.
- Preserve original spelling and capitalisation, including ALL CAPS
headers if they appear that way.
- Preserve orthography exactly as written, including all diacritics
and accent marks (e.g., keep "Chocó" as "Chocó", never "Choco";
keep "Ramón" as "Ramón", never "Ramon").
- Do not strip accents, tildes, cedillas, or umlauts. If a mark is
visible, keep it.
- Include every visible text element — headers, body, marginalia, stamps,
signatures (transcribe the signed name as written), printed labels,
handwritten annotations.
- For text you cannot confidently read, use explicit uncertainty markers:
[ilegible] for unreadable text and [uncertain] for plausible-but-low-
confidence readings. Place the marker inline at the uncertain span.
Do not guess. Do not fill in.
- Do NOT invent dates, numbers, names, or words that are not legibly
present. Do not normalise dates ("23/7/1999" stays "23/7/1999", not
"1999-07-23").
- Do NOT repeat any portion of the transcription. Output each visible
passage exactly once.
- If the image contains no legible text, output the single token
[sin texto].
9. Reference Corpus Search
Tool: Search — Find files matching a search query
Settings this step uses:
| Option | Value |
|---|---|
limit |
5 |
search_type |
hybrid |
status_filter |
completed |
10. Transcribe — Paleography Pass 2 (Review)
Tool: Transcribe Review — Second-pass QA of a prior transcription against the image
Settings this step uses:
| Option | Value |
|---|---|
language |
auto |
lens |
abbreviations,orthography,names-and-dates |
thinking_mode |
long |
update_page_content |
no |
vision_mode |
llm |
What this step asks the model:
You are a paleographer reviewing an EXISTING transcription of the manuscript image you are shown. The prior transcription is provided as context. Take your time: think through every hard line against the strokes before accepting or changing a reading. Change only what the strokes justify.
This review makes no assumption about the language, the genre or the conventions of the documents: it reads the strokes, not what the words are expected to be.
**This pass: abbreviations, contractions and formulary; orthography and consistency; names, dates and the final check.**
- Verify every expansion. Brackets belong ONLY around the letters or signs supplied by expansion; remove bracket noise around signs that are visible on the page.
- Expand an abbreviation the draft left unexpanded only where the expansion is unambiguous; otherwise keep the abbreviated form exactly as written.
- Identify the document's genre and check the draft against the formula that genre usually follows. Where the draft breaks the formula AND the image is damaged or ambiguous there, propose the formula reading as [UNCERTAIN: ...]; never insert formula text as if it were read.
- Sign-shape audit (the letters, characters or syllables this script writes): where the draft's reading implies a stroke the image does not show, correct it. Watch for signs of this script that look alike, and for ligatures or joined strokes read as two signs (or the reverse).
- Division: where this script divides its words, transcribe them as the scribe divided them and check the split the draft chose against the pen lifts and the gaps; where it does not, add no spaces. Keep the order the lines and signs run in on the page.
- Consistency: the same word or name must be spelled the way the scribe spells it elsewhere on the page; where two readings conflict, keep the better-attested one and mark the other [UNCERTAIN: ...].
- Collect every personal name, place name and date on the page (and in any reference material provided). The same person must read the same way throughout; a date must agree with the document's own heading or docket. Mark what cannot be resolved [UNCERTAIN: ...].
- Uncertainty audit: every [UNCERTAIN: ...] either resolves (the strokes decide) or stays with its best single reading; every [ILLEGIBLE] is confirmed truly unreadable at this resolution.
- Layout check: line breaks, marginalia, insertions, deletions, signatures and seals are where the page has them.
Rules for every pass: keep the writing as the scribe wrote it, in its own spelling, script, direction and diacritics; never modernise, translate or transliterate it. Brackets belong only around letters or signs supplied by expansion. Keep the conventions the draft already uses for what cannot be read ([UNCERTAIN: ...], [ILLEGIBLE], [deleted: ...]). Do not invent text that is not visibly present. If a line is already correct, return it unchanged.
Output the COMPLETE corrected transcription (same line breaks), nothing else.