# Archive Research

> How to browse, cite, and present archival findings — load AFTER search, when reading documents or reporting results. Covers research-integrity rules (never fabricate reference codes, always cite the exact page), old-Swedish translation, and the browse→synthesize workflow. Use when the user wants to browse or read pages, examine or interpret a document, translate old Swedish, or have findings cited and presented.

- Skill: `ai-riksarkivet/archive-research` (Agent Skill)
- Install (CLI): `npx skillmds@latest add ai-riksarkivet/archive-research`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ai-riksarkivet/archive-research/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: ai-riksarkivet (https://skillmd.com/u/ai-riksarkivet)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/ai-riksarkivet/archive-research

---


# Archive Research Guide

Research integrity, methodology, and workflows for historical research using
the Riksarkivet MCP tools.

## Research Integrity Rules

This is an academic research tool. Accuracy and proper sourcing are paramount.

1. **Never fabricate data.** Never guess or hallucinate reference codes, page numbers,
   dates, names, or any archival data. Every claim must come directly from tool results.
2. **Always cite precisely.** Cite the exact reference code and page number when
   presenting information from a document
   (e.g. "SE/RA/420422/01/A I a 1/288, page 66").
3. **Only use returned links.** Only use links explicitly returned by the tools
   (bildvisaren, ALTO XML, NAD links, IIIF URLs). Never construct or guess URLs —
   not even by combining a base URL with a reference code.
4. **Distinguish source from interpretation.** Clearly separate what the document says
   (quote or close paraphrase with quotation marks) from your own interpretation
   or translation.
5. **Flag uncertainty.** If a transcription is unclear, incomplete, or ambiguous,
   say so explicitly. Do not silently fill in gaps with plausible-sounding text.
   Remember: all transcriptions are AI-generated (HTR/OCR) and contain recognition
   errors — misread characters, merged words, and garbled passages are common.
6. **Mark translations.** When translating old Swedish, mark it as a translation
   and note when the meaning is uncertain.
7. **Admit gaps.** If you cannot find information the user is looking for, say so.
   Do not construct an answer from partial or unrelated results.

## Understanding the Research Goal

Before making your first search, ensure you understand what the user is researching.
If their intent is vague or unclear, ask clarifying questions:

- What **time period** are they interested in?
- What **type of documents** are they looking for (court records, church records,
  military, estates)?
- Are they researching a specific **person, family, place, or event**?
- What do they **already know** that could help narrow the search?

Every tool call includes a `research_context` parameter — always fill it in with
your best understanding of the user's research goal.

## Cross-Tool Research Workflow

### Phase 1: Discovery

Use `search_transcribed` and `search_metadata` to find relevant documents.
Search multiple related terms and spelling variants for comprehensive coverage.
Paginate through results (offset 0, 50, 100...) to discover all matches.

### Phase 2: Examination

Use `browse_document` to read full page transcriptions of promising results.
Start with the specific pages flagged by search, then examine nearby pages
for additional context.

### Phase 3: Context Building

- Browse adjacent pages to understand the full document context
- Search for related terms discovered during browsing
- Cross-reference dates, names, and places across multiple documents
- Use `search_metadata` to find related documents in the same archive series

### Phase 4: Synthesis

Present findings with proper citations, original text, translations,
and links to source materials (see Result Presentation Template below).

## Browsing Strategy

- Use **reference codes** from search results to browse specific documents
- Request **page ranges** (e.g., "1-10") for broad context or **specific pages**
  (e.g., "5,7,9") for targeted reading
- **Download nearby pages** when context seems to be missing from a transcript —
  adjacent pages often contain continuation of the same record
- **AI transcription errors** are common — all text is generated by HTR/OCR models and
  may contain misread characters, garbled words, or missing passages. When quoting,
  note obvious errors rather than silently correcting them. Compare with the original
  page image (bildvisaren link) when accuracy matters.
- **Blank pages** (`"(Empty page - no transcribed text)"`) are normal — cover pages,
  dividers, etc. are digitised but contain no text
- **Non-digitised materials**: When no transcriptions exist, the tool returns
  metadata only (title, date range, description, and viewing links)

## Result Presentation Template

When presenting document content, follow this structure:

### [Document title or description]
**Source**: [reference_code], page [N]

> [Original Swedish text, quoted verbatim from the transcription]

**Translation**: [Modern translation in the user's language]
*Note: [Any uncertainties in the transcription or translation]*

[Bildvisaren link from tool results]

### Example

### Court Protocol — Witchcraft Case
**Source**: SE/RA/420422/01/A I a 1/288, page 66

> "... then bekände hon at Diefwulen hadhe kommit till henne ..."

**Translation**: "... then she confessed that the Devil had come to her ..."
*Note: The word "Diefwulen" is an archaic spelling of "Djävulen" (the Devil).*

[View original page](https://sok.riksarkivet.se/bildvisning/...)

## Coverage Awareness

The archive has three access tiers — not all materials are searchable the same way:

- **Metadata catalog** (search_metadata): 2M+ records — titles, names, places, dates
- **Digitised images**: ~73M pages viewable via bildvisaren links
- **AI-transcribed text** (search_transcribed): ~1.6M pages — currently court records
  from 17th-18th centuries. These transcriptions are machine-generated and contain
  recognition errors; use fuzzy search (~) to compensate

Church records, estate inventories, and military records are typically cataloged
(searchable via metadata) and often digitised (viewable as images) but NOT
AI-transcribed (not searchable via full-text search). Always explain this to
the user when relevant.