Mapping Legal Cases

Maps a folder of heterogeneous legal documents into a structured analysis brief in a .casebase/ subfolder.

What it does

Given a folder containing any mix of legal case documents, this skill produces 10 cross-referenced analysis documents in a .casebase/ subfolder:

1. DOCUMENTS.md — full inventory of every file with status
2. GLOSSARY.md — bilingual keyword index (Which terms appear where?)
3. PRIVACY_FLAGS.md — sensitive data locations (IDs, accounts, medical, etc.)
4. PARTIES.md — who's who (client, opposing, judges, experts, witnesses)
5. TIMELINE.md — chronological events with source citations
6. CLAIMS.md — legal claims, arguments, basis, status
7. EVIDENCE.md — evidence mapped to each claim
8. GAPS.md — missing documents, contradictions, unanswered questions
9. RISKS.md — case weaknesses, opposing-side-favorable material
10. DEADLINES.md — forward-looking dates with OVERDUE/URGENT flags

Plus MAPPING_LOG.md — audit trail listing processed files, pending OCR/transcription items, and environment diagnostics.

When to use

• Case intake (first time touching a case folder)
• Building a case brief before drafting a motion
• Preparing for a hearing (verify every date is known)
• Onboarding a colleague to a case
• Identifying missing documents or contradictions
• Refreshing analysis after new documents arrive
• Auditing a case folder for sensitive data before sharing

How it works

Two-wave architecture keeps token usage low:

Wave 1 — Inventory (runs first, alone):

1. Scans the folder via scripts/extract.py --scan
2. Extracts text from each file once using scripts/extract.py --cache
3. Caches all extractions to .casebase/.cache/<hash>.json (persistent across runs)
4. Loads the right keyword pack(s) based on detected source languages
5. Builds DOCUMENTS.md, GLOSSARY.md, PRIVACY_FLAGS.md

Wave 2 — Analysis (4 parallel tasks, if platform supports; otherwise sequential):

1. parties-timeline → PARTIES.md + TIMELINE.md
2. claims-evidence → CLAIMS.md + EVIDENCE.md
3. risks-gaps → GAPS.md + RISKS.md
4. deadlines → DEADLINES.md

Wave 2 tasks read from the cache and glossary — they never re-extract files.

Step 1 — Pre-flight

Before doing anything else, run the pre-flight check to know what extraction tools are available in this environment:

python scripts/extract.py --preflight --pretty

Parse the JSON output. Key fields:

• capabilities.pdf_text — can we read PDFs?
• capabilities.word — can we read Word documents?
• capabilities.excel — can we read spreadsheets?
• capabilities.ocr_images — will scanned content auto-OCR?
• capabilities.transcription — will audio/video auto-transcribe?
• install_hints — what's missing

The skill always works for plain text, CSV, JSON, markdown. For PDF/Word/Excel, at least one backend must be available — the extractor falls back through multiple options.

If critical capabilities are missing, show install_hints to the user but continue — flag affected files rather than blocking.

Step 2 — Resolve target folder

If the user provided a path, use it. Otherwise use the current working directory.

Verify it is a directory and is not a system location (e.g., C:\Windows, /etc) — if so, ask the user to specify a folder.

Step 3 — Check for existing map

If .casebase/ already exists with documents, ask:

.casebase/ exists with {N} documents:
- DOCUMENTS.md ({X} lines)
- ...

What would you like?
1. Refresh — regenerate outputs (keep cache for speed)
2. Full refresh — also delete cache and re-extract every file
3. Update — regenerate only specific documents
4. Skip — use existing map as-is

Wait for user choice.

Step 4 — Pre-scan and threshold

python scripts/extract.py --scan "<case_folder>" --pretty

Show the user:

Case folder: {path}
Found {total} files:
  - PDFs: {pdf}
  - Word documents: {word}
  - Spreadsheets: {spreadsheet}
  - Emails: {email}
  - CSVs: {csv}
  - Markdown: {markdown}
  - Text: {text}
  - Images: {image}       ← needs OCR (auto if tesseract installed)
  - Audio: {audio}         ← needs transcription (auto if whisper installed)
  - Video: {video}         ← needs video analysis + transcription
  - Other: {other}
Total size: {size}

If total > 150 files, ask for confirmation:

Large case folder detected ({total} files). Estimated time: {estimate}.
Proceed?
1. Yes — map everything
2. Sample — most-recent 50 files only (preview)
3. Filter — skip audio/video/images, text-bearing only
4. Cancel

Step 5 — Create structure

mkdir -p "<case_folder>/.casebase/.cache"

Step 6 — Wave 1: Inventory + glossary + privacy

Do this work in order. If your platform supports subagents (e.g. Claude Code, OpenCode), this can be a dedicated task; otherwise do it inline in the current agent.

6a. Extract every file (populate cache):

For each file in the pre-scan file list:

python scripts/extract.py "<relative/path>" \
    --cache "<case_folder>/.casebase/.cache" \
    --case-root "<case_folder>" \
    --max-chars 200000

Track extraction results — note files with needs_ocr, needs_transcription, needs_video_analysis, or success: false for MAPPING_LOG.md.

6b. Detect corpus language(s):

Read the language_detected field from each cache JSON. Take the mode (most frequent) as the primary language. Possible values: en, he, ar, ru, es, fr, de, pt, zh.

6c. Load ONLY the relevant keyword packs:

Always load: references/keywords/common-en.md (baseline)

Also load based on detected languages:

• Primary language he → also load references/keywords/he.md
• Primary language ar → also load references/keywords/ar.md
• Primary language ru → also load references/keywords/ru.md
• Primary language es → also load references/keywords/es.md (if present)
• Primary language fr → also load references/keywords/fr.md (if present)
• Primary language de → also load references/keywords/de.md (if present)
• Primary language pt → also load references/keywords/pt.md (if present)

If a file is absent, skip it silently — that jurisdiction pack hasn't been contributed yet.

6d. Build GLOSSARY.md by scanning every cached extraction for terms in the loaded keyword packs. For each term, record the files where it appears. Include:

• Terms by category (dates, money, parties, procedures, etc.)
• Regex-matched patterns (date strings, currency amounts, case numbers)
• Auto-detected entities (capitalized names/orgs appearing 3+ times)
• File-name signals (dates, versions, case numbers in filenames)

6e. Build PRIVACY_FLAGS.md by scanning for privacy patterns defined in the keyword packs:

• National IDs, passports, credit cards, bank accounts, medical record numbers, phone numbers
• Mask values in the output (e.g., 123-45-****) — do not echo full sensitive values

6f. Build DOCUMENTS.md with per-file rows: path, type, size, pages (if PDF), language, one-line summary, status tags ([PENDING OCR], [PENDING TRANSCRIPTION], etc.).

See references/output-templates.md for the exact template structures.

Step 7 — Wave 2: Analysis (4 tasks, can run in parallel)

Each Wave 2 task reads from the cache + GLOSSARY.md. None of them re-extract files.

If your platform supports parallel subagents (Claude Code, OpenCode, OpenHands, etc.), spawn these 4 as parallel tasks. Otherwise run them sequentially.

Task 2a — PARTIES.md + TIMELINE.md — read GLOSSARY to find files with party names / dates, read those cached JSONs, build structured outputs.

Task 2b — CLAIMS.md + EVIDENCE.md — identify legal arguments (by party) and cross-reference evidence to each claim.

Task 2c — GAPS.md + RISKS.md — identify missing documents, contradictions, unfavorable facts. GAPS covers absence; RISKS covers present-but-problematic.

Task 2d — DEADLINES.md — extract every forward-looking date, sort ascending, flag OVERDUE / URGENT (<14 days).

See references/output-templates.md for templates.

Step 8 — Write MAPPING_LOG.md

Orchestrator-level audit trail:

• Timestamp, case folder path
• Total files scanned / successfully extracted
• Pending OCR list (files with needs_ocr: true)
• Pending transcription list (files with needs_transcription: true)
• Pending video analysis list
• Failed / unsupported list
• Environment (tools available, from preflight)
• Documents generated (names + line counts)

Step 9 — Summarize

Show the user:

• All 10 documents generated with line counts
• Privacy warning prominently if PRIVACY_FLAGS.md has content
• Pending extraction count if any items need OCR / transcription / video analysis
• Language-appropriate output — write this summary in the corpus primary language (Hebrew, English, etc.)
• Next-step suggestions: review TIMELINE, DEADLINES, refresh after changes

Critical rules

1. Cite every finding — every fact in every output document must reference a source file path like file.pdf:p3 or email.eml.

1. Match corpus language — write output documents in the language of the source documents. Use the mode of language_detected values from cache JSONs. Default to Hebrew if mixed and Hebrew is present; otherwise English.

1. Mask sensitive values — in PRIVACY_FLAGS.md, never quote full ID numbers, credit cards, account numbers. Mask with asterisks.

1. Extract once, analyze many times — never re-run extract.py on a file already in cache. Wave 2 tasks read cache JSONs.

1. Pending files are not ignored — tag them [PENDING OCR], [PENDING TRANSCRIPTION], [PENDING VIDEO ANALYSIS] in relevant documents and log them in MAPPING_LOG.md.

1. Forward slashes in paths — use / not \ in all documented file references (Unix-style, works everywhere).

1. Lazy-load keyword packs — load common-en.md always, but only load language-specific packs when the corpus language matches.

Supporting files

• scripts/extract.py — universal extractor (PDF, Word, Excel, email, image, audio, video, text, CSV, JSON, markdown). See references/document-handlers.md.
• references/document-handlers.md — per-format extraction details and fallback chains.
• references/output-templates.md — exact templates for all 10 documents.
• references/keywords/common-en.md — base English legal terminology.
• references/keywords/{lang}.md — language-specific packs (loaded on demand).

Platform notes

• Claude Code / OpenCode / Cursor / OpenHands / similar — this skill uses the Agent Skills standard and works in any agentskills.io-compatible platform.
• Subagent support — if available, Wave 2 tasks parallelize. If not, they run sequentially. Output is identical.
• OCR / transcription — used automatically when tesseract / whisper are installed. Never required; always falls back to [PENDING...] flags.
• Network access — the skill does not require internet access. All extraction is local.

Output language override

The user may explicitly request an output language even if it differs from corpus language:

• "map this case in English" → force English output
• "in Hebrew" → force Hebrew
• "same as documents" → default behavior (auto-detect from corpus)