Case Summary

Case Summary

A user arrives with a pile of legal documents — often hundreds or thousands — and asks for a summary. This skill is the playbook for doing that well, across any practice area.

The work is map-reduce: map the corpus across the dimensions that matter, reduce the hits into a memo. The skill does not prescribe a specific search tool — it describes the work. references/BACKENDS.md lists concrete implementations the agent can use today.

The core describes what every case has (parties, timeline, evidence, claims, exposure, defenses, procedural status, red flags). Practice-area playbooks (references/PLAYBOOK-*.md) specialize the core for specific domains. Specialized document types (depositions, medicals, discovery, liens) route out to dedicated sibling skills per references/ROUTING.md.

Related skills

• case-chronology — when the deliverable is a dated timeline rather than a full memo
• case-intake-initial-fact-memo — for commercial-litigation intake memos
• case-viability-report — for go/no-go screening before accepting a matter
• pi-demand-summary — when the objective is a PI settlement package, not a case evaluation
• demand-letter — when the objective is a pre-suit demand letter
• deposition-summary — for deposition transcripts; route per ROUTING.md
• discovery-summary — for interrogatory/RFP/RFA responses; route per ROUTING.md
• medical-record-chronology — for dense medical-records sets; route per ROUTING.md
• medical-treatment-summary — for causation-focused treatment narratives
• evidence-liability-summary — for plaintiff-side liability element breakdowns
• lien-resolution-summary — for post-settlement lien tracking
• legal-strategy-summary — for motion-and-discovery roadmap deliverables
• settlement-summarization — for post-memo settlement activity

Checkpoint A: pre-ingest intake (mandatory)

Ask every time unless the user says "use defaults" or "just go." Gather:

1. Practice area guess — PI, commercial, IP, regulatory, criminal, family, bankruptcy, etc. The agent can update this after Inventory, but a guess narrows the initial playbook loader.
2. Expected document types — pleadings, discovery, depositions, medicals, contracts, corporate records.
3. Custodians / sources — who produced the files, and in what posture (client production, opposing-counsel production, third-party subpoena).
4. Date range — anchor the timeline.
5. User objective — evaluation memo, viability screen, demand prep, settlement prep, strategy roadmap. Different objectives trigger different sibling skills per ROUTING.md.
6. Privilege posture — work-product only, attorney-client privileged, or litigation-ready.
7. Venue / jurisdiction — affects SOL, statutory-notice requirements, governing law.

Defaults (if the user doesn't answer): PI/tort playbook; evaluation memo; attorney work-product; U.S. state-court posture. Label every default explicitly in the memo header.

The loop

1. Ingest. Put the corpus into something searchable. See references/WORKFLOW.md for the phase description and references/BACKENDS.md for implementations.

2. OCR. Any PDF that isn't already searchable will not be indexed usefully. Handle OCR before searching on files that need it. Asynchronous — keep moving.

3. Inventory. Pull the object list. Skim filenames to build a mental model: what practice area, what kinds of documents, how many of each. This is the only time the agent cares about every file.

4. Diagnose practice area. From the inventory plus a single global-overview query, pick one or more playbook modules from references/PLAYBOOK-*.md. Each has triggers in its frontmatter that match filenames and content signals. Record the choice in the memo header. When no module fits, proceed core-only and note it.

5. Map — iterative search per dimension. For each of the eight core dimensions in references/CORE-DIMENSIONS.md:

• Run the base queries from references/SEARCH-PLAYBOOK.md.
• Run the module queries from each loaded playbook for that dimension.
• Read the top 3–10 chunks per query. Cite everything the memo will rely on by object name + page.
• Where the corpus contains specialized document clusters (≥1 full deposition, dense medical records set, discovery responses, lien correspondence), delegate per references/ROUTING.md and consume the sibling skill's output rather than summarizing inline.

Don't brute-force read every file. Drive the search with focused queries, pull the 3–10 most relevant chunks per query, follow up when hits are load-bearing, route when a cluster warrants a sibling skill.

6. Reduce — synthesize. Assemble the memo using references/OUTPUT-TEMPLATE.md's core skeleton, plus the output sections contributed by each loaded playbook. Every claim cites an object + page (or chunk ID). No citation = cut or label as assumption.

7. Review. Run the quality checklist below (core items + playbook-specific items). Flag any unsatisfied items.

Checkpoint B: post-memo alignment (mandatory)

After delivering the draft, ask:

1. Did the right playbook module(s) get loaded? (List them.)
2. Which sibling skills were invoked, and did their outputs get integrated correctly?
3. Are there coverage gaps — dimensions the agent couldn't populate from the corpus?
4. Any claims in the memo that depend on unverified legal citations flagged [VERIFY]?
5. Is attorney review explicitly required before any downstream use?

If the user doesn't answer, default to: list the loaded playbooks and invoked siblings; flag any dimension that returned "no evidence found"; require attorney review.

What this skill does not do

• It does not value the case. Ranges belong in the memo when a loaded playbook supplies the methodology, but the skill doesn't substitute for attorney judgment about jurisdiction, venue, or negotiation posture.
• It does not resolve liens, draft pleadings, or send demands. It identifies them and routes to the sibling skill that does.
• It does not replace an attorney review of the final memo. Always label the memo AI-generated and require sign-off.

Scale notes

• < 20 files: skip the search loop; read every document directly. See references/BACKENDS.md small-corpus recipe.
• 20–200 files: one pass through the search playbook, one memo. ~1 hour.
• 200–2,000 files: expect 40–80 targeted searches across the dimensions, checkpoint notes as the agent goes. Multi-hour run.
• 2,000+ files: expect 100+ targeted queries; tighten scoping with module-specific filters and entity-mode searches; budget multi-hour run time. One corpus — the agent works it harder, not by fragmenting it.

Core quality checklist

Every matter:

• Every file is either searched-over or explicitly noted as out-of-scope
• Every OCR job finished (or its absence is noted in the memo)
• Parties, venue, posture correct with citations
• Timeline continuous; gaps >30 days flagged
• Each claim element has at least one evidentiary citation or explicit gap note
• Each asserted defense has supporting-evidence citation or a "pled without support" note
• Exposure/remedies section shows methodology, not an attorney valuation
• Red flags section is honest, not decorative — each item names a next action
• Every claim in the memo traces to an object + page or is labeled an assumption
• Unverified statutes/rules/cases marked [VERIFY]
• Memo labeled AI-generated, requires attorney review

Plus every playbook-specific item from each loaded PLAYBOOK-*.md's Quality checklist section.

Troubleshooting

• OCR didn't run on a file. The file isn't searchable. Record the file in the memo's out-of-scope list, try a different OCR engine if the backend supports one, or ask the user to re-upload a text-searchable version.
• Search returns noise, not relevant hits. Run a global-overview query first to discover the vocabulary the documents use; re-issue targeted queries using that vocabulary. Narrow with keyword anchors (party names, dates, case numbers) once known.
• Practice area ambiguous; no playbook fits cleanly. Proceed core-only (core dimensions + SEARCH-PLAYBOOK.md base queries). Note the ambiguity in VIII Strengths / Weaknesses / Red Flags. Offer the user a recommendation to narrow.
• Wrong playbook picked. Re-enter step 4 (Diagnose practice area) with the new information. Load the correct module; keep hits from the previous pass that still apply.
• Corpus too large for a single pass. Prioritize the dimensions that matter for the user's stated objective. Run a triage pass (parties, posture, top 3 claims, headline exposure) first, then expand to full coverage if time allows. Surface to the user which dimensions are deep vs. triage in the memo header.
• Sibling skill unavailable or returns nothing useful. Note it in the memo header (Sibling skills invoked), record the cluster the agent wanted to route, and summarize the cluster inline with reduced depth. Attorney review should consider whether a focused sibling-skill run is warranted before the memo is used downstream.
• Multi-matter corpus (the folder mixes unrelated cases). Surface to the user before proceeding. The skill summarizes one matter at a time; don't silently merge.