ai-tool-handoff

Mendeteksi ketika Luminance, Kira, atau alat peninjau massal serupa sedang digunakan, menyerahkan ekstraksi klausa volume tinggi ke alat tersebut, dan memeriksa kualitas keluarannya sesuai dengan kepercayaan…

npx skills add https://github.com/anthropics/claude-for-legal --skill ai-tool-handoff

AI Tool Handoff

Matter context

Matter context. Check ## Matter workspaces in the practice-level CLAUDE.md. If Enabled is (the default for in-house users), skip the rest of this paragraph — skills use practice-level context and the matter machinery is invisible. If enabled and there is no active matter, ask: "Which matter is this for? Run /corporate-legal:matter-workspace switch <slug> or say practice-level." Load the active matter's matter.md for matter-specific context and overrides. Write outputs to the matter folder at ~/.claude/plugins/config/claude-for-legal/corporate-legal/matters/<matter-slug>/. Never read another matter's files unless Cross-matter context is on.


Purpose

Luminance and Kira are good at one thing: reading 500 contracts and finding every change-of-control clause. They're less good at judgment — deciding whether a particular CoC provision is actually triggered by this deal structure.

This skill hands off the bulk extraction to the right tool, then runs the QA layer on what comes back.

Before you hand off: try tabular-review first (/corporate-legal:tabular-review). For anything the user's environment can handle — a few hundred documents, a defined column schema — native tabular review is faster to set up, has no per-document cost, and keeps the work product local. Hand off to Luminance/Kira when the corpus is genuinely too large, the team already has a license and workflow, or the matter requires a tool with a validated provenance chain.

Load context

~/.claude/plugins/config/claude-for-legal/corporate-legal/CLAUDE.md → AI-assisted review:

  • Tool in use (Luminance / Kira / none)
  • What it's used for (which clause types)
  • Trust level (use as-is / spot-check / full re-review)
  • Handoff process (who loads, who QAs)

If ~/.claude/plugins/config/claude-for-legal/corporate-legal/CLAUDE.md says no AI tool → this skill is a no-op. Everything goes through diligence-issue-extraction directly.

When to hand off

Hand off when all of:

  • Category has >50 documents (below that, faster to just read them)
  • Extraction target is a clause type the tool is good at (CoC, assignment, exclusivity, MFN, termination, auto-renewal)
  • Documents are reasonably uniform (all customer contracts on similar paper — not a mix of contracts, letters, and board minutes)

Don't hand off:

  • Bespoke or heavily negotiated documents
  • Side letters and amendments (context-dependent, tools miss the interaction with the main agreement)
  • Anything where the question is "what does this mean for the deal" not "does this clause exist"

The handoff

Step 1: Prepare the batch

  • Identify documents for the batch (from VDR inventory)
  • Specify extraction targets per ~/.claude/plugins/config/claude-for-legal/corporate-legal/CLAUDE.md (which clause types)
  • Note the materiality threshold so tool output can be filtered

Step 2: Load (or instruct the loader)

Per ~/.claude/plugins/config/claude-for-legal/corporate-legal/CLAUDE.md — who loads. If it's you, generate the load instructions. If it's someone else, generate the request:

## [Tool] Load Request — [Deal code] — [Category]

**Documents:** [N] docs from VDR folder [path]
**Load to:** [Tool workspace/matter]
**Extraction targets:**
- Change of control / assignment
- Exclusivity
- [etc. per `~/.claude/plugins/config/claude-for-legal/corporate-legal/CLAUDE.md`]

**Filter output:** Flag only where extraction target is present — no need for "no CoC clause found" for every doc.

**Return by:** [date]

Step 3: QA the output

When the tool returns results, apply the trust level:

"Use as-is": Ingest directly into diligence findings. (Only if ~/.claude/plugins/config/claude-for-legal/corporate-legal/CLAUDE.md says this — it's rare.)

"Spot-check X%": Randomly sample X% of flagged documents. For each, read the actual clause and compare to the tool's extraction. If error rate is low, accept the batch. If errors found, widen the sample.

"Full human review of flagged": Tool narrows the universe (500 docs → 80 with CoC clauses). Human reads all 80. Tool saved the time of reading the 420 clean ones.

Step 4: Judgment layer

The tool found the clauses. Now apply judgment:

For each flagged CoC provision: is it actually triggered by this deal?

  • Stock sale vs. asset sale vs. merger — different triggers
  • "Change of control" defined how in the contract — majority ownership? board control? something else?
  • Is there a carve-out for this type of transaction?

This is the part the tool can't do. Output goes to diligence findings in house format.

Output

The QA summary below is derived from VDR documents that are privileged, confidential, or both. It inherits the sources' privilege and confidentiality status — distribution beyond the privilege circle can waive privilege. Store with the matter's privileged files.

## AI Tool Handoff Summary — [Category]

**Tool:** [Luminance / Kira]
**Documents processed:** [N]
**Extraction targets:** [clause types]

### QA

**Trust level:** [per `~/.claude/plugins/config/claude-for-legal/corporate-legal/CLAUDE.md`]
**Sample size:** [N] docs spot-checked
**Error rate:** [X]% — [Accepted / Widened sample / Full re-review triggered]

### Results

| Clause type | Docs flagged | After judgment layer | Material |
|---|---|---|---|
| Change of control | [N] | [N actually triggered by deal structure] | [N above threshold] |
| Assignment | [N] | [N] | [N] |

**→ [N] findings added to diligence issues**
**→ [N] consents added to closing checklist**

Close with the next-steps decision tree

End with the next-steps decision tree per CLAUDE.md ## Outputs. Customize the options to what this skill just produced — the five default branches (draft the X, escalate, get more facts, watch and wait, something else) are a starting point, not a lock-in. The tree is the output; the lawyer picks.

What this skill does not do

  • It doesn't run Luminance or Kira — it manages the handoff and QA. A human (or the tool's own interface) runs the extraction.
  • It doesn't replace the tool's output with its own judgment entirely — if ~/.claude/plugins/config/claude-for-legal/corporate-legal/CLAUDE.md says spot-check 10%, check 10%, not 100%.
  • It doesn't decide the trust level — that's in ~/.claude/plugins/config/claude-for-legal/corporate-legal/CLAUDE.md, set at cold-start based on the team's experience with the tool.

Lebih banyak skill dari anthropic

analyzing-financial-statements
anthropic
Keterampilan ini menghitung rasio dan metrik keuangan utama dari data laporan keuangan untuk analisis investasi.
applying-brand-guidelines
anthropic
Skill ini menerapkan pencitraan merek dan gaya korporat yang konsisten pada semua dokumen yang dihasilkan, termasuk warna, font, tata letak, dan pesan.
creating-financial-models
anthropic
Skill ini menyediakan rangkaian pemodelan keuangan tingkat lanjut dengan analisis DCF, pengujian sensitivitas, simulasi Monte Carlo, dan perencanaan skenario untuk investasi…
board-minutes
anthropic
Menyusun notulen rapat dewan atau komite dalam format rumah Anda. Mendeteksi secara otomatis rapat dewan dan komite yang akan datang dari kalender Anda, meminta agenda dan…
crm-cleanup
anthropic
Memindai HubSpot untuk deal basi, kontak duplikat, dan kolom yang hilang, lalu memperbaiki apa yang disetujui pemilik. Menerima argumen scope opsional untuk deal, kontak,…
redshift-api
anthropic
Jalankan SQL terhadap Amazon Redshift — kirim pernyataan, periksa status, ambil hasil per halaman, dan jelajahi basis data/skema/tabel. Gunakan ini setiap kali pengguna ingin…
ticket-deflector
anthropic
Membaca email atau tiket pelanggan yang diteruskan, mengambil status pesanan/pengembalian dana dari PayPal dan riwayat akun dari HubSpot, menyusun balasan dengan nada yang sesuai dengan pemilik...
reg-feed-watcher
anthropic
Periksa umpan regulasi sekarang dan laporkan apa yang baru sejak pemeriksaan terakhir, difilter berdasarkan ambang materialitas Anda. Gunakan saat pengguna mengatakan "periksa umpan",…