playbook-as-code

Negotiation playbooks as data: an open YAML schema for "our position on each clause", a validator, a clause locator, and a Claude skill that reviews a contract against the playbook and emits a memo plus a redline-ready changeset.

- id: liability-cap
  topic: Limitation of liability
  severity: deal_breaker
  ideal:      Cap at 12 months' fees; exclusions only where required by law.
  acceptable: Cap at 12 months' fees; carve-outs for confidentiality and IP indemnity; 2x super-cap for data.
  walk_away:  Uncapped, or data-protection breach carved out with no super-cap.
  detect:
    any_of: ["aggregate liability", "shall not exceed", "limitation of liability"]
  fallback_language: |
    ...each party's total aggregate liability ... shall not exceed the Fees paid or payable
    in the twelve (12) months immediately preceding the event giving rise to the claim.
  authority: General Counsel
  jurisdiction_notes:
    SG: UCTA s 11 reasonableness applies; an unreasonably low cap risks being struck out.

Every legal team has a playbook. It lives in a Word table nobody opens, so every review starts from memory. The commercial products that solve this (Luminance, Robin, Spellbook) keep the playbook format proprietary. There is no open schema, which means no portability, no diffing your positions against a counterparty's, and no way for an AI assistant to review against your actual standards rather than generic ones.

The schema is the contribution. The tooling around it is deliberately thin.

pip install -e ".[dev]"

As a Claude skill: clone into .claude/skills/playbook-review/. SKILL.md at the repo root is the skill.

playbook validate playbooks/saas-vendor.yaml
playbook scaffold playbooks/saas-vendor.yaml examples/sample-msa.md     # what to read, per position
# ... assess (by hand or via the skill) into assessments.json ...
playbook memo playbooks/saas-vendor.yaml examples/sample-assessment.json
playbook changeset playbooks/saas-vendor.yaml examples/sample-assessment.json --contract examples/sample-msa.md

examples/sample-msa.md is a deliberately hostile MSA (3-month one-sided cap, uncapped data and IP carve-outs, Delaware law, 30-day termination for convenience inside a 36-month term). Running the vendor playbook against it produces five walk-aways, which is the point.

A playbook has a side (vendor / customer / mutual / ...), a contract_type, a jurisdiction, and a list of positions. Each position has:

Full schema: playbook schema.

The skill grades each position IDEAL, ACCEPTABLE, BELOW, WALK_AWAY or MISSING, always with a verbatim quotation as evidence. The vocabulary is fixed so memos are comparable across reviewers and over time.

{
  "changes": [
    {
      "position_id": "liability-cap",
      "severity": "deal_breaker",
      "verdict": "WALK_AWAY",
      "clause_ref": "7.2",
      "find": "<verbatim text in the contract>",
      "replace": "<replacement>",
      "comment": "<one line for the counterparty>",
      "insert_if_missing": false
    }
  ],
  "problems": []
}

find is checked against the contract text; a non-verbatim find is reported in problems and the command exits 1, so nothing reaches the redline tool that would fail there. Feed changes to docx-redline-mcp, legal-redline-tools, or a person.

Copy the closest bundled playbook. Keep acceptable honest: it is the tier your commercial team will sign at without asking you, so if you would actually want to be asked, move that wording up to ideal. Give every position a walk_away even if it is "none"; a missing walk-away reads as "anything goes".

• Not a contract analyser. It does not decide what a clause means; the skill (a reader with the contract in front of it) does.
• Not a redliner. It produces the changeset; other tools apply it.
• Not legal advice. The bundled playbooks are illustrative positions with Singapore-law notes, not a substitute for your own.

citecheck, oblig-register, sg-deadline.

MIT.