We taught AI to forget the contract and keep the position

Every lawyer who has built a negotiation playbook knows where the real work is. You open three or four precedent contracts and a blank spreadsheet, and then you start the slow march. One row per clause. Your standard position. Your fallback. The line you will not cross. Where the risk sits. It takes hours, and the knowledge it captures usually lives in one senior lawyer's head, gone the moment they are on holiday.

The obvious wish is to point AI at your contracts and get the playbook out the other side. Plenty of tools claim exactly that. Most are quietly doing something far less useful.

What does a real playbook failure look like?

Here is the failure mode nobody advertises. You feed a contract in, and what comes back reads like "Clause 8.1 states that the Supplier shall indemnify the Client against...". Row after row. It looks like a playbook. In truth it is a summary of one contract wearing a playbook's clothes, and it breaks the day you try to use it on a new deal, because every row cites clause numbers and party names from a document that has nothing to do with the contract in front of you. A position that only works next to the contract it came from is not a playbook. It is a memo. And a memo does not travel.

What does a transferable playbook row actually require?

So we set one rule above all others. A playbook row has to be transferable. No party names. No source clause numbers. No quoting the source. State the position itself, usable on the next deal as readily as the one it was learned from. That is genuinely hard to make a model do, because a model's instinct is to narrate what it sees. Left alone, the first pass drifts straight back into a summary.

How do we enforce abstraction in code?

Which is why we do not trust the model to abstract on its own. Every proposed row runs through hard validators before it is allowed into your playbook:

• A clause number: rejected
• A party name: rejected
• Eight or more words quoted verbatim from the source: rejected
• Comparative voice ("version A says, version B says"): rejected

On one real workspace, the first attempt produced 42 criteria and the validators rejected 41 of them. That is not a failure. That is what careful looks like. You receive the version that survived the guardrail, not the model's first instinct.

How is Playbook Builder built?

Under the surface it is a multi-agent system built on the Claude Agent SDK. One agent reads your contracts and supporting material, works out which document is doing what, and extracts findings clause by clause, pointing at the exact source paragraph rather than quoting it. A second agent compiles those grounded findings into the structured playbook: clause families, criteria, a recommended primary position, a fallback, a risk level. The sheet you open is rendered deterministically from a single structured source of truth, and a QA agent checks that structured data before it reaches you. Give it one contract and you get a narrower playbook, and it tells you so. Give it two versions of the same deal, one pro-client and one pro-supplier, and it abstracts them into a market-practice range with a balanced baseline, rather than a list of both sides.

What can you do once the playbook exists?

A playbook sitting in a folder is just a nicer spreadsheet. The reason to make it transferable is what happens next. Point it at a new contract and it reviews that contract clause by clause: risk level, deviation from your position, fallback language. You build the standard once, and you reuse it on every review after. That loop only works because the playbook was built to travel in the first place.

What is Playbook Builder and what is it not?

It is AI-generated work product, designed to be reviewed by a lawyer, not to replace one. The review loop runs when you ask for it, not as a background autopilot. It captures the positions in the precedent you give it. It does not quietly learn your firm over the months. What it does is take the playbook that used to live in one senior lawyer's head and turn it into a structured asset the whole team can reuse, without the row-by-row afternoon.

Playbook Builder is available now in Genie 3.0.