Choreography: drafts and clauses in your firm's style
Generic AI gives you a generic clause. With the house context, it generates in your style, and what turned out good becomes a new entry in your clause library.
It's four in the afternoon and you need a new document for a case similar to about ten you've already handled. You ask AI without giving it any context and out comes text that's correct and soulless, in textbook format, not yours. The second time, you gather the three examples you trust most, paste them in as reference, and ask again: now it comes out almost in your style, just tweaking two sentences. You save that version in a named folder, so you don't have to rebuild the context again next week. The third time you need that kind of document, you don't start from zero, you start from what already worked, and the text comes out ready in minutes, not hours.
You need the draft of a loan agreement to close the deal today, and you ask AI with no context: out comes text that's correct and soulless, with interest and collateral terms outside your standard. You stop, gather the two loan agreements your team already uses and trusts, paste them in as reference, and ask again, saying what's different in this case. The draft comes out almost ready, just adjusting the grace period. You read it line by line, because AI doesn't know this borrower already missed a payment once, and you adjust the early-termination clause with that detail. The final version goes into the folder of financial templates, and the next similar loan starts from there, not from the generic textbook text.
It's four in the afternoon and a request for a service agreement draft for a new client lands on your desk. You ask a generic AI and out comes a contract that's correct and soulless: clauses that aren't your firm's, a venue that isn't the one you use, a tone that isn't yours. You sit down and rewrite everything, just like last week.
You need the draft of an image-rights assignment for the campaign launching tomorrow, and you ask AI with no context: out comes text that's correct and soulless, with a usage term outside your standard. You gather the two assignment contracts legal already approved before, paste them in as an example, and ask again, specifying this campaign's term and channel. The draft comes out almost in your format, just adjusting the territory clause. You read and adapt it, because AI doesn't know this influencer negotiated category exclusivity, and you add that into the right clause. The adjusted contract becomes a template in the image-rights folder, and the next similar campaign starts from that point, not from zero.
You need the offer letter for the candidate who accepted the role today, and you ask AI with no context: out comes text that's correct and soulless, with a benefit that isn't in your package and a probation period outside your standard. You gather the last two offer letters the company already used, paste them in as an example, and ask again, with this candidate's actual role and package. The letter comes out almost ready, just adjusting the start date. You read and adapt it, because AI doesn't know this candidate negotiated two remote days a week, and you add that into the right clause. The final version goes into the hiring templates folder, and the next similar offer starts from there.
You need the PRD for the next feature to align the team today, and you ask AI with no context: out comes a document that's correct and soulless, with a section that isn't from your template and a metric the squad has never used. You gather the last two PRDs the team approved, paste them in as reference, and ask again, explaining this feature's problem. The document comes out almost in your format, just adjusting the metrics section. You read and decide, because AI doesn't know this feature depends on an architecture decision made last sprint, and you adjust that dependency yourself. The final version goes into the approved PRDs folder, and the next similar feature already starts from that template.
You need the commercial proposal for the account that asked for a reply by end of day, and you ask AI with no context: out comes text that's correct and soulless, with terms that aren't yours and a pitch that isn't your pipeline's. You gather the two proposals that closed the most deals this quarter, paste them in as reference, and ask again, with this account's real numbers and history. The proposal comes out almost in your format, just adjusting the discount. You read and adapt it, because AI doesn't know this client already complained about implementation timelines in an earlier negotiation, and you reinforce that point in the proposal. The final version goes into the template proposals folder, and the next similar negotiation already starts from there.
You need the standard operating procedure for the new shipping flow going live tomorrow, and you ask AI with no context: out comes text that's correct and soulless, with a step that doesn't follow your process and an SLA outside your standard. You gather the two SOPs the operation already validates and trusts, paste them in as reference, and ask again, explaining what's different in this new flow. The document comes out almost in your format, just adjusting the final check step. You read and adapt it, because AI doesn't know this hub has a space constraint the others don't, and you adjust the step with that detail. The final version goes into the SOP folder, and the next similar flow already starts from that template.
You need the internal data-processing policy to file by end of day, and you ask AI with no context: out comes text that's correct and soulless, with controls that aren't yours and regulatory references outside your standard. You gather the last two policies the committee already approved, paste them in as reference, and ask again, specifying the process this policy covers. The document comes out almost in your format, just adjusting the legal-basis section. You read and adapt it, because AI doesn't know this process already had a non-conformance flagged in the last audit, and you reinforce that control yourself. The final version goes into the approved policies folder, and the next similar policy already starts from that template.
You need the deploy runbook for the new service shipping today, and you ask AI with no context: out comes a document that's correct and soulless, with a step that isn't from your architecture and a command outside your infra standard. You gather the last two runbooks the team already validated in production, paste them in as reference, and ask again, explaining what's different about this service. The document comes out almost in your format, just adjusting the rollback step. You read and adapt it, because AI doesn't know this service depends on a queue the others don't use, and you add that step yourself. The final version goes into the runbooks folder, and the next similar deploy already starts from that template.
You need the user research script for the session starting tomorrow, and you ask AI with no context: out comes text that's correct and soulless, with a question that doesn't cover the flow you want to test and a format outside your discovery standard. You gather the two scripts that already produced good findings in past research, paste them in as reference, and ask again, explaining what's different about this session. The script comes out almost in your format, just adjusting two questions. You read and adapt it, because AI doesn't know the last test already revealed confusion at the same point in the flow, and you reinforce that question yourself. The final version goes into the research scripts folder, and the next similar session already starts from that template.
You need next quarter's OKR plan ready for tomorrow, and you ask AI with no context: out comes a document that's correct and soulless, with a metric your team has never tracked and a cascade structure that isn't yours. You gather the last two OKR cycles that worked, paste them in as reference, and ask again, explaining this quarter's priority. The document comes out almost in your format, just adjusting two metrics. You read and decide, because AI doesn't know the expansion goal got deprioritized in yesterday's meeting, and you rewrite that objective yourself. The final version goes into the OKR cycles folder, and next quarter already starts with a tested template, not a document rebuilt from scratch.
How many times have you already rewritten the same clause from scratch? Confidentiality, termination, venue: you know the house way by heart, but you type it all again with every draft because generic AI doesn't know that way. And here's the point most people miss: the problem isn't AI, it's what you didn't give it. You asked for a contract and it gave you some contract. If you give it the house context, your templates, the clauses you trust, your way of drafting, it starts generating in your style. It's not magic, it's feeding it. Fair?
The core idea of this lesson. Generic AI gives you a generic clause. With the house context, it generates in your style. The choreography has four beats: you give the context and the request, AI generates the draft in the firm's format, you adapt with judgment (because AI doesn't know the case strategy), and what turned out good becomes a new entry in your clause library. The danger to name right away: pasting a generated clause without reading it is a recipe for disaster. AI speeds up the draft; the responsibility for the text stays yours, with your name underneath.
Back in lesson 3.2 you saw that a flow that works shouldn't be rediscovered every time, it should be packaged. Here it's the same economics applied to legal text: the clause you fine-tuned and trust shouldn't be rewritten from scratch with every draft. It should be stored, feeding the next generation.
01Why generic AI gives you a generic contract
Think of AI as a newly graduated lawyer, competent, but who never set foot in your firm. You ask for "a service agreement" and it hands you the textbook contract: technically correct, but without your venue, without your termination clause, without your tone. It nailed the generic because that's all you gave it. The text comes out soulless not from incompetence, but from lack of context.
The cure isn't switching AI or fine-tuning the isolated request. It's giving the house context before asking: your templates, the clauses you trust, examples of what "good" means to you. When AI has that in front of it, it stops inventing the standard from scratch and starts copying yours. At the end of the day, it's the difference between asking a stranger for a contract and asking your partner, who already knows how the house works.
02The clause library feeds the generation
The fuel for generating in your style is your clause library: the set of pieces you've already drafted, reviewed, and trust. Confidentiality, termination, venue, penalty, warranty. Each one is a ready piece. When you give this library to AI as context, it assembles the draft by fitting in your pieces, not the textbook's. It's the library becoming the factory.
The clause library as input, AI as the assembler, the draft in your style as the output:
Notice the dashed arrow that closes the loop: the new clause that turned out good goes back into the library. That's where the system grows on its own. Every well-made draft doesn't end at the client; it deposits a new piece in the stock, and the next generation already comes out better. At the end of the day, you're building a firm asset while you work.
03The four-beat choreography
Generating a draft in your style has four beats, and the order matters. Skip one and it falls apart.
- Beat 1, you give the context and the request. Present the library (or the templates you trust) and say what you need: type of contract, parties, what's specific about this case.
- Beat 2, AI generates the draft in the firm's format. With the context in front of it, it fits in your pieces and hands back a draft in your style, not the textbook's.
- Beat 3, you adapt with judgment. AI put together the skeleton, but it doesn't know the case strategy: the relationship with the client, the risk you want to cover, the gap you want to close. This beat is yours, and it's non-negotiable.
- Beat 4, what turned out good becomes a new entry. The clause you adjusted and approved goes into the library. The loop closes and the stock grows.
AI speeds up beats 1 and 2, which used to eat up the most typing. Beat 3 stays human work, with a name and responsibility attached. That's where the next section lives.
04The danger: pasting without reading is a recipe for disaster
I'm going to be direct, because this is the part that separates who uses it well from who gets burned. AI writes fast and writes well, and that fluency is a trap: the text looks so ready that you feel like pasting it and sending it. Don't. AI doesn't know that client already stiffed you, doesn't know that in this contract the venue has to be his home court because the deal requires it, doesn't know that the penalty it wrote is unenforceable in your case. It nails the form and misses the strategy, because the strategy was never in the context.
The rule is simple and has no exception: AI speeds up the draft, the responsibility for the text is yours. Your name goes under the draft, not AI's. Think of it as the intern who speeds up the skeleton: you'd never sign the intern's work without reading it line by line, and it's the same here. The speed of beats 1 and 2 only becomes a real gain if beat 3 is taken seriously. Cutting the reading to save ten minutes is trading ten minutes for a risk that costs the client relationship, or worse. Good.
Do it now
Take a draft you redo often for your real task and run the choreography once, start to finish:
- Build the context. Gather 2 or 3 clauses you trust and always use (confidentiality, termination, venue, whatever fits your case). This is the embryo of your library.
- Give the context and ask. Paste those clauses into AI, explain your standard, and ask for the draft for this specific situation, naming the parties and what's particular about it.
- Read everything and adapt with judgment. Mark what AI got right in form but wrong in the case's strategy. Adjust with your own head, not AI's.
- Deposit what turned out good. The clause you adjusted and approved: save it in a "clause library" folder. In the next draft, it already comes in as context.
Time it. Compare it to the time it would take rewriting from scratch. That delta is what you stop paying every single time.
You just ran the choreography once by hand. Once your standard matures, you can wire all of this into a system that carries the library, generates, and hands you back a draft ready for your review. That's what lesson N.jur.8 builds: the legal OS that orchestrates this choreography without you rebuilding the context every time.
Practice
1. Why does generic AI deliver a contract that's 'correct, but soulless,' outside the firm's standard?
2. In the four-beat choreography, which step is human work and can't be delegated to AI?
3. What makes using AI to generate drafts safe, and what is a recipe for disaster?
For the board
On the diagnosisthe problem is not the AI, it is what you did not give it. You asked for a contract and got any old contract.
On the librarythe house style has to be in the context. Without it the output comes out correct and soulless.
On what is yoursthe AI gets the form right and ignores the strategy, because the strategy was never in the context.
Thanks for the feedback. It helps sharpen the next lesson.