AI Completion Prompts: Making AI Draft Legal Documents Without Embarrassing You
This post uses hypothetical scenarios for illustrative purposes only. It does not describe any actual client, transaction, or representation, and is not legal advice.
A predictable failure mode has emerged in AI-assisted legal drafting, and it shows up within the first five pages of virtually every agreement generated by a generic prompt. The instruction is something like “draft a stock purchase agreement for a Series A financing,” and the output is a thirty-page document with a four-page definitions section (defining terms used once), mutual indemnities (in a transaction where only the seller has indemnification exposure), a notice provision requiring delivery by “nationally recognized overnight courier” (for a company where the CEO checks Slack, not the mailroom), and scattered throughout, hedging phrases like “it is important to note that” and “for the avoidance of doubt” deployed as structural filler. The document is complete in the sense that it covers every topic a stock purchase agreement should cover. It’s also wrong in the sense that it covers those topics generically rather than for this specific transaction, these specific parties, and this specific set of risks.
The problem isn’t the AI model. The problem is that a generic instruction to “draft this agreement” provides no constraints on what the output should contain, no guidance on what it should omit, and no framework for the model to evaluate whether its output is correct. The fix is to embed AI completion prompts, role-specific directives, directly in the document template itself.
What a Generic Prompt Produces and Why
Understanding the failure mode is the first step to designing the fix. When an AI model receives an instruction to draft a legal agreement, it draws on patterns in its training data to produce a document that looks like the agreements it has seen. Those agreements include large-company M&A purchase agreements, public-company proxy statements, venture-capital financing documents, commercial real estate leases, and everything in between. The model’s job, absent further instruction, is to produce something that resembles the statistical average of those documents.
That statistical average includes provisions that are common across agreement types even when they’re inappropriate for a specific transaction. Defined-terms sections appear because most training-data agreements have them, regardless of whether this agreement’s complexity warrants one. Mutual indemnities appear because they’re common in commercial contracts, even though most investment transactions have asymmetric risk profiles. Boilerplate notice provisions appear because they’re present in virtually every training-data agreement, unchanged since the era when physical mail was the primary delivery mechanism.
The result is a document that passes a superficial completeness check but fails a fitness-for-purpose review. Every provision is technically valid. Many provisions don’t belong in this specific agreement. And the guardrails needed to prevent these errors can’t be applied after the fact nearly as well as they can be built into the drafting process itself.
The Document-Embedded Prompt Approach
The core insight is that the AI completion prompt should not be a separate instruction that precedes the document. It should be part of the document’s template, embedded at the points where the AI will make drafting decisions. This approach treats the prompt as quality infrastructure: a set of constraints and directives that travel with the document and govern how any AI tool interacts with it.
A well-designed embedded prompt operates at several levels. At the document level, it establishes the role the AI is playing, the transaction context, and the boundaries of the task. At the section level, it specifies what each section must contain, what it must not contain, and what cross-references it must maintain. At the provision level, it directs the AI to make specific drafting choices rather than defaulting to generic forms.
Consider a hypothetical template for an early-stage stock purchase agreement. A document-level prompt might read: “You are drafting a stock purchase agreement for a seed-stage preferred stock financing. The company has fewer than ten employees and no general counsel. The transaction involves a single lead investor and three smaller participants. Draft for this specific context. Do not include provisions appropriate for larger or more complex transactions.” This single directive, embedded as a comment or instruction block at the top of the template, eliminates the most common category of AI drafting errors: provisions that don’t belong in this type of agreement.
Five Directives That Change the Output
Through iterative testing with legal document templates, five specific directive categories have proven effective at producing AI output that approximates what experienced counsel would draft.
First, the evidentiary-control directive. This tells the AI to treat every row in a schedule, every entry in a disclosure, and every item in a closing checklist as an evidentiary record. The practical effect is that the AI stops producing generic placeholder language (“such items as may be reasonably requested by Purchaser”) and instead produces specific, verifiable entries. A closing-conditions schedule generated under this directive will list specific deliverables rather than catch-all categories.
Second, the named-owner directive. This tells the AI to assign a named human owner to every action item, deliverable, and approval in the document. Instead of “the Company shall deliver,” the output becomes “the Company, through [Name/Title], shall deliver.” This directive prevents the common AI failure of producing documents where every obligation is assigned to an abstraction (“the Company,” “the Purchaser”) without identifying the person responsible for performance. In a transaction context, knowing that the CFO is responsible for delivering the capitalization certificate and the CEO is responsible for the officer’s certificate is operationally critical.
Third, the defined-terms reconciliation directive. This tells the AI to use a defined term only if the term appears at least three times in the operative provisions, and to expand any term used fewer than three times into its full definition at the point of use. This single directive eliminates the bloated definitions sections that are the most visible artifact of generic AI drafting. It also forces the AI to evaluate whether each defined term earns its place in the document, which is exactly the analysis an experienced drafter performs instinctively.
Fourth, the failure-simulation directive. This tells the AI to evaluate each covenant, condition, and deliverable by asking: “What happens if this obligation is not performed? Who notices, what is the consequence, and how is it enforced?” This directive produces documents with more thoughtful remedies provisions and, more importantly, surfaces obligations that the AI included by default but that serve no functional purpose. If the answer to “what happens if this isn’t performed?” is “nothing, because nobody would notice or care,” the provision probably doesn’t belong in the agreement.
Fifth, the counsel-sign-off directive. This tells the AI to never mark a legal conclusion, a representation, or a compliance determination as complete without flagging it for human counsel review. The practical effect is that the AI’s output includes inline markers (brackets, comments, highlight) at every point where a legal judgment is required. Instead of producing a representation that “the Company is in compliance with all applicable laws” as final text, the AI produces “[COUNSEL: confirm scope of compliance representation, including environmental, employment, and data-privacy carve-outs].” This transforms the document from a finished draft that might contain errors into a working draft that explicitly identifies where human judgment is needed.
Why Embedding the Prompt in the Document Matters
The reason to embed these directives in the template rather than in a separate prompt has to do with durability and portability. A prompt typed into a chat interface exists for one session. A directive embedded in a document template persists across every use of that template, across every AI tool that processes it, and across every user who works with it. When a firm maintains a library of transaction document templates, embedding AI completion prompts in those templates means that every AI-assisted draft produced from the library inherits the firm’s quality standards automatically.
This approach also addresses a governance concern that has become increasingly relevant as AI tools proliferate in legal practice. When an AI completion prompt is embedded in the document, it becomes part of the document’s version history. It can be reviewed, updated, and approved through the same process the firm uses for template maintenance. It creates a record of the instructions that governed the AI’s drafting decisions, which is relevant to both malpractice risk management and to the that practitioners owe their clients.
The broader insight is that AI-assisted legal drafting will improve not primarily through better models, though models will certainly improve, but through better prompts. And the most impactful prompts aren’t the ones typed into a chat window. They’re the ones built into the documents themselves, traveling with the template, enforcing quality constraints at the point of generation, and producing output that reflects the judgment of the practitioners who designed the template rather than the statistical average of the model’s training data.
For practitioners evaluating their own AI workflows, the question isn’t whether to use AI for drafting. The question is whether the AI is receiving instructions that produce work product you’d be willing to send under your name. If the answer is no, the fix isn’t to stop using the tool. The fix is to give it better instructions, embedded in the document, where they’ll do the most good.
If you are building or refining AI-assisted legal drafting workflows and want to discuss how embedded completion prompts can improve your corporate transaction documents, feel free to reach out to our firm manager, Magda, at Magda@montague.law, or fill out our contact form. Mention you read this post.
