There is no shortage of generic advice about using AI for legal work, most of it amounting to "be careful." That is true and useless. What follows is the opposite: a set of opinionated practices earned from watching legal and compliance drafts succeed and fail, each one stated as a rule with the reasoning that makes it worth following. Some of these will feel stricter than necessary. They are strict on purpose, because the cost of being wrong in this domain is asymmetric.
These practices assume you are using a model to draft and accelerate, with a qualified human always in the loop. They are not a substitute for legal judgment. They are the operating discipline that makes the model's output safe to put in front of that judgment without wasting the reviewer's time.
Read these as positions, not platitudes. Each one exists because skipping it produced a problem worth not repeating.
Rule One: Ground or Do Not Draft
The reasoning
The model fabricates authority with total fluency. The only reliable defense is to supply the governing text and forbid reliance on anything else. If you cannot provide the source, you are not ready to draft, because the model will invent the authority your document depends on.
How to apply it
- Paste the controlling regulation, clause, or policy into every prompt.
- Instruct the model to quote the governing language before applying it.
- Treat any claim not traceable to provided text as unverified.
This rule is the foundation; the others build on it. The full grounding rationale appears in Everything That Matters When You Prompt for Legal Writing.
Rule Two: Make Uncertainty a First-Class Output
The reasoning
A model that hides its uncertainty produces drafts that look finished but rest on guesses. You want the opposite: a draft that tells you exactly where it is shaky. Honest uncertainty is more valuable than smooth completeness here, because it directs scarce reviewer attention.
How to apply it
Instruct the model to flag gaps explicitly, mark its riskiest assumptions, and decline to fill anything the provided text does not support. Reward the flag as a success. A draft full of honest "[GAP]" marks is better than a polished one that buried the same gaps in confident prose.
Rule Three: Protect Operative Language Religiously
The reasoning
The model's instinct toward readable prose quietly weakens obligations: shall becomes should, including without limitation gets trimmed. These edits look like improvements and change the legal substance. Operative language is not style, it is the content.
How to apply it
Instruct the model to preserve operative terms and defined terms exactly, then verify by hand. Do not delegate this check to the model alone, because it does not perceive the change as an error. This is the failure mode detailed in Seven Prompting Habits That Sink Legal and Compliance Drafts.
Rule Four: Fix Jurisdiction and Audience Up Front
The reasoning
A model with no jurisdiction blends legal regimes into an average that fits nowhere, and a model with no audience defaults to a dense register that may fail a plain-language requirement. Both errors are silent and both are preventable with one sentence each.
How to apply it
Open every prompt by stating the governing jurisdiction, the reader, and any comprehension standard the document must meet. Instruct the model to flag anything that depends on jurisdiction rather than resolving it silently. These two lines prevent a disproportionate share of errors.
Rule Five: Build Self-Critique Into Every Draft
The reasoning
The cheapest reviewer is the model itself, used correctly. Asking it to attack its own draft surfaces issues before a human spends time, and it routes the human straight to the risky parts. This roughly halves the effort a qualified reviewer needs to spend.
How to apply it
After drafting, run a pass that asks the model for its three riskiest assumptions, any unsupported claim, and the spots a reviewer should scrutinize. Confirm defined-term discipline in the same pass. Deliver this critique alongside the draft so the reviewer starts where the risk is.
Rule Six: Never Let Speed Eat the Human Gate
The reasoning
The model's biggest danger in this domain is that it is fast and looks finished, which tempts teams to ship without qualified review. That temptation is exactly when mistakes get through. The gate is non-negotiable, and treating it as optional is how all the other defenses fail at once.
How to apply it
Make qualified human review a required step for every legal and compliance output, with no exceptions for small or urgent items. Use the model's speed to make the review fast, not to eliminate it. A worked example of holding this line is in Inside One Compliance Team That Rebuilt Drafting Around Prompts.
Rule Seven: Keep the Record
The reasoning
Compliance work invites the question "how was this produced?" A clean record of prompts, sources, and human edits answers it cleanly and also lets you improve your prompts over time. The absence of a record turns a reasonable question into a problem.
How to apply it
Retain the inputs and the review trail for legal and compliance outputs. Confirm your data handling permits processing any confidential material before you paste it. Treat the audit trail as part of the deliverable, not an afterthought.
Rule Eight: Separate Drafting From Judgment
The reasoning
The model can draft, structure, and even critique its own work, but it cannot exercise legal judgment or take responsibility for being right. Blurring those two roles is how teams drift into trusting the model with decisions it cannot own. Keeping a clear line between what the model produces and what a human decides preserves accountability where it belongs.
How to apply it
Frame the model's output as a proposal, never a conclusion. The model proposes language; a qualified person decides whether it is correct and appropriate. State this framing in your process so no one mistakes a fluent draft for a vetted decision. The model's confidence is a drafting artifact, not a sign-off.
Making the Rules Stick
Wire them into a template
Rules that depend on memory fail under deadline. Encode the grounding instructions, jurisdiction and audience fields, gap-flag convention, and self-critique request into a reusable template so the safeguards fire by default. The discipline then survives busy days and new team members rather than relying on anyone remembering each rule.
- Build the grounding and gap-flag instructions into every template.
- Make jurisdiction, reader, and standard required fields to fill.
- Include the self-critique and term-discipline checks as fixed steps.
Revisit the rules as models change
A practice that held for one model version can need adjusting when the model updates. Treat each major model change as a reason to re-test your templates against known-good outputs. The rules themselves are durable, but how aggressively a given model fabricates or softens language can shift, so verify rather than assume continuity.
Frequently Asked Questions
Aren't these rules stricter than necessary?
Deliberately, because the cost of error in legal work is asymmetric. A draft that is slightly over-cautious wastes a little reviewer time; a draft that misstates an obligation can create lasting liability. The rules are calibrated to that asymmetry, trading a bit of friction for a large reduction in tail risk.
Which rule matters most if I can only follow one?
Ground or do not draft. Without supplied authority, the model fabricates the very foundation of the document, and no downstream practice fully recovers from that. Grounding is the rule the others depend on.
How does self-critique save reviewer time?
By surfacing the model's riskiest assumptions and unsupported claims before a human looks, it points the reviewer straight to where attention is needed. Instead of reading the whole draft cold, the reviewer starts at the flagged risk. That focus roughly halves the effort a careful review requires.
Can I trust the model to protect operative language itself?
No. The model perceives smoother prose as better and does not register that softening shall to should changes the obligation. Instruct it to preserve operative terms, but verify by hand. This is one check you cannot delegate to the model alone.
Why keep a record if a human reviewed everything?
Because compliance work invites the question of how a document was produced, and a clean trail of prompts, sources, and edits answers it without friction. The record also feeds prompt improvement over time. Human review and record-keeping serve different purposes; you want both.
Do these rules slow the work down?
A little, and that is the point in a domain where speed is the main risk. In practice, grounding and self-critique often speed the overall cycle by reducing rework and focusing review. The rules trade raw drafting speed for a faster, safer path to a shippable document.
Key Takeaways
- Ground every draft in supplied authority; if you cannot provide the source, you are not ready to draft.
- Make uncertainty a first-class output by rewarding explicit gap flags over smooth, false completeness.
- Protect operative and defined language by hand, since the model weakens it while thinking it improves the prose.
- Fix jurisdiction, audience, and comprehension standard up front to prevent a disproportionate share of silent errors.
- Build self-critique into every draft to surface risk early and focus the human reviewer where it counts.
- Treat the human review gate and the audit trail as non-negotiable parts of the deliverable, not optional extras.