Knowing that legal prompting requires grounding and review is one thing. Sitting down with a blank prompt and an actual deadline is another. This article is the do-this-then-that version: a sequence you can follow today to take a legal or compliance writing task from request to a draft that is safe to hand to a reviewer. No theory beyond what you need to perform each step.
The sequence assumes you are producing a first draft for a qualified person to refine, which is the only responsible way to use a model for this work. Each step builds on the last, and skipping one tends to surface as a problem two steps later. Work through them in order the first few times until the rhythm becomes habit.
We will use a running example throughout: drafting a data-retention clause for a client's vendor agreement. The steps generalize to policies, disclosures, and other legal documents.
Step One: Gather the Authority Before You Prompt
Collect the governing text
Before opening the model, assemble the material that governs the task: the relevant regulation, the client's existing policy, the counterparty's template, whatever actually controls. For the retention clause, that means the applicable data-protection rule and any retention commitments the client has already made.
- Pull the exact governing language, not a summary.
- Note any prior commitments the new text must stay consistent with.
- Identify what you do not have, so you know the gaps going in.
Confirm you can use the material
Check that your tooling permits pasting this content. Confidential contracts and client data carry handling obligations. Resolve that question now, not after you have fed sensitive text into the model.
Step Two: Set the Context in the Prompt
State jurisdiction, reader, and document type
Open the prompt by fixing the frame: the governing jurisdiction, who reads the output, and what kind of document this is. For the example: "Governing jurisdiction is [X]. This is a clause in a business-to-business vendor agreement. The reader is the counterparty's legal team."
Supply the structure
Give the model the container. If the agreement uses defined terms and a numbering scheme, provide them and instruct the model to follow them exactly. This keeps the draft inside the document's conventions instead of inventing its own that a reviewer must rip out. The reasoning behind these context controls is detailed in Everything That Matters When You Prompt for Legal Writing.
Step Three: Draft With Strict Grounding Rules
Paste the source and constrain the model
Now paste the governing text and add the grounding instructions: work only from the provided material, quote the controlling language before applying it, and flag any point the provided text does not cover rather than filling it in. Ask for the retention clause to be drafted within that constraint.
- "Use only the text I supplied; do not rely on outside knowledge."
- "Quote the controlling provision before drafting from it."
- "Where the provided text does not resolve a question, mark it [GAP] for human input."
Read the gaps first
When the draft comes back, read the flagged gaps before the prose. Those marks are the model telling you where it could not stand on solid ground. They are the most valuable part of the output because they direct the reviewer's attention precisely.
Step Four: Run a Self-Critique Pass
Make the model attack its own draft
In a follow-up prompt, instruct the model to identify the three riskiest assumptions in its draft, any claim it cannot support from the provided text, and the specific places a reviewer should scrutinize. This converts the model into its own first reviewer and surfaces issues before a human spends time on them.
Check term discipline
Ask the model to confirm it used every defined term exactly as defined and introduced no synonyms for them. Stray synonyms for defined terms create ambiguity, and this targeted check catches a common, subtle error in legal drafts.
Step Five: Tighten Language and Form
Enforce precision
Review the draft for the words that carry weight: must versus should, shall versus may, including versus including without limitation. If the model softened an obligation or scoped a term loosely, prompt for a tightened version with the precise operative language. Precision is the product here, so this step is not optional polish.
Match plain-language requirements
If the document has a plain-language or conspicuousness requirement, prompt explicitly for it and check the result against that standard. The model's default register runs dense, so plain language usually needs to be requested directly. The common failures to watch are listed in Seven Prompting Habits That Sink Legal and Compliance Drafts.
Step Six: Hand Off to a Qualified Reviewer
Package the draft for review
Deliver the draft together with the flagged gaps, the model's self-critique, and the source text you grounded it in. A reviewer who sees the inputs and the model's own risk flags can verify far faster than one handed only prose. This is the mandatory gate; nothing ships without it.
Record the trail
Save the prompts, sources, and edits. For compliance work especially, a clean record of how the document was produced is worth keeping. It also lets you refine your prompts the next time a similar task comes around. A full narrative of this sequence in action appears in Inside One Compliance Team That Rebuilt Drafting Around Prompts.
Turning the Sequence Into a Reusable Template
Capture what stays constant
Once you have run the sequence a few times, the constant parts become obvious: the grounding instructions, the gap-flag convention, the self-critique request, and the term-discipline check rarely change. Lift those into a fixed template so you are not retyping them under deadline. What varies, the source text, jurisdiction, reader, and structure, becomes a short set of fields you fill in per task.
- Keep the grounding and gap-flag instructions identical across runs.
- Turn jurisdiction, reader, and document type into fill-in fields.
- Store the template where the whole team can reach it.
Why templating reduces errors
A template removes the chance that someone forgets the grounding rule or the jurisdiction line on a busy day. The safeguards happen by default rather than depending on memory. This is the same reasoning that makes the workflow hand-off-able: discipline that lives in a template survives turnover and deadlines in a way discipline living in one person's habits does not.
Adapting the template per document type
A clause, a policy, and a regulatory response share the core safeguards but differ in structure. Maintain small variants of the template for each common document type, each pre-loaded with the right skeleton and defined-term conventions. The shared grounding core stays identical; only the structural scaffolding changes between them.
Frequently Asked Questions
What if I do not have the governing text to paste?
Then you stop and get it before drafting. Prompting without the source invites the model to fabricate the very authority your document depends on. Gathering the controlling text is step one precisely because everything downstream relies on it.
How many times should I iterate with the model?
Usually a draft pass, a self-critique pass, and one tightening pass. More iterations help only if each has a specific instruction. Aimless "make it better" rounds tend to drift and introduce new problems, so give every pass a concrete job.
Why quote the controlling language before applying it?
Quoting forces the model to anchor its draft to real provided text rather than its memory. It also gives the reviewer an immediate way to check whether the application matches the source. The quote is a small step that prevents a large class of grounding errors.
Can I automate this sequence?
You can template the prompts and the structure, but keep the human review gate manual for anything that ships. The drafting and self-critique steps automate well; the judgment about whether the result is correct and safe stays with a qualified person.
What does a good gap flag look like?
A specific note like "[GAP] The provided regulation does not state a maximum retention period; a human must supply the client's required period." It names exactly what is missing and what input resolves it. Vague flags are nearly useless; specific ones direct the reviewer straight to the decision.
How do I keep the model from softening obligations?
Call it out in the tightening step. Ask the model to use precise operative language and to preserve must, shall, and similar terms exactly. Then read the draft yourself for softened obligations, because the model's instinct toward smoother prose can quietly weaken a requirement.
Key Takeaways
- Gather the governing authority and confirm you can use it before opening the model.
- Set jurisdiction, reader, document type, and structure in the prompt so the draft starts in the right frame.
- Draft with strict grounding rules and read the model's flagged gaps before the prose.
- Run a self-critique pass that surfaces risky assumptions and confirms defined-term discipline.
- Tighten the operative language by hand, since precision is the product and the model tends to soften it.
- Hand a packaged draft with gaps and self-critique to a qualified reviewer, and record the trail.