AI Agents for Lawyers

Glossary

System Prompt — Definition for Lawyers

A system prompt is the standing instruction an AI model reads before your message. Where it lives in Claude, ChatGPT and Skills, a legal template, and its limits.

By Dr. Niklas SchmidtUpdated 13 November 20264 min read
On this page
  1. Where the system prompt lives in the tools you already use
  2. A system prompt for a legal drafting assistant
  3. What a system prompt cannot do
  4. What to do this week

A system prompt is the standing instruction a language model reads before it sees your message: who it is, what it may do, in what format, under what constraints. In legal tools it lives in Claude Projects, custom GPTs, Skills files and firm platforms. It shapes behaviour reliably; it does not guarantee it, and it changes nothing about where your data goes.

Where the system prompt lives in the tools you already use

Every chat product has one whether or not you can see it. The vendor's own system prompt sets the model's default manners and refusals; anything you add sits on top of it and cannot override it. What you can control goes by different names.

ProductWhere you write itScopeNotes (as of September 2026)
ChatGPT custom instructionsSettings, under personalisationEvery new chat on the accountAccount-wide; easy to forget it is there
Custom GPTThe Instructions field in the GPT builderEveryone who uses that GPTThe builder edits it; users do not see it
Claude ProjectProject instructions plus uploaded filesEvery chat inside the ProjectUploaded files are static context, not instructions
Claude SkillA SKILL.md file with YAML frontmatter and instructions, optionally scripts and reference filesFires only when a prompt matches its trigger; works across Claude.ai, Claude Code and the APIBest for one repeatable task, not a whole practice
Claude for Legal pluginA CLAUDE.md practice profile written by a 10–20 minute cold-start interviewRead by every other skill in the pluginContext resets per session; upkeep is manual
API callThe system parameterThat request onlyWhat every vendor product uses underneath

The Skill and Project rows come from the layer3labs guide to Claude Skills for law firms (updated 7 Sep 2026): Skills are triggered, Projects are static uploaded context in one workspace (layer3labs). The plugin row comes from Anthropic's open-source Claude for Legal repository and from GC AI's review of it, which lists context resets with manual CLAUDE.md upkeep, no citation verification without a connector, and single-user markdown instead of team playbooks as the three gaps (GC AI, Jun 2026). The practical difference shows up when you set up a matter: an instruction to always cite the clause number applies to every chat in a Project, while a Skill applies only when its trigger fires. The Claude Projects playbook and the custom GPT playbook walk through both, and the Skills and plugins entry covers the file format.

The five-part structure that dominated legal prompt guides in 2026, role and context, task, sources, output format, constraints, works as well for a standing instruction as for a single prompt. Two refinements from the same guides: give positive examples rather than lists of prohibitions, and set hard limits on length (Legalfly, 17 Sep 2026).

You are a drafting assistant for a commercial practice advising under Austrian
and English law. Work only from the documents and authorities in this Project.
If a point needs authority that is not in the Project, write "authority needed"
and stop; do not supply case citations or statutory references from memory.

Output for every drafting request: the original clause quoted in full, then the
proposed replacement, then a comment of at most 60 words explaining the change.
Use British English. Where the agreement defines a term, use the defined term.

At the end, list every clause touched that has a governing-law, limitation-of-
liability or indemnity effect, in the order the clauses appear.

If a request is ambiguous, ask one clarifying question before drafting.

The "authority needed" line follows the safe-research pattern the legal prompting guides settled on in 2026: the model frames issues and drafts, the research database supplies the authorities, and the model then works "only with the authorities I have provided" (The Legal Prompts, 18 Feb 2026). The prompt engineering guide explains why each of the five parts is there.

What a system prompt cannot do

Four limits matter more than the rest.

It does not change where your data goes. Retention and training are set by the plan and the vendor's terms; consumer tiers train unless you opt out, business and API tiers do not by default. A line reading "do not store this conversation" has no effect on either.

It can be overridden by content. Text inside a document, web page or email competes with your instruction, and Zenity's researchers put it plainly in 2026: the agent "can't reliably tell the difference between content it was asked to read and hidden instructions" (Dark Reading, 2026). The prompt injection entry covers what that means for contract review.

It is not a verification step. Claude for Legal tags any citation not checked against a connected research tool as [verify] (Claude for Legal); the tag is a convention the prompt asks for, not a check the prompt performs. GC AI lists the missing verification as a gap for the same reason.

It decays. Long conversations dilute it, new sessions reset it, and model updates change how literally it is read. An instruction that behaved on one model version has not been tested on the next.

What to do this week

Write one standing instruction for your most repeated task and put it in a Project or custom GPT instead of retyping it. Include the "authority needed" stop rule verbatim. Test it on three documents you know well, one of them with a deliberately misleading footnote, and read what comes back before you trust it on a live matter. Module 1 of the AI Lab for Lawyers course covers prompting fundamentals, including how to write and test standing instructions for a practice.

Frequently asked questions

Is a system prompt the same as custom instructions in ChatGPT?

In effect, yes. ChatGPT's custom instructions, a custom GPT's Instructions field, Claude's Project instructions and the system parameter of an API call all do the same job: text the model reads before every message. The differences are scope and persistence. Custom instructions apply account-wide; a custom GPT or Project applies only inside that workspace; an API system prompt applies to one request.

Can a system prompt stop an AI tool from hallucinating case law?

It can reduce the occasions. An instruction to work only from supplied authorities and to write 'authority needed' instead of citing from memory changes behaviour in most cases. It does not verify anything. Anthropic's own legal plugins tag citations not checked against a connected research tool as [verify], and reviewers list the absence of built-in verification as a gap. Read every citation yourself.

Does a system prompt make an AI tool confidential?

No. Whether your prompts are retained or used for training is set by the plan and the vendor's terms, not by anything you write in the prompt. Consumer tiers of ChatGPT and Claude train unless you opt out; Team, Enterprise and API tiers do not by default. A line saying 'do not store this conversation' has no legal or technical effect.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion