OVERVIEW
Moving to Claude Sonnet: CA-V2 and LA-V2 will be removed at the end of July. Any prompts remaining on these models will be moved to Claude Sonnet, which is being updated to Anthropic’s newest frontier and most powerful model, Claude Sonnet 5.
Testing: Prompts were originally tested with Claude Sonnet 4.6 and it was found that this was the most stable model for most prompts. We then tested the new model Claude Sonnet 5, which won 7 of 9 head-to-head prompt tests, including complex prompts.
Legal research: For complex prompts involving legal research, testing indicated Claude Opus or GPT Thinking are also good agentic models.
Prompt Templates: Testing indicates that the existing prompt templates work well with the agentic models; no rewrite needed. New prompts can follow the same template headings but can be simplified and do not require added persona or jurisdiction context.
TESTING AND FEEDBACK ON PROMPTS FROM VERTICAL TEAMS
TESTING AND FEEDBACK ON PROMPTS FROM VERTICAL TEAMS
The legal innovation team ran structured head-to-head testing across nine types of prompts spanning the Prompts Library: affidavits, basic and complex letters, basic and complex summaries, file notes, subpoenas, letters with legal research, and legal documents with legal research. Testing compared GPT Instant 5.5, GPT Thinking 5.5, Claude Sonnet 4.6, Claude Opus 4.8 and then Claude Sonnet 5, predominantly across Australian family and criminal law test matters.
Claude Sonnet 4.6 was initially found to be the most stable model for prompts.
We then tested the new Claude Sonnet 5, which won 7 of the 9 categories and performed best overall across the prompt set, including affidavits, basic and complex letters, basic and complex summaries, file notes, subpoenas, letters with legal research, and legal documents with legal research.
Some notes on Claude Sonnet 5:
It was generally accurate, disciplined in avoiding fabricated content and reliable in applying template requirements.
As an agentic model, it also recognised requirements in a template and applied them itself, where some older models suffered template leak or confusion.
It is more cautious than other agentic models, so you may see more placeholders where it is uncertain about the information. This is a good thing as being accurate is more important than fully populated details that are not entirely correct. Where there are placeholders, users can ask the AI to prepare a checklist or correspondence to collate the missing information.
Two categories went to other models. GPT Thinking 5.5 won Complex Summaries for its more nuanced treatment of inconsistencies, and Claude Opus 4.8 won Legal Documents with Legal Research for stronger accuracy and case-law citations. For prompts involving substantial legal research or nuanced inconsistency analysis, you can therefore test and consider defaulting to Claude Opus or GPT Thinking.
Vertical feedback received so far is consistent with these results and supports the update to Claude Sonnet as the default: it is the stronger choice across most legal workflow prompts.
On the strength of the above results, we confirmed that prompts should be defaulted to Claude Sonnet 5.
WHERE TO FOCUS YOUR PROMPT EFFORT
WHERE TO FOCUS YOUR PROMPT EFFORT
Product framing and the purpose of the prompt library
Prompt Library and Matter AI (All and Matter Tabs) serve different purposes and should not be seen as competing. Matter AI (All and Matter Tabs) is best suited to exploratory, analytical, basic drafting and ad hoc legal tasks where users want flexibility. Prompt Library is best suited to repeatable legal workflows where quality, consistency, structure, and legal drafting standards matter.
The Prompt Library can become the source of trusted legal workflows that Matter AI (All and Matter Tabs) may not reliably reproduce on demand. The competitive advantage is not the existence of a prompt itself, but the legal expertise, safeguards, and consistency embedded within it.
Not all prompts carry equal long-term value. The question is not whether prompts can be replaced by Matter AI (All and Matter Tabs); it is which legal tasks benefit from an open-ended AI conversation, and which require structured legal workflows. Prompts broadly fall into two categories.
1. Analytical and basic drafting prompts (low priority)
These prompts ask the AI to summarise, analyse, or produce straightforward content without a prescribed legal structure. Examples include:
Summarising correspondence or offers
Identifying key issues from documents
Preparing matter summaries
Extracting chronology events
Comparing or analysing documents
Drafting basic letters or other straightforward content
For these tasks, the value comes mainly from the AI's reasoning capability. An experienced user can often obtain a strong result from a short instruction in Matter AI (All and Matter Tabs), so these tasks are less likely to justify extensive prompt engineering.
2. Content-led legal workflow prompts (high priority)
These prompts require the AI to produce a legal document that follows a recognised legal structure, drafting conventions, court or practice requirements, specific wording conventions, logical sequencing, and embedded safeguards and instructions. Examples include:
Affidavits
Subpoenas
Complex legal correspondence with a prescribed structure
File notes with required sections
Costs notices
Procedural documents
The value of these prompts is not simply that they tell the AI what to write; it lies in the prompt engineering that controls how the document is produced. They encode legal expertise most users would not know to include in an ad hoc instruction, ensuring consistency, completeness and quality while reducing the risk of omitted information, incorrect structure, or poor drafting.
Accordingly, the more a prompt encodes legal expertise and a repeatable legal workflow, the more enduring its value.
YOUR PROMPT TEMPLATES
YOUR PROMPT TEMPLATES
You do not need to update your existing prompt templates. Retain them unless testing identifies a specific issue. Claude Sonnet 5 is an agentic model: it reviews the template and picks up requirements automatically, so retained templates will continue to work.
Going forward, new prompts do not need to be as detailed. You do not need to include persona information in the prompt (e.g. "you are a lawyer") or jurisdiction and matter context (e.g. "this is a family law matter in Australia"), as the system prompt already provides this. Focus detail on legal structure, consistency and safeguards.
Prompt Structure
When drafting a new content-led legal workflow prompt for the Prompts Library, use the following five-part structure:
Title: A short, descriptive name that identifies the document type
Purpose: A one or two sentence statement of what the prompt produces and when it should be used.
Clarifying Questions (previously referred to as Preliminary Questions): Any questions the AI should ask the user before drafting, where information is missing or ambiguous, to ensure the output is accurate and complete. Note that: agentic AI models may automatically ask clarifying questions of their own accord where they identify a gap. However, if you definitely want a particular question asked — for example, to route the response to a specific part of the template — this should be explicitly included in the prompt rather than left to the model to identify on its own.
Template: The structure, headings, sequencing and drafting conventions the output must follow, reflecting the required legal or court structure.
Rules: The safeguards and instructions that control the AI's behaviour — for example, restrictions on fabricated content, verification requirements, tone, or matters that must always or never be included. In the system prompt we provide the matter context, so you should not need to confirm language or currency conventions.
WHAT YOU NEED TO DO FOR PROMPTS
WHAT YOU NEED TO DO FOR PROMPTS
You do not need to do anything to your existing prompts for this update to take effect, and they will continue to work as they are.
When creating new prompts, we recommend considering:
Whether the prompt is necessary or could be run by the lawyer in Matter AI (All and Matter Tabs)
On 31 July 2026, CA-V2 and LA-V2 will be removed and all prompts remaining on those models will be moved to Claude Sonnet, which is being updated to the Claude Sonnet 5 model. If you are happy for your prompts to default to Claude Sonnet 5 then you do not need to do anything. If you would prefer for your prompts to default to a different agentic model (e.g. GPT Instant, GPT Thinking or Claude Opus) then you need to change the prompt default to your preferred model manually. As usual, users can manually move between AI models when running the prompts.
Whether the prompt template can be simplified, as set out above.
FEEDBACK ON PROMPTS
FEEDBACK ON PROMPTS
Please continue to raise anything you find through the Internal Portal (https://provider.docorto.ai):
Bugs or technical issues: log a case in the Internal Portal and attach debug information, any screenshots, prompt text, and output where possible so the issue can be reproduced.
General feedback or suggestions: submit via the feedback option in the Internal Portal, rather than replying to this update, so all input is captured and tracked in one place.
Using the Internal Portal for both bugs and feedback (rather than email or Slack) ensures issues are logged, prioritised, and actioned consistently across the legal innovation and product teams.
NEXT MATTER AI UPDATE
The next Matter AI session will be held in the week of 24 August 2026, covering updates to Matter AI (All and Matter Tabs) and the Prompt Library. Further details, including exact date and time by region, will follow by calendar invite.