Playbooks
Set Up Claude Projects for a Matter: Instructions, Files and Memory
Set up a Claude Project for a legal matter in 30 minutes: sample instruction text, which files to upload, Skills, memory, Team sharing and three prompt flows.
On this page
A Claude Project for a matter is a workspace with three parts: instruction text that fixes role, jurisdiction, citation rules and confidentiality; a curated set of matter files as project knowledge; and chats that inherit both. Set it up once, in about thirty minutes, on a Team or Enterprise plan that does not train on your data, and every later prompt starts from the same brief.
Step zero: the plan and data-policy check
Nothing about a matter goes into a Project until the plan is right. Anthropic's consumer plans (Free, Pro at $20 a month, Max at $100 or $200) have required users since 28 August 2025 to choose whether chats train future models; opted-in data is retained for five years, opted-out data for 30 days (Anthropic consumer terms update). Team ($25 per standard seat, $125 premium, five-seat minimum) and Enterprise are never used for training (Anthropic privacy centre).
| Plan (September 2026) | Price | Trains on chats? | Client files in a Project? |
|---|---|---|---|
| Free | $0 | Only if the training toggle is on; Fable 5.1 not available | No: consumer terms |
| Pro | $20/month ($17 annual) | Only if the toggle is on | No: consumer terms, see Heppner below |
| Max | $100 or $200/month | Only if the toggle is on | No: consumer terms |
| Team | $25 standard / $125 premium per seat, 5-seat minimum | Never | Yes, with a signed data processing agreement; chat is not ZDR-eligible |
| Enterprise | Custom | Never | Yes; Compliance API and local session transcripts (Cowork, Claude Code) are retained six years by default |
The reason the consumer rows say no is not only training. In United States v. Heppner (S.D.N.Y., bench ruling 10 February 2026, opinion 17 February 2026) Judge Rakoff held that a defendant's research memos generated with consumer Claude were neither privileged nor work product: "Claude is not an attorney", and the terms allow logging and disclosure. Privilege never attached, although the court left open that it might where counsel directs the AI use. The UK Upper Tribunal reached a similar practical conclusion in UK v SSHD [2026] UKUT 81 (IAC): pasting client letters into public AI tools puts them "in the public domain". Germany's BRAK guidance treats tools that store or forward client data for training as a breach of § 43a Abs. 2 BRAO, and ÖRAK's four Austrian practice rules include EU servers and no client data for training. The fuller analysis is in attorney–client privilege and AI and the confidentiality checklist.
Two practical notes. If a personal Pro account is all you have, the training toggle is under your name, then Settings, then Privacy ("Help improve Claude"); the walkthrough is in AI privacy settings step by step. And API users should know that Fable 5.1 and Mythos are "Covered Models" that require 30-day retention and are not eligible for zero data retention without express authorisation (Anthropic data retention docs); Opus 5 and Sonnet 5 are not so restricted.
Write the instruction text
Project instructions are a standing system prompt for every chat in the workspace. The five-part formula that dominates legal prompting guides in 2026, role and context, task, sources, output format, constraints, maps directly onto it. What follows is a complete sample for a contentious commercial matter; adapt the bracketed parts.
ROLE AND MATTER
You are a senior associate supporting [Firm] on [Matter name, file no.].
We act for [Client], the [claimant/defendant], against [Counterparty].
The supervising lawyer is [Name]. Everything you produce is a draft for
that lawyer's review, never a final document.
JURISDICTION AND LANGUAGE
Governing law: [Austrian law / the laws of England and Wales / New York
law]. Forum: [court or tribunal]. Write in [British English / German].
Do not import concepts from other jurisdictions without saying so.
SOURCES, IN ORDER OF PRIORITY
1. Files in this Project's knowledge. Cite them by file name and page or
paragraph, e.g. [2025-03-14_SPA_signed.pdf, cl. 9.2].
2. Authorities in the folder "authorities-verified". These have been
read and checked by a lawyer.
3. Nothing else. If a point needs a source that is not in 1 or 2, write
[NO SOURCE IN PROJECT] and continue.
CITATION RULES
- Never cite a case, statute, article or URL that is not in the Project.
If you believe one exists, describe it in words and tag it [TO SEARCH].
- Quote only text you can see in a file, verbatim, with file and page.
Paraphrase everything else without quotation marks.
- Do not add a closing "verification" or "I have double-checked" passage.
Verification happens outside this tool.
FACTS
- Distinguish what a file says from what you infer. Tag inferences
[INFERRED] and gaps [NOT IN FILE]. Do not fill gaps with plausible
detail.
- Where two files conflict, say so and cite both.
CONFIDENTIALITY AND OUTPUT
- Use party names exactly as they appear in the files; do not search
for or speculate about individuals.
- Flag anything that looks like a deadline, a limitation issue, a
conflict of interest or a privilege question at the top of the reply.
- Default format: headings, numbered paragraphs, tables for comparisons,
no preamble. Default length: as short as the task allows; state the
word count for anything over 800 words.
- If a task is ambiguous, ask one clarifying question before drafting.
Three of those lines earn their place because of documented failures. The citation ban exists because even Anthropic's own counsel apologised on 15 May 2025 after Claude, asked only to format a citation, kept the link but produced the wrong title and authors. The [INFERRED] and [NOT IN FILE] tags exist because models fill gaps by design, and an Austrian, German or English court will not distinguish a confidently invented date from a lie; the sanctions record runs from Mata v. Avianca (S.D.N.Y. 2023) to the Austrian OGH's dismissal of an AI-drafted appeal in 14 Os 95/25i (7 October 2025). And the instruction not to add a verification passage is new. Anthropic's release notes for Claude Opus 5 (24 July 2026) say the model "verifies its own work without being told to" (What's new in Opus 5); prompts written for 2025 models that demand a "final verification step" now produce reassurance paragraphs that look like checking and are not. Real verification follows the citation-checking protocol.
Choose the files
Project knowledge is static: what you upload is what every chat sees. Capacity tracks the model's context window, which is one million tokens for Opus 5, Sonnet 5 and Fable 5.1, roughly 555,000 words by Anthropic's estimate (models overview). That is the ceiling, not the target. A Project stuffed with a data room answers slowly, costs more per reply on metered plans and dilutes attention across documents that do not matter to the question. Fifteen to twenty curated files is the usual sweet spot for a contentious matter; bulk review belongs in Claude Cowork or a DMS connector.
| Upload | Leave out | Why |
|---|---|---|
| Pleadings, key contracts, the engagement letter's scope paragraph, your own precedent memos, the firm style guide | The entire data room, email archives, working drafts with tracked changes you have not accepted | Attention and cost; a chronology from 20 documents is checkable, from 2,000 it is not |
| Verified authorities as PDFs, in a folder named authorities-verified | Any authority nobody has opened in a database | The citation rules above only work if the folder is clean |
| OCR-processed text of scanned documents | Raw scans | Claude reads scans by vision, and one 2026 comparison reports it is likelier to say "illegible" than invent a figure, but a dedicated OCR pass first is the safer route for financial pages (CounselPro) |
| Documents the engagement covers | Third-party personal data beyond what the task needs; anything from a matter behind an ethical wall | GDPR minimisation; a Project does not enforce information barriers |
Name files so that dates sort: 2025-03-14_SPA_signed.pdf, 2025-11-02_letter_counterparty_breach.pdf. The chronology flow below depends on it. Anthropic's help centre lists the current per-file and per-Project ceilings; check it before uploading anything unusual, because those limits change more often than this page will.
Skills, plugins and memory: what each adds
Projects are one of four context mechanisms, and the setup is stronger when you know which does what.
| Mechanism | What it holds | When it applies | Best legal use |
|---|---|---|---|
| Project | Instructions plus static uploaded files | Every chat inside the Project | Matter context, matter rules |
| Skill | A folder with SKILL.md (YAML frontmatter plus instructions), optional scripts, references and assets | Only when a prompt matches the Skill's trigger; works across Claude.ai, Claude Code and the API | House citation style, clause-numbering conventions, a redlining procedure |
| Claude for Legal plugin | Open-source practice-area bundles; 12 plugins as of 12 May 2026, 150+ skills and 20+ connectors as of 18 September 2026 | After a 10–20 minute cold-start interview that writes a practice profile | Tabular diligence, DPA review, claim charts, with [verify] tags on unchecked citations |
| Memory | Summaries Claude carries between chats, where your plan offers it | Across chats, according to your settings | Personal working preferences, not matter facts |
Skills are described in Layer 3 Labs' guide of 7 September 2026, which also makes the key contrast: Projects are "static uploaded context in one workspace", Skills are triggered procedures (Layer 3 Labs). The plugins come from Anthropic's Apache-licensed repository, where every output is described as "a draft for attorney review" and any citation not checked against a connected research tool is tagged [verify] (Claude for Legal on GitHub). GC AI's critique of the plugins listed three gaps: context that resets per session, no citation verification without a connector, and single-user markdown instead of team playbooks (GC AI). A Project closes the first gap for chat-based work; the second is closed only by a database and a lawyer; the third is why the instruction text above lives in the firm's precedent system, not on one person's laptop. The distinction between Skills and plugins is set out in the glossary entry and the plugin line-up in Claude for Legal plugins explained.
Memory deserves one rule. Whatever memory options your plan exposes as of September 2026, matter facts do not belong in them: memory is cross-matter by design, and a preference learned on one client's file surfacing in another's draft is the kind of contamination an ethical wall exists to prevent. Keep matter facts in the Project's files and instructions, where they are scoped, and use memory for how you like your output formatted.
Sharing on Team and Enterprise
A shared Project is a small case team. Set it up as one.
- The supervising lawyer owns the Project; associates and paralegals are members. Ownership is who answers for the instruction text.
- Membership equals the matter team, checked against the conflicts system before anyone is added. A Project has no concept of an information barrier; the connectors do. iManage's MCP server, released 14 May 2026, is permission-bound and respects ethical walls, and NetDocuments' ndMAX runs under the user's own credentials so documents "never leave the platform". Where the firm has either, connect it rather than uploading: the setup is in connecting Claude to iManage or NetDocuments with MCP.
- Name Projects with the matter number first, so a partner with forty of them can find the right one.
- Tell the team that chats are supervisable. On Enterprise, the Compliance API and local session transcripts are retained for six years by default; that is what allows the supervision that ABA Formal Opinion 512 and the SRA's August 2026 warning notice both expect, and it is also why nobody should treat a Project chat as a private notebook.
- Remove leavers the day they leave, and archive the Project when the matter closes, with the verification tables and research logs exported to the file.
The pay-off of treating the Project as a team is what Quinn Emanuel's Christopher Kercher described to Fortune in May 2026 as "treating Claude like a case team member", fed chronological materials rather than one-off questions (Fortune, 12 May 2026). Freshfields, which rolled Claude out to 5,700 staff across 33 offices, reported roughly 500% usage growth in six weeks on the same principle.
Three prompt flows that use the setup
Each of these assumes the instruction text and files above. Because the rules are already in the Project, the prompts are short.
FLOW 1 — CHRONOLOGY
Build a chronology from every file in project knowledge. One row per
event: date | event | source (file, page/para) | actors | open question.
Sort by date. Tag inferred dates [INFERRED]. Where sources conflict, add
a row for each version and mark [CONFLICT]. End with a list of dates
that look like deadlines or limitation triggers.
FLOW 2 — ISSUE LIST
From the chronology and the pleadings, list the legal issues in dispute.
For each: issue | facts from the files (with sources) | our position |
counterparty's likely position | evidence we do not yet have | the kind
of authority needed, described in words and tagged [TO SEARCH].
Do not cite any case, statute or rule.
FLOW 3 — FIRST DRAFT
Draft [the letter before action / section 3 of the defence / the
memorandum on limitation] for [supervising lawyer]'s review. Use only
the authorities in authorities-verified, cited by short name and
pinpoint. Tag any proposition without support [NO SOURCE IN PROJECT].
Put a bullet list of every file relied on at the end.
Flow 2's [TO SEARCH] tags become the research task; the authorities you then read and verify go into the authorities-verified folder; Flow 3 draws only on that folder. The loop is the point: at no stage does an unverified authority reach a draft, and the verification protocol runs once, on the folder, rather than after every chat. The instruction text belongs in the firm's prompt library as a template.
A setup that works, in order
- Confirm the plan: Team or Enterprise, with the data processing agreement signed and the model choice made (Opus 5 is the default on Max; Sonnet 5 the default on Free and Pro; Fable 5.1 is not available on Free).
- Check the engagement letter for any AI restriction and, where your bar requires it, the client disclosure position described in AI disclosure to clients.
- Create the Project, named matter-number first.
- Paste the instruction text, adapted, and have the supervising lawyer read it once.
- Upload the curated files with dated names, OCR done, authorities-verified folder empty for now.
- Run Flow 1 and read the chronology against two source documents to calibrate how the model reads your files.
- Run Flow 2; take the [TO SEARCH] list into Westlaw, Lexis, RIS or juris; verify; upload the PDFs to authorities-verified.
- Run Flow 3. Review in tracked changes. Export the verification table to the matter file.
- Add the matter team, brief them on the tags, and archive the Project at closing.
Module 2 of the AI Lab for Lawyers course builds a matter Project like this one live, from the instruction text to the first chronology. The setup is not sophisticated; its value is that the rules are written once, by the lawyer who answers for them, and then apply to every prompt a junior types at eleven at night.
Frequently asked questions
Is it safe to upload client documents to Claude Projects?
On Team or Enterprise plans, where Anthropic states data is never used for training, and subject to your own engagement terms and bar rules, yes with ordinary care. On Free, Pro or Max, no: those are consumer plans where chats may train the model unless you have opted out, and a federal court in Manhattan held in February 2026 that a defendant's consumer-Claude research memos were not privileged.
What should go in Claude Project instructions for legal work?
Five blocks: who Claude is acting for and on which matter; the governing jurisdiction and language; the hierarchy of sources it may rely on; citation rules, above all a ban on any authority not in the uploaded files; and confidentiality and output rules. Keep it under 400 words. Longer instruction text gets followed less consistently, not more.
How many files can I put in a Claude Project?
Capacity tracks the model's context window, which is one million tokens on Claude Opus 5, Sonnet 5 and Fable 5.1, roughly 555,000 words by Anthropic's estimate, but a Project that fills its capacity answers more slowly and less precisely. For a matter, fifteen to twenty curated files beat a whole data room. Use a document-management connector or Cowork for bulk review.
What is the difference between a Claude Project and a Skill?
A Project is a workspace with fixed instructions and uploaded knowledge that every chat inside it inherits. A Skill is a folder with a SKILL.md file and optional scripts and references that Claude loads only when a prompt matches its trigger, and it works across Claude.ai, Claude Code and the API. Projects hold matter context; Skills hold reusable procedures such as a house citation style.
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.
Related reading
Tools
Claude for Lawyers: An Honest Review of Anthropic's Models for Legal Work
Claude for lawyers, reviewed as of September 2026: Fable 5.1, Opus 5 and Sonnet 5, plan prices, data retention by plan, legal plugins, benchmarks, weaknesses.
9 min read
Tools
Claude for Legal Plugins Explained: The 12 Practice-Area Plugins and 20+ Connectors
Claude for Legal gives paid Claude users 12 open-source practice-area plugins, 150+ skills and 20+ connectors to iManage, NetDocuments and CoCounsel.
8 min read
Guides
Which AI Tools Train on Your Data? A Confidentiality Checklist for Lawyers
Which AI tools train on your data, plan by plan as of September 2026: ChatGPT, Claude, Gemini, Copilot, legal vendors, plus bar rules and a 10-point checklist.
11 min read
Regulation & Ethics
Attorney–Client Privilege and AI: Heppner, Warner, the NYT Logs and Public Chatbots
Consumer chatbot memos are not privileged. US v. Heppner, Warner v. Gilbarco and the 20-million-log OpenAI order, plus rules that keep privilege intact.
10 min read
Playbooks
How to Verify AI Legal Research: A Citation-Checking Protocol
Verify AI legal research with a five-check protocol: open every authority, match the quote, confirm holding and status, log it. Template, time budgets, tools.
11 min read
Glossary
Skills and Plugins (AI) — Definition for Lawyers
A skill is a packaged, reusable instruction an AI tool loads for one task; a plugin bundles skills with tools, connectors and automations. How they differ, and governance.
3 min read