Set up Claude Projects for workers compensation practice

What this sets up

A Claude Project is a workspace with two persistent parts: project instructions that apply to every chat inside it, and project knowledge, a file space where, in Anthropic's words, "Anything you upload to this space will be used across all of your chats within that project". This configuration turns that structure into a workers compensation drafting assistant: the instructions below carry the de-identification rules and drafting discipline, the file pack carries your templates and legislation map, and the first three prompts put it to work on determination briefs, return to work summaries and file notes. The method comes from the AI Setup module LM-S01; this page packages the WC playbook for Claude specifically.

What never goes in

No real claimant data ever enters the platform: no names, no claim numbers, no dates of birth. De-identify at the source with the placeholders [CLAIMANT_NAME], [CLAIM_NUMBER] and [DATE_OF_BIRTH], and re-identify only inside your controlled claims system after human review. That rule is absolute regardless of plan. The platform's own documentation adds these reasons to hold the line:

• On consumer plans (Free, Pro and Max), Anthropic's Privacy Policy states: "We may use your Inputs and Outputs to train and improve Anthropic AI models, unless you opt out through your account settings." The opt-out is not absolute: "Even if you opt-out, we will use Inputs and Outputs for model improvement when: (i) your conversations are flagged for safety review... or (ii) you've explicitly reported the materials to us".
• Where training is enabled, Anthropic "may retain your data in a de-identified format for up to 5 years"; chats flagged by automated trust and safety systems are retained "for up to 2 years" with classification scores "for up to 7 years". Deleted conversations are removed from back-end systems "within 30 days", not instantly.
• On Claude for Work (Team and Enterprise): "By default, we will not use your inputs or outputs from our commercial products... to train our models." The stated exception is material you explicitly submit through feedback or bug reports, so do not use thumbs up or down on any chat containing claim material.
• Anthropic's documentation does not state whether project knowledge files are covered by consumer training in either direction, so treat every upload as if it were in scope.
• On Team and Enterprise plans a shared project exposes everything: members with view access "can see project contents, knowledge, and instructions". Never put anything in project knowledge you would not show every colleague the project could be shared with.
• Anthropic's own advice is blunt: "we encourage our users not to use our products and services to process personal data."

Sources, all verified 14 August 2026: Anthropic Privacy Policy (effective 8 July 2026), Is my data used for model training?, How long do you store my data?, commercial products training default, personal data in model training, What are Projects?.

The instruction text

Create a project, open its settings, and paste this into the project instructions field:

You support a workers compensation practitioner. Context files in
this space: writing-style.md, wc-legislation-map.md,
determination-brief-template.md, rtw-plan-summary-template.md,
de-identification-checklist.md, glossary.md.

Hard rules:
1. All inputs are de-identified. Expect and preserve the
   placeholders [CLAIMANT_NAME], [CLAIM_NUMBER] and
   [DATE_OF_BIRTH]. If I ever paste material containing a real
   name, claim number or date of birth, stop immediately and tell
   me to de-identify before you do anything else.
2. You draft; you never decide. Frame every output as material for
   my assessment, not as a determination. Use deterministic
   language in drafts of decision documents and never present an
   outcome as if it were yours to make.
3. Every reference to legislation must name the specific section.
   If you are not certain a section applies, say so and list what
   needs checking.
4. Follow writing-style.md for tone and structure. Determination
   briefs follow determination-brief-template.md exactly.
5. End every draft with "For human review:" and the highest-risk
   points a delegate must verify against the claim file.

The files to attach

Write these six markdown files and upload them to project knowledge. Each links to a published artefact to build it from:

• writing-style.md: your voice profile. Build it with the voice-profile interview in module LM-S01, Part 1.
• wc-legislation-map.md: the sections you use most, each with a one-line plain-language summary and the tests it sets. Start from the scheme architecture in module LM-W01.
• determination-brief-template.md: your team's brief structure, headings and evidence table. The determination drafting recipe shows the shape, and is downloadable as raw markdown.
• rtw-plan-summary-template.md: the fields a return to work summary must cover, informed by recovery at work and suitable duties.
• de-identification-checklist.md: what gets replaced, with the placeholder for each item, built from the de-identification toolkit.
• glossary.md: scheme terminology and your organisation's preferred usage; the site glossary's SRC Act entry is a starting point.

The first three prompts

Run these in order in a new chat inside the project.

Interview me one question at a time so you can support my workers
compensation practice. Ask:
1. Which scheme and legislation do I work under, and which
   sections come up most often?
2. Which claim stages and decision points do I handle: initial
   liability, ongoing entitlements, return to work, disputes?
3. Who reads my work: delegates, injured workers, employers,
   tribunals?
4. What does a strong determination brief look like in my team,
   and what do reviewers most often correct?
5. What is my de-identification practice before any material
   enters an AI tool?
6. Which templates and precedents do I reuse, and where do they
   live?
Then summarise my answers as folder-ready context notes for my WC
project space.

I am pasting a de-identified evidence pack for a determination
brief. Extract the evidence items and map each one to the relevant
legislative test from wc-legislation-map.md, flagging any item
that supports no test and any test with no supporting evidence.
Then structure the brief skeleton against
determination-brief-template.md and stop for my approval before
drafting.

I am pasting de-identified call notes. Structure them into facts,
actions and follow-ups, then draft a file note in my register's
required format, following writing-style.md. End with "For human
review:" and anything a delegate must verify against the claim
file.

Governance guardrails

• De-identification is mandatory, not optional, even in an enterprise workspace with training excluded. De-identify at the source, keep the placeholders through every step, and re-personalise only inside your controlled claims system after human review.
• The AI drafts; the delegate decides. Every decision document keeps its outcome with the human who holds the delegation, and the "For human review:" footer is part of every draft, never trimmed.
• On a consumer plan, turn off the model-improvement setting (Settings, then Privacy, then "Help Improve our AI models") before first use, and do not submit feedback from work chats on any plan.
• Verify every section reference the assistant produces against the current compilation before it leaves your desk.
• Check your organisation's AI use policy before setting this up; a personal consumer account is not an approved channel for work material in most regulated workplaces.

About this configuration

Built against: Claude Projects on claude.ai (web), August 2026. Platform behaviour and plans change; re-check the sources above before relying on any data-handling claim. TheAICommand is independent: no platform vendor paid for, reviewed or influenced this configuration, and we accept no payment for coverage. See ownership and funding.