LegalTax-controversy law firm

Designing a Confidential Claude Setup for a Tax-Litigation Firm

The founder had built his own Claude system. The job was making it work for the whole firm without breaking confidentiality rules. A research brief changed the plan before anything got built. The retainer didn't sign.

3architecture options weighed before any build

1workspace per case, no access across cases

6month retainer did not sign

July 9, 2026

The ask

A tax-controversy firm litigating dozens of federal refund cases. The founder spent about four hours a day in Claude. He'd built his own system: 14 skills, 40 commands, 29 hooks and hundreds of scripts, wired into email, his practice software and his CRM.

It worked. For him. Nobody else at the firm could use it.

What I did

The platform. I recommended Claude's enterprise plan over the team plan, because a law firm needs audit logs of who touched client data. I also caught that they were two seats short of the minimum before the upgrade stalled.
Audit before building. The founder estimated 60 to 70% of the planned work already existed. Two weeks later he said 95%. So the first step was measuring the hole. Their chat logs were privileged, so the firm ran a self-audit from a prompt I wrote, and reported back what they found.
A research brief that changed the plan. Anthropic's own guidance said Cowork wasn't built for regulated work that needs audit trails. Tax-return confidentiality law and the bar's confidentiality rule both applied here. I laid out three options and recommended a split: Cowork for low-sensitivity work, Claude Code or the API for anything with raw taxpayer data, and a de-identification skill controlling what crosses between them. Cowork plus activity logging would cover the first 30 to 60 days while that got built. Data retention was set per tool, not once for the whole firm.
One folder per case. Every case got its own workspace and a running log, with no access across cases. Synced folders instead of a data warehouse, because the case files already lived there.
The plan, the policy and the deck. I wrote the product spec, the firm's AI use policy and the executive deck. The policy doubled as the build spec for the skills. The offer was phased: three weeks at a fixed price, with everything after it priced from that review.

The brief also corrected two numbers I'd given them earlier, the seat price and the cost difference between models. I put both corrections in writing.

What happened

The firm moved to the enterprise plan. The six-month retainer didn't sign.

What I learned

One person going 10x faster isn't the hard part. The organization going 10x faster is. I later turned the policy method from this engagement into a reusable skill that drafts AI policies for other firms.

Designing a Confidential Claude Setup for a Tax-Litigation Firm

3architecture options weighed before any build

1workspace per case, no access across cases

6month retainer did not sign

July 9, 2026

The ask

It worked. For him. Nobody else at the firm could use it.

What I did

The platform. I recommended Claude's enterprise plan over the team plan, because a law firm needs audit logs of who touched client data. I also caught that they were two seats short of the minimum before the upgrade stalled.

Audit before building. The founder estimated 60 to 70% of the planned work already existed. Two weeks later he said 95%. So the first step was measuring the hole. Their chat logs were privileged, so the firm ran a self-audit from a prompt I wrote, and reported back what they found.

A research brief that changed the plan. Anthropic's own guidance said Cowork wasn't built for regulated work that needs audit trails. Tax-return confidentiality law and the bar's confidentiality rule both applied here. I laid out three options and recommended a split: Cowork for low-sensitivity work, Claude Code or the API for anything with raw taxpayer data, and a de-identification skill controlling what crosses between them. Cowork plus activity logging would cover the first 30 to 60 days while that got built. Data retention was set per tool, not once for the whole firm.

One folder per case. Every case got its own workspace and a running log, with no access across cases. Synced folders instead of a data warehouse, because the case files already lived there.

The plan, the policy and the deck. I wrote the product spec, the firm's AI use policy and the executive deck. The policy doubled as the build spec for the skills. The offer was phased: three weeks at a fixed price, with everything after it priced from that review.

The brief also corrected two numbers I'd given them earlier, the seat price and the cost difference between models. I put both corrections in writing.