Bringing Claude
into Production: Attorney Share Success Story
How Distillery and Claude cut Attorney Share’s ticket-to-staging time from about ten days to four, with 40% fewer active engineers and 88% fewer reverts and hotfixes per line of code, and put Claude to work inside the product.
Summary
Ten-day tickets to four-day releases.
Attorney Share is a digital marketplace where attorneys refer cases to one another. Distillery has built and evolved the platform alongside Attorney Share’s in-house team since its MVP. Since August 2025, Distillery engineers have made Claude (Claude Code, Claude chat, Cowork and MCP integrations) the center of how the team plans, builds, tests and documents. Tickets now reach staging in about four days instead of ten while growing nine times larger, end-to-end tests grew from 13 to 566 and cover 12 of 14 core user flows, reverts and hotfixes per line of code fell 88%, and the number of active engineers fell 40% (9.2 to 5.5 a month) as Distillery’s delivery team went from eight roles to four, while the live platform kept growing. Distillery also built three Claude features that run inside the product: case intake, referral scoring and an incident monitor. Case intake and referral scoring call the Claude API directly, on Claude Sonnet, through Attorney Share’s own Anthropic account, and Claude has scored 29,513 referrals since October 2025. It also built an MCP server that lets users manage their cases from Claude Desktop.
In production
Claude in production at a glance
Claude runs inside Attorney Share’s product through the Claude API, on Attorney Share’s own Anthropic account, and Distillery’s engineers also use it every day to build the platform. Case intake and referral scoring call the Claude API directly, with no cloud reseller or gateway in between.
Distillery worked alongside Attorney Share’s in-house team and was responsible for taking the platform live with Claude: Distillery deployed Claude into production, rolled out the features described below, and updated the development lifecycle around Claude, from planning and coding to testing and documentation.
29,513
Since October 31, 2025, each with a 0-100 quality score and a written rationale
97%
Of 24,346 new cases since January 2026
About 6.6 seconds
On Claude Sonnet 5.5, about half the latency of the previous model
at the same cost
Every hour
20 business metrics against 168 weekday-and-hour baselines; an anomaly surfaces in about 1.5 hours
21
10 read and 11 action tools, usable from Claude Desktop, Claude Code and claude.ai
3,146 of 5,842 (54%)
Since August 2025, across 8 repositories
Referrals scored by Claude per quarter
Each with a written rationale
Results at glance.
Every metric below compares the project before and after Claude became part of the team’s daily workflow.
George Durzi
CTO at Attorney Share
The challenge
Background and challenge
Attorney Share had to build a secure, compliant legal marketplace from scratch and keep it moving as the business changed. The platform lets any licensed lawyer create an account, integrates with legal case management systems so cases can be referred in a few clicks, and had to meet demanding security expectations from day one. Attorneys post cases they cannot take themselves, for example outside their jurisdiction, and another attorney buys the case. The client’s personal details are released only when the case is purchased. When a case is posted, the seller puts the initial commission or fee, but other lawyers are allowed to counter-offer, and when an agreement is reached, the case and the information about it is sold.
Distillery and Attorney Share’s in-house team delivered the MVP in two months using the JavaScript ecosystem and AWS. From there the product kept reshaping: it pivoted from single attorneys to law firms with multiple users and roles, and Distillery’s team grew to eight. The challenge became sustaining speed and quality as scope grew, without growing headcount to match.
The approach
How Claude is used on the project
Distillery engineers have used Claude subscriptions on Attorney Share since January 2026, using both Distillery’s own subscriptions (3 developers and 1 QA) and Attorney Share’s (2 developers, 1 CEO, and 1 PM/PO), across the full delivery cycle rather than just coding.
Claude Code
Solution design, technical research, implementation planning, coding, integrations, testing, root-cause analysis and refactoring, with engineers supplying context on the codebase, business rules, architecture and conventions
Claude Code Skills
Custom skills the team built for code quality, testing, and how user stories are written, plus technology skills for PostgreSQL, Stripe and Playwright, and documentation shipped with every ticket
MCP integrations
Linear, so Claude reads and updates tickets directly, and Figma, so designs flow straight into implementation. As well we integrate Datadog and AWS MCP servers to investigate and debug different scenarios and errors during development and production.
Claude (chat)
Analyzing the product to produce user stories and the full backlog, which are then pushed to Linear
Claude Cowork
Automating routine tasks and answering questions from the team's knowledge base
The team keeps building a shared framework of best practices and a knowledge base that Claude Code draws on. As it grows, errors drop and both quality and velocity improve.
Since August 2025, Claude has co-authored more than half of all commits on the platform: 3,146 of 5,842. Test selection in CI was rebuilt on the Claude Agent SDK in about two days in July 2026, so each pull request runs only the end-to-end suites it affects.
powered features
Claude inside the Attorney Share product
Beyond how Distillery works, Claude also runs inside the product itself. Distillery built three Claude-powered features for the platform, working with Attorney Share’s product and engineering stakeholders: case intake, referral scoring and an incident monitor. All three are live in production. Distillery’s role covered prompt design, the PII-scrubbing pass, evaluation and production rollout, and Distillery still maintains all three today, from fixing errors and optimizing usage to tuning the prompts used for classification.
The goal of the first two features is the same: turn messy input from an attorney into a clean, structured, PII-safe referral listing, and help judge how good the lead is. The third keeps the platform itself healthy.
Case intake
Cases reach the platform in several ways: a web form, partner systems, and a bot that takes a case by phone. When the bot takes a call, it transcribes the conversation and asks questions to create the case, and Claude decides which field each piece of information belongs to. When a partner firm submits a plain-text referral summary, Claude writes the marketplace listing title, then a second pass checks that no personal information (names, phone numbers and so on) leaks into it. When a partner submits a raw intake-call transcript, Claude extracts practice area, jurisdiction, lead contact details, incident date and a summary, and the summary is scrubbed of personal information before it becomes the listing. Personal details stay hidden until a lawyer buys the case.
Referral processing pipeline
Cases from every source (phone, forms and mostly third party systems known as LCMS (Legal Case Management Software)) pass through a background pipeline before they are published. Claude reads the case and returns a 0-100 quality score for how good it is for a lawyer to take, including how likely the lead is to sign and how complete the information is, along with whether the client is likely still looking for a lawyer and a short rationale, so lawyers can see why. The score and rationale are stored with the referral.
Incident monitor
Claude reviews application logs, with access through an MCP server, to produce reports for internal use and to help fix errors that customers run into. It checks whether the app or one of its modules is failing, sends messages in Slack, and lets engineers chat with it to work through an incident, suggesting hypotheses about what may be happening and possible fixes. It checks production every hour against weekday-and-hour baselines for 20 business metrics (cases, proposals, payments, sign-ups), so anomalies surface within about 90 minutes. Live in production since August 31, 2026.
Referral scoring has run on Claude since October 31, 2025 and case intake since November 2025; the incident monitor went live on August 31, 2026. Case intake and scoring run on Claude Sonnet (scoring on Claude Sonnet 5.5 since October 2026, which cut its latency by about half at the same cost) through Attorney Share’s own Anthropic API account. The incident monitor runs its hourly checks on Claude Haiku 4.5 and its investigations on Claude Code with Claude Sonnet 5.5. Field mapping for case management systems such as Clio and Litify is deterministic code, and other AI features in the product do not use Claude.
Since October 31, 2025, Claude has scored 29,513 referrals, each with a 0–100 quality score and a written rationale (average score 62).
Since January 2026, 97% of all new cases on the platform have been scored, at a median of about 6.6 seconds per score.
MCP server for Claude Desktop
Distillery also built an MCP server for the Attorney Share platform. Connected to Claude Desktop, it lets a user search the marketplace, track their referrals and case updates, review and accept proposals, submit proposals, post a referral from an intake transcript and report case status by asking Claude, without opening the app. It extends the platform into the user's own Claude Desktop, so it is separate from the three features above. It launched in June 2026, with a seller-focused second version in August 2026, exposes 21 tools (10 read and 11 action), and is open to any attorney with an Attorney Share account through a Direct Share API key or OAuth, from Claude Desktop, Claude Code or a claude.ai connector.
Three wins that show the impact
Test automation went from a bottleneck to a by-product of every feature
Automated coverage sat at about 10% (13 end-to-end tests), and writing a full test suite for one feature took roughly a week.
Distillery's test automation engineer used Claude Code to build a Playwright framework from the ground up, covering page models, fixtures, authentication workflows and CI integration. Now every feature spawns multiple agents that write unit tests, Playwright tests and more, and Claude Code analyzes tickets and pull requests, validates against preview environments and investigates failures through logs.
End-to-end tests grew from 13 to 566 across 37 product areas and now cover 12 of 14 core user flows (86%), alongside 3,549 front-end unit tests. End-to-end tests written per month went from about 4 to about 87. The Playwright framework itself was then modernized (1.56 to 1.64, 13 pull requests merged) in about a day and a half. The QA lead no longer writes most tests and instead builds the agents, skills and harnesses that do. Fewer issues bounce between QA and developers, and QA can fix simple bugs directly with Claude Code.
Months-long refactors shipped in days
Adopting a new technology partway through the project required large refactors that the team expected to take months. In 2024, a comparable rework of the payment flow (refunds, capture on signing and changing the payment method, about 11,000 lines of code) took about two months.
Once end-to-end tests covered the core user flows, the team could let Claude Code drive the refactor and verify behavior continuously.
In 2026, rebuilding Stripe invoicing and webhooks (about 12,000 lines) took about two weeks, roughly four times faster for a change of similar size, and moving billing to firm-level Stripe accounts took about a week. Other refactors followed the same pattern: moving the product's case-intake AI onto Claude Sonnet took one day (November 4, 2025), and modernizing the Playwright test framework took about a day and a half.
A lean team covering the full delivery cycle
Distillery's delivery team had eight roles.
Claude chat produced user stories and the backlog from analysis of the product, the Linear MCP server let Claude manage tickets directly, and the Figma MCP server connected design to code. Skills generate documentation with every ticket.
Distillery's delivery team now runs with four roles instead of eight, while ticket-to-staging time, test coverage and defects all improved at the same time (see the results table above). Active engineers fell 40% (9.0 to 5.5 a month) against the pre-Claude average, and the team merged 96 pull requests a month against 120 before, with about twice the code per engineer (5,627 lines against 2,690).
CONTROLS
Quality and
security controls
Claude accelerates the work, but people remain accountable for every change.
An engineer reviews every pull request.
George Durzi reviews every user story.
Development happens in dev environments, with access to the live database restricted.
Personal information in a referral stays hidden until a lawyer buys the case, and a second pass checks listings for leaked personal details.
The team follows SOC 2 guidelines throughout.
Outcomes and what’s next
The platform is live in production and growing: Attorney Share’s account base grew 85% in users (3,405 to 6,311) and 87% in firms (2,638 to 4,920) between August 2025 and November 2026.
This growth runs alongside the project and is not attributed to Claude.
NOTES ON THE FIGURES
"Before Claude" is the 2024 to first-half 2025 average and "with Claude" is the second half of 2026 (July to October). Ticket-to-staging is the median time a ticket spends from In Progress to On Staging in Linear, by the quarter the ticket started (9.6 days in Q3 2025, 4.0 days in Q3 2026). Ticket size is the median number of changed lines per ticket (73 before, 659 now). An active engineer is one with three or more merged pull requests in a month; bot, release and sync pull requests are excluded. Reverts and hotfixes are pull requests titled or branched "revert" or "hotfix" (114 in 2,162 pull requests before Claude, 7 in 383 in the second half of 2026). Lines of code measure volume, not value. Functional test coverage counts a core user flow when an end-to-end test exercises it from start to finish; line coverage is not measured. Account growth runs alongside the project and is not attributed to Claude. Sources: Linear (2,194 tickets), GitHub (8 repositories, 4,088 merged pull requests), git history and Attorney Share's reporting warehouse, pulled on November 6, 2026.