Claude Code business workflows for law firms

TaskChad sells the $250 Business Diagnostic Session and the $2,000 14-Day Implementation Sprint described on this page. This page is provider-written implementation guidance from TaskChad's product team, not independent research, a bar-association publication, or a customer case study. Every workflow named below is a scoping hypothesis until a real law firm pays for a Session, accepts a scope, and TaskChad holds terminal evidence for the result.

The expensive problem: an open Claude Code session doesn't know when drafting became advising

A firm that hands an associate Claude Code gets a fast first draft: a client status email built from a case-management export, a discovery production index built from a document list, a billing narrative cleaned up from a terse time entry. The associate who found the prompt that works well keeps it in their own chat history, not somewhere the paralegal or the next associate can reuse it, so the firm re-solves the same drafting problem every time someone new tries the tool. The sharper risk shows up one message later, when that same open session gets asked whether a client should take a settlement offer, because answering feels like the next question in the conversation rather than a different kind of question that requires a bar license.

Drafting from verified case data is repeatable once matter conventions, house phrasing, and the no-legal-advice boundary live in a file Claude Code reads automatically. Without one, "each Claude Code session begins with a fresh context window," so even safe drafting restarts from nothing, and the line between paperwork and advice gets re-explained by whoever is typing that day (Claude Code, "How Claude remembers your project"). Comment 8 to Model Rule 1.1 ties competence to staying abreast of "the benefits and risks associated with relevant technology" (ABA Model Rule 1.1, Comment 8), and ABA Formal Opinion 512, issued July 29, 2024, extended that duty to generative AI specifically, requiring a lawyer to reasonably understand a tool's capabilities and limitations before using it on a matter (ABA Formal Opinion 512). A workflow that drafts paperwork around a licensed attorney's decisions belongs in this lane; a workflow that makes the decision does not.

Two Claude Code primitives, and the boundary this firm needs a third to hold

CLAUDE.md files and Claude's own auto memory are Claude Code's two complementary memory systems, and "both are loaded at the start of every conversation," so a firm's practice-area list, house phrasing, and no-legal-advice boundary stop depending on whichever associate remembers the details that day (Claude Code, "How Claude remembers your project"). A skill is a folder holding a SKILL.md file; "Claude uses skills when relevant, or you can invoke one directly," and unlike CLAUDE.md content, "a skill's body loads only when it's used," so a firm can hold a discovery-index skill, a status-letter skill, and a billing-narrative skill side by side without any of them consuming context on a day nobody needs them (Claude Code, "Extend Claude with skills"). The Sprint turns those primitives into four artifacts: a context contract carrying the practice-area list, house phrasing, and the no-legal-advice boundary; a skill scoped to one work-packet type; a test fixture built from synthetic or staged matter facts, never a live client file; and an operator runbook naming who routes output to a reviewing attorney.

Neither primitive holds a boundary alone. "Permission rules are enforced by Claude Code, not by the model. Instructions in your prompt or CLAUDE.md shape what Claude tries to do, but they don't change what Claude Code allows" (Claude Code, "Configure permissions"). A CLAUDE.md line telling Claude Code never to evaluate a matter's merits is forgettable at 11 p.m. before a filing deadline; this cell keeps that boundary out of Claude Code's reach with a deny rule, not wording alone.

Where a law firm's paperwork actually gets produced today

The Business Diagnostic Session maps where each candidate work packet gets produced before any build starts. This table is a scoping instrument, not a recommendation; during the paid Session, every cell gets replaced with the firm's real owner, system, and gap.

Candidate work packet	Owner today	Where the facts live	What breaks today
New-matter intake summary	Intake coordinator or whichever associate answers	Intake form or call notes, case-management system	Matter facts get typed from memory before a conflicts check is confirmed
Discovery production index	Paralegal	Document management system, the request list	Bates ranges and request items drift out of sync when tracked in a shared spreadsheet
Client status update draft	Associate or paralegal	Case-management system entries, calendar	Wording varies by whoever drafts it, and a factual update can drift into a prediction
Billing narrative cleanup	Timekeeper or billing coordinator	Raw time-entry log	Terse entries get rewritten from memory days later, thinner or vaguer than the work performed

A firm running a busy litigation calendar likely ranks the discovery production index highest; a firm with high-volume, unqualified consultation requests ranks the intake summary first instead. The billing narrative packet sits closest to a client-facing financial record, so a first build there gets extra scrutiny on accuracy over speed. Ranking is Session work, not a default.

Baseline and the KPI: accepted work packets without re-entering business context

TaskChad records the baseline before any build: how often the work packet recurs, how much attorney or paralegal time goes into producing it from scratch, and whether a named reviewer currently checks a draft before it reaches a client, court, or opposing counsel.

The KPI is accepted work packets completed without re-entering business context. A packet counts only when the operator invoked the skill without retyping the practice-area list or the no-legal-advice boundary, because the context contract already carried them; the named reviewer approved the output against a verified case-management entry or time-entry record with no correction needed to a fact the contract should have supplied; and the disposition — sent, filed, or logged — got recorded against that acceptance.

Work packet type	Volume source	Counted as accepted when	Disposition owner
New-matter intake summary	Skill log plus intake form or call log	Facts tie to the intake record, conflicts check confirmed first	Intake coordinator or managing attorney
Discovery production index	Skill log plus document management system export	Bates ranges and request items match the source list exactly	Paralegal
Client status update draft	Skill log plus case-management system entries	Status matches the case-management record, with no advice added	Associate or reviewing attorney
Billing narrative cleanup	Skill log plus time-entry system	Narrative maps to the actual time entry and task, with nothing added	Timekeeper or billing coordinator

A skill that ran without erroring is not evidence the KPI moved. Only a reviewer's logged acceptance counts.

Six states between a source record and a filed work packet

1. Request — the operator names one work packet and supplies verified source data: an intake form, a document management system export, a time-entry log. Facts are never typed from memory.
2. Load — Claude Code reads the context contract automatically, so the practice-area list, house phrasing, and the no-legal-advice boundary are already present.
3. Draft — the skill produces the work packet, citing which source record each figure came from and flagging anything it cannot verify.
4. Security check — permission rules confirm no credential is embedded, and the draft never touched a merits-evaluation, conflicts-clearance, or engagement-promise field.
5. Business review — the named reviewer checks the draft against source data and the no-legal-advice boundary, marking it approved, needs revision, or rejected.
6. File, send, or log — an approved packet goes to the client, court, opposing counsel, or billing system, and the repository owner logs whether it counted toward the KPI.

Systems this Sprint touches, and the fields it never touches

This cell's working systems are Claude Code, the repository holding the context contract, and an approval queue. Firms also run a case-management system, a document management system, a calendar, and a time-and-billing platform — but this Sprint keeps Claude Code reading exported records rather than connecting live to production data, holding the build to two or three connected systems. A live connection is possible later through the Model Context Protocol, which "gives Claude Code access to your tools, databases, and APIs" (Claude Code, "Connect Claude Code to tools via MCP"), but that stays out of scope until the file-based workflow has a track record.

Two categories stay outside the context contract entirely: secrets, blocked by a deny rule on .env files and a secrets/ directory so a case-management API key never reaches a prompt; and any merits-evaluation, conflicts-clearance, or engagement-promise field. If a matter's file already carries a conflicts-check result, the skill reads that value as a fact, the same way it reads a case number, and a deny rule blocks any write path back into it.

Three named approvers, and why sign-off can't be optional

Three roles carry approval authority. The repository owner maintains the context contract, skill, and test fixture — usually a managing attorney or intake director. The business reviewer checks a draft against verified source data before it reaches a client, court, or opposing counsel: a designated attorney for a new-matter summary, the supervising attorney for a discovery index or status update, the supervising partner for a billing narrative. The security owner controls which credentials the workflow can touch and confirms the deny rules covering secrets, the no-legal-advice field, and the conflicts field.

That review step does real work under the rules governing the profession. Formal Opinion 512 ties a lawyer's use of a generative AI tool to Rule 5.3's existing supervisory duty, which requires reasonable efforts to ensure assistance — human or otherwise — behaves compatibly with the lawyer's own professional obligations (ABA Model Rule 5.3). Confidentiality carries equal weight for a matter-adjacent draft: Rule 1.6 protects "information relating to the representation of a client," from whatever source, which is why any test fixture uses synthetic or staged facts rather than a live client file (ABA Model Rule 1.6).

The billing narrative packet carries its own stakes. Formal Opinion 512 addresses billing directly: a lawyer billing by the hour must charge only for time actually spent, even when a generative AI tool completes the work faster, and a flat fee reflecting little or no actual work performed can be an unreasonable fee under Model Rule 1.5 (ABA Formal Opinion 512, "Generative Artificial Intelligence Tools"; ABA, "Formal Opinion 512 and the Reasonableness of Fees When Using AI"). This Sprint's billing-narrative skill rephrases a timekeeper's own entry for clarity, never lengthening or embellishing it, and a deny rule blocks any output describing work the time-entry record doesn't already show.

Failure tests before this workflow ships

A workflow is not ready because it worked once in a demo. It is ready once TaskChad has tried to break it and watched it fail safely.

• Secret in a prompt or context file. A case-management API key gets pasted into chat or committed to a context file; the deny rule must block it before it reaches a shared skill or repository history.
• Merits or engagement-promise write attempt. The skill is asked to evaluate a matter's strength, predict an outcome, or imply representation has begun; the deny rule blocks the write, and the request routes back to a licensed attorney.
• Fabricated figure. The skill is asked to draft a status update or discovery index with no verified export attached; it must flag the missing figure rather than guess one.
• Unreviewed send. A work packet reaches a client, court, or opposing counsel without sign-off; the approval state has to be a logged decision, not prompt text asking Claude to check first.
• Embellished billing narrative. A draft describes work broader or longer than the underlying time entry supports; the skill halts the line item and routes it back to the timekeeper rather than sending an unsupported narrative.

Each test must produce a visible failure state and a named next action. Silence is not an acceptable outcome.

The 14-day Sprint for this cell

Days	Phase	What happens
1–3	Preflight and baseline	Confirm which work packet recurs most and has a named reviewer; pull sample intake records, document-management exports, or time-entry logs for the test fixture
4–7	Build	Write the context contract and first skill for the highest-volume work packet, wired to permission rules blocking secrets, merits-evaluation fields, and unsupported billing narratives
8–11	Failure and approval tests	Run the five failure tests above; confirm the three named roles can act on a draft inside the observation window
12–14	Release and handoff	Hand the skill and context contract to the repository owner with the operator runbook, document safe-disable, record the baseline receipt

For this technical example, the working scope is one work packet type, at most two or three connected systems, one named KPI, one owner, one release, one acceptance decision. A live MCP connection into production systems, a rollout across every work packet at once, any client-facing send without attorney review, and any workflow letting Claude Code evaluate a matter's merits or promise representation sit outside this technical example. When a request exceeds that boundary, TaskChad narrows the scope or declines the engagement. The purchased Sprint is scoped to the agreed business result, which may address one big problem or several connected problems.

Fit and wait conditions

Likely a fit:

• The firm already has one recurring work packet — intake summaries, discovery indexes, status updates, billing narratives — that several people draft inconsistently today.
• A specific attorney or paralegal is willing to be named business reviewer and can commit to checking drafts inside the observation window.
• Source data already exists as a record the firm can hand over: an intake form, a document-management export, a time-entry log.
• The firm already keeps merits evaluation, conflicts clearance, and engagement decisions inside its attorneys' own judgment; this Sprint drafts paperwork around those decisions, it does not replace them.

Reasonable to wait, or fix something else first:

• No one at the firm has tried Claude Code for this work yet, so there is no baseline to rank a first build against.
• The firm wants one workflow to touch every practice area and every work packet type at once — that is a program, not a two-week Sprint.
• The actual request is for the workflow to evaluate a matter, draft client-facing legal analysis, or send communication without an attorney reading it first.

Terminal evidence: what "working" is allowed to mean

A drafted status update is not a result. A skill that ran without erroring is not a result. A discovery index sitting unreviewed in a chat transcript is not a result. The only evidence this cell treats as terminal is a work packet the named business reviewer accepted and sent, filed, or logged — counted only when the operator did not have to re-enter business context to reach that acceptance — observed over a stated window against the dated baseline.

That distinction holds across every TaskChad Session and Sprint. Activity does not get promoted into a customer result; a reported outcome comes from the system of record, observed over a stated window, with the baseline written down first.

See the pattern before you pay for it

TaskChad publishes three controlled demonstrations so a firm can see the mechanics before paying for anything. The AI Workflow Audit demonstration shows what this page describes in miniature: naming candidate work packets, scoring which is worth building first, and producing one bounded Sprint recommendation. The lead-to-booking demonstration shows a capture-to-receipt path with a human-approval hold before anything reaches a prospect, comparable to the business-review state used here before a draft reaches a client or court. The SEO and GEO improvement loop demonstration is unrelated to this lane's build, but shows how TaskChad treats a measurement claim generally.

Before booking a Session, a firm can run the Revenue Leak Score for law firms, a free, deterministic check across visibility, trust, capture, response, follow-up, and owner dependency. It names the single highest-priority leak without a call — a starting point for deciding whether an intake summary, discovery index, status update, or billing narrative workflow deserves the first Sprint.

To scope this specific route, book the $250 Business Diagnostic Session for Claude Code business workflows, law firms. Paid Sessions are contacted within one business day to schedule; paying does not book a specific time automatically.

Frequently asked questions

Will Claude Code ever evaluate our case or predict an outcome?

No. Matter evaluation, outcome prediction, and engagement decisions stay with a licensed attorney, and the skill carries a deny rule so it has no write path into those fields. This Sprint drafts the paperwork that follows a decision an attorney already made or a fact the case-management system already records; it never makes that decision.

What happens if the source record for a work packet is missing or incomplete?

The skill is required to flag the gap, not guess. Drafting a status update without a matching case-management entry, or a discovery index without the underlying document list, is one of the failure tests run before the workflow ships — the draft stops, names the missing figure, and routes back to whoever can supply it.

Will this workflow touch our billing without a timekeeper's review?

No. A billing-narrative draft rephrases a timekeeper's own entry for clarity, and every draft routes back to the timekeeper or billing coordinator before it reaches a client invoice. A narrative that describes work broader than the underlying time entry is one of the failure tests run before launch, consistent with the fee-reasonableness guidance in Formal Opinion 512.

What does the $250 Session produce, and how does it connect to the $2,000 14-Day Implementation Sprint?

The Session produces a written brief within two business days: the ranked work packet, the KPI and baseline source, the context-contract and skill scope, the failure tests, and the named repository owner, business reviewer, and security owner. It does not touch production systems. The Sprint builds the agreed solution, with acceptance tests and an operator handoff. The Session fee credits toward an accepted Sprint for 30 days.

Sources

• Claude Code, "How Claude remembers your project" — CLAUDE.md and auto memory both load at the start of every conversation.
• Claude Code, "Extend Claude with skills" — a skill's body loads only when it's used.
• Claude Code, "Configure permissions" — permission rules are enforced by Claude Code itself, not the model.
• Claude Code, "Connect Claude Code to tools via MCP" — MCP as the mechanism for a live case-management connection.
• ABA Model Rule 1.1, Comment 8 — competence includes staying abreast of the benefits and risks of relevant technology.
• ABA Formal Opinion 512, "Generative Artificial Intelligence Tools", issued July 29, 2024 — the ABA's formal ethics guidance on lawyers' use of generative AI, including billing.
• ABA Model Rule 5.3, "Responsibilities Regarding Nonlawyer Assistance" — the supervisory duty Formal Opinion 512 extends to AI tools.
• ABA Model Rule 1.6, "Confidentiality of Information" — protects information relating to a client's representation from whatever source.
• ABA, "Formal Opinion 512 and the Reasonableness of Fees When Using AI" — billing must reflect time actually spent, and a flat fee reflecting little work performed can be unreasonable under Rule 1.5.