Search "Claude for lawyers" or "Claude for accountants" and you get two kinds of results: breathless lists of prompts, and disclaimers about hallucination. Neither tells you the thing you actually need, which is how to turn Claude into the production line of a one-person service business — a repeatable setup where one careful person produces client-grade deliverables on a monthly rhythm and gets paid a retainer for them.
That is what this guide is. It assumes the model already decided in the one-person AI service business guide: you sell professionals a finished deliverable they already pay for — a regulatory alert, an R&D-credit memo, a farm campaign — produced on an AI production line you run yourself, with no code and no integration into their systems. This page is the operational version of "run that production line on Claude," feature by feature, plus an honest account of where Claude does and doesn't matter.
First, the honest part: the model is not the moat
The flagship guide says it plainly and this page will not contradict it: the model is the least differentiated part of the stack. The mega-prompts in this business are structured plain text; they run on any frontier assistant, and if you already work fluently in a different one, the migration cost is a copy-paste. Your source-checking and your editorial review — not your model choice — are what make the output client-grade. Anyone selling you "the Claude secret" is selling you the wrong variable.
So why a whole guide on Claude? Because how the tool is organized around your work changes how fast and how reliably one person can run ten client production lines at once — and on that axis, four of Claude's features map unusually cleanly onto the productized-service model. This isn't "Claude is smarter." It's "Claude's workspace shape happens to be the shape of this business." Read it as: here is the sharpest way to run the model-agnostic playbook, using the features that reduce your per-client overhead. If you run it somewhere else, port the pattern, not the buttons.
The four features that map to the production line
A productized AI service has four moving parts: a per-client context (their practice area, their voice, their prior work), a repeatable procedure (the mega-prompt plus the review checklist), a finished artifact (the document you hand over), and a verification step (every dated claim checked against a primary source). Claude has a native home for each.
1. Projects = the per-client account file
A Claude Project is a persistent workspace with its own custom instructions and its own uploaded knowledge. Create one Project per client, not one per task. Into it goes: the client's practice area and jurisdictions, their tone and formatting conventions, samples of work they've published under their own name, and the standing instruction for how you review before shipping. Every conversation inside that Project inherits all of it, so you are not re-explaining "this is a boutique employment firm in Illinois that writes in plain English" on the fifth of every month — the Project already knows.
This is the single biggest overhead-killer in the whole setup. The reason one person can carry ten retainers is that ten Projects hold ten clients' context for you, and switching clients is switching Projects, not rebuilding context from memory. Name them by client, keep the knowledge current, and your marginal cost per client-month falls exactly the way the service model promises it will.
2. Skills = the production line, made reusable
The kit ships fifteen mega-prompts — five per market. Pasting a 900-word prompt works at three clients and hurts at twelve; the fix is to stop pasting. Claude's Skills let you package a procedure — the instructions, the output format, the review gates — as a named capability Claude loads on demand, so "run the monthly alert" becomes an invocation instead of a wall of pasted text. Your fifteen prompts become fifteen skills; the production line stops living in your clipboard and starts living in the tool.
This is also the bridge to scale that the ladder is built around. Hand-run skills are the manual transmission; when the volume justifies it, the hosted LibSkills packs ($99/mo per vertical) run the same five skills per market as browser run-forms with the current week's market data injected and QA guardrails on every output — the automatic transmission to this page's manual. You do not have to start there. You start by turning your prompts into skills you invoke by name, and you graduate to hosted packs when twelve clients make hand-running them the bottleneck.
3. Artifacts = the deliverable, client-ready
The thing the client pays for is a document — a brief, a memo, a campaign, a ready-to-send email. Claude's artifacts render that deliverable as a standalone document beside the chat, which you refine in place and export. The practical win is that the boundary between "AI output" and "the file I send" gets thin: you are editing the deliverable itself, in its final shape, not copying raw chat text into a Google Doc and reformatting. For work that recurs monthly, shaving the reformatting tax off every unit compounds.
4. Long context + web search = verification you can actually do
This business lives or dies on dated accuracy — the entire pitch is "the rule changed on this date and here is what it means for your clients," and a wrong effective date is a fired client. Claude's large context window lets you hold the source material, the prior month's deliverable, and the draft in a single conversation, so verification happens in-window instead of across a dozen tabs. With web search enabled, you can pull the current state of a rule and check the draft's dates against a primary source before you ship. Use it as a discipline, not a crutch: the model drafts, you confirm every date against the statute, the IRS page, or the agency notice. That review step is your actual product. (When you scale sourcing beyond hand-checking, Claude's MCP support is the no-code path to connect live data sources — but that's a month-twelve problem, not a week-one one.)
A concrete build: one law-firm client, end to end
Abstract features are useless without a walk-through, so here is a full client, set up once and run monthly. The client is a boutique employment-law firm that wants a monthly regulatory client alert — the highest-scoring legal service in the law-practice niche guide, because employment law is in the middle of a dated AI-regulation wave.
Setup (once, ~90 minutes):
- Create the Project — "Client — [Firm] Employment Alerts." Custom instructions: jurisdiction (Illinois + federal), audience (the firm's HR-leader clients), voice (plain English, no legalese), and the hard rule: every effective date cited must be confirmed against a primary source and quoted with that source.
- Load the knowledge — two alerts the firm previously published, their style guide if they have one, and a one-page list of the practice's focus areas.
- Install the skill — the kit's regulatory-alert mega-prompt, packaged as a skill: gather the month's relevant changes, produce a change summary with effective dates, an affected-client matrix, and two ready-to-send drafts (short email + long-form) in the firm's voice.
Each month (one session):
Invoke the skill. The dated drivers are real and current — Illinois HB 3773 took effect January 1, 2026, and the Colorado AI Act's deployer duties came into force June 30, 2026 — so every affected firm's clients need those changes explained this quarter. Claude drafts the alert against the Project's context. You then do the part that is actually the job: confirm each effective date against the primary source with web search, cut anything you can't verify, edit for the firm's voice, and export the artifact. Ship it on the fifth. Fee band for this service is $600–$1,200/mo; $800 is the standard opening price. The second month takes a fraction of the first, because the Project, the skill, and the review checklist are already built — which is the entire economic argument for the model.
That pattern — Project per client, skill per service, artifact per deliverable, human verification per date — is the whole system. Everything below is the same pattern pointed at a different market.
By market: where the Claude setup differs
The workspace shape is identical across verticals; the dated drivers and the deliverable change. Each of these has its own deep-dive:
- Law firms — regulatory alerts, intake-response drafting, and AI-governance memos, driven by the Illinois/Colorado/NYC AI-law wave. See Claude Skills for law firms.
- CPA & accounting firms — Section 174 R&D catch-up memos, BOI-reporting workflows, and tax-planning briefs, driven by the OBBBA retroactive amendment window and state-level BOI restarts. See Claude for accountants.
- Real-estate agents — ADU feasibility and resale-valuation packets, insurance-and-disclosure briefs, and farm campaigns, driven by California AB 1033, Florida SB 4-D, and the boomer downsizing cohort. See Claude for real estate agents.
What Claude does not fix
Three honest limits, because the disclaimers-only articles are right about one thing even if they're useless about everything else:
It does not verify itself. Every dated claim is your responsibility to confirm. The model will state a plausible effective date with total confidence and be wrong; the affected-client matrix is only as good as your source-check. Build the verification step into the skill and never skip it — it is not overhead, it is the product.
It does not find the client. The setup above produces the deliverable; it does not produce the retainer. Landing the first client is outreach plus a finished free sample, and it takes weeks, not days — the sequence is in how to get your first AI client without an audience. Claude helps you produce the free sample fast; it does not send the emails.
It does not choose your niche. The whole model depends on picking a market you can score with dated evidence before you enter it, and that decision is upstream of any tool. Start where the scoring already exists: the free Sub-Niche Opportunity Report hands you 25 scored sub-niches across legal, accounting, and real estate, each with the dated demand signal and a fee range. Pick the row that fits your background, verify its dates yourself, then build the Project.
Where this sits in the stack
Run this on Claude and you have the manual version of the production line: Projects hold your clients, skills hold your services, and your review holds the quality. When hand-running twelve clients becomes the bottleneck, the hosted LibSkills packs ($99/mo per vertical) are the same five skills per market as guardrailed run-forms — you graduate onto them, you don't rebuild. And to keep the top of the funnel fed, Niche Radar ($19/mo) re-scores the market every Tuesday and hands you the one sub-niche worth moving on that week, so you reach affected clients the week a trigger lands rather than the quarter after.
The kit is where the fifteen prompts and the playbook live: the AI Service Business Kit ($79 once) is the manual for everything above — the prompts you'll package as skills, and the playbook that turns each one into a named, priced service.
FAQ
Do I have to use Claude for this?
No. The prompts are structured plain text and run on any frontier assistant — the model is the least differentiated part of the stack, and your source-checking and editorial review are what make the output client-grade. This guide uses Claude because its workspace shape — Projects, Skills, artifacts — maps unusually cleanly onto a per-client, per-service production line, which lowers the overhead of running many retainers at once. If you already work fluently elsewhere, port the pattern (one workspace per client, one reusable procedure per service, verify every date) rather than the specific buttons.
What's the difference between Projects and Skills here?
A Project is a client — persistent context, their voice, their prior work, held so you never rebuild it. A Skill is a service — a named, reusable procedure (the mega-prompt plus the review gates) you invoke inside any client's Project. Ten Projects × the right Skills is how one person runs ten retainers without re-explaining anything.
Won't clients object to "AI-generated" work?
They are not buying "AI" — they are buying the finished deliverable on their desk on the fifth of the month. The AI is your production line, invisible to the client, exactly as it would be if you drafted with any other tool. What earns the retainer is that the work is correct, on time, and in their voice — which is why the verification step, not the model, is the thing you protect.
How does this connect to the hosted packs?
The packs run the same five skills per market as this manual setup, but as guardrailed browser run-forms with the current week's market data injected — the automatic transmission to this page's manual. You start by hand-running skills in Claude Projects and move onto hosted packs when volume makes hand-running the bottleneck. Same skills, less hand-labor; you graduate, you don't rebuild.