Claude Code can do more for a business than draft a document: it can read project instructions, use connected tools, assemble a deliverable and check parts of its own work. The useful distinction is between generating an output and making that output dependable. The workflows below show both the labor an AI agent can take on and the points where a person still needs to inspect, decide and approve.
Start with a workspace, not a one-off prompt
Claude Code has a desktop application, so using it does not require working in a command-line window. A project begins with a local folder that the tool can access. That folder can hold reference documents, brand assets and instructions for the work ahead. Claude Code can read and organize those files, which makes the choice of project folder and access permissions part of the setup.
A file named CLAUDE.md can serve as the project’s instruction guide. Markdown, the plain-text format used by that file, lets an operator organize rules with headings and lists. The guide can describe the business, point to more detailed documents and state what must happen before a deliverable is complete. For example, a rule can require a verification pass and screenshots of visual work before the agent presents it for approval.
For a recurring task, a separate SKILL.md file can hold a reusable procedure: what to gather, how to structure the result and which checks to run. The distinction matters. Fixing one report improves that report; updating the saved skill gives the next run the same instruction. Building and refining such a procedure can take longer than doing a single task manually, so it makes the most sense for work a business expects to repeat.
From lead research to a spreadsheet awaiting approval
One business workflow connected Claude Code to Clay, a tool for finding and enriching business leads. The connection used an API key, a private credential that lets one tool access another service under granted permissions. Existing project files supplied context about the intended customers, while the task asked for 50 suitable businesses, tailored outreach drafts and a Google Sheets deliverable.
The agent gathered company information and checked attributes relevant to each draft, including website details, reviews, advertised service availability and online booking. It then organized the results in a sheet with outreach subjects, message bodies and a review-status field. In this case, the 50-record sheet took under 10 minutes to produce. That is a result from this particular workflow, not a timetable a business can assume for its own data or connections.
▲ Checks before outreach approval
The important step came before any message was sent. Nine research agents gathered site information, and two separate verification agents compared the outreach claims with the underlying facts. Those checks found and corrected five discrepancies, including an outdated promotion. The drafts still needed a person to judge whether the businesses were appropriate prospects, whether the messages represented the offer accurately and whether outreach should proceed. Connecting an approved sheet to a sending workflow is a separate decision from generating it.
Use business intent to shape the output
A second workflow assembled YouTube Analytics into a quarterly Google Sheets report. Its instructions went beyond “make a spreadsheet”: the business wanted to understand viewership trends, identify performance drivers and plan the next quarter’s content. Claude Code gathered metrics and organized an overview, monthly trends and quarterly tabs, including views, watch hours, traffic sources and top-performing videos.
The reported run took about 10 minutes for work estimated at three to four hours by hand. The saving came from gathering, calculating and formatting the figures together. It did not remove the need to understand what the measures meant or decide which content to make next. An operator can use the finished sheet to investigate a spike or compare formats, but the business judgment remains human.
A smaller automation illustrates the same boundary. Claude Code assembled an n8n workflow—n8n is a tool for connecting automated steps—after roughly two minutes of spoken instructions. The full task took about 10 minutes rather than the usual 30 to 45 minutes of manual construction. It still required variable mapping and a run with sample data before the workflow operated end to end.
Match each check to the kind of error
An agent that reviews its own output needs more than one kind of check. A market-analysis workflow shows why. Claude Code researched and built a formatted HTML report, then used browser screenshots to inspect how it rendered. The visual review could catch layout or contrast problems, but it could not establish whether a market figure was sound. A separate fact-checking agent reviewed 21 claims: 18 were confirmed, one was outdated and two had source or attribution problems.
▲ Visual and factual review
The checks serve different purposes:
- Screenshot inspection checks whether a report, document or graphic looks right when rendered. It does not verify its claims.
- Source comparison checks figures and statements against the material used to support them. It is needed even when the page looks polished.
- Browser testing checks whether an application’s buttons, forms and other user flows work as intended. In one scheduling-app project, 50 automated agents simulated interactions to uncover problems before human testing.
The report workflow became more reusable when its corrections were added to the saved skill, including deeper market-size analysis and a required source-checking pass. During early runs, an operator needs to watch what the agent does and correct its procedure, not merely accept or reject the final file. Later runs can reuse the improved procedure, but they still warrant review appropriate to the task.
Put one repeatable task to work
For a service business, the practical starting point is a bounded, recurring job: preparing a report, assembling research or drafting material for approval. Give Claude Code the relevant project files, state the business purpose and define what a satisfactory deliverable must contain. Grant the tool access to the connections that job needs, then observe its first runs closely.
Before making the workflow routine, check its output against the original data, inspect any visual deliverable and decide which actions require human sign-off. Save the corrections in the project instructions or skill rather than repeating them in every prompt. Claude Code can shorten the path from request to draft, as the lead, analytics and automation cases illustrate. The business value comes from making that path repeatable without treating an unchecked draft as a finished decision.