You don't code, and you want Claude to finish a real piece of work in your business, not just answer questions. In our view, the way in is one document you already make by hand. Claude builds it from your own past versions, you check it line by line, and you time the fixing.
By the end you will have the seven steps that take that document from your files to a finished version, the fixing-time test that tells you whether to keep it, and a straight answer on whether you still need a developer.
Seven steps take one document from your own files to a finished version
The task in this playbook is one document you already make by hand: a quote, a proposal or a monthly client report. In our view, a document is a good first task for an owner who doesn't code, because you can check every line of it yourself. Each step below has one thing to do and one test that tells you it worked. Steps one to six need no code. Step seven is optional and comes only after the document works.
Step 1: Pick one document you already make by hand from the same kinds of inputs
Do this. Pick one document you make by hand at least once a month from the same kinds of inputs, such as a quote, a proposal or a monthly client report, and write down how many minutes the last one took you.
It worked when You can name the document, the files and notes it is built from, and the minutes it takes you by hand.
Why. Our rule of thumb is that a good first task repeats often, takes the same kind of input each time, gives the same kind of output, and saves more time than it takes to set up. In a preregistered experiment with 758 knowledge workers, run with Boston Consulting Group, people using GPT-4 on tasks within AI's capabilities completed 12.2% more tasks, 25.1% more quickly, with significantly better quality 1. On a task chosen to lie outside AI's capabilities, the people using AI in that experiment were less likely to produce a correct solution than those without it 1. The researchers call this a "jagged technological frontier": AI improves performance on some tasks and worsens it on others, even within the same workflow and at a seemingly similar level of difficulty 1. We read that as a reason to start with a task you know well enough to judge line by line.
Example. In Anthropic's June 2026 Economic Index report, the most common outputs of chat and Cowork conversations were explanations (17%), documents and reports (15%) and guidance (11%) 2. We read that as a sign that documents are everyday work for the tool, not a stretch.
Step 2: Write down how you make it now, before Claude sees it
Do this. Write the steps you follow today as a numbered list, then add one sentence that says what the finished document must contain and who will read it.
It worked when Someone else in your business could make a passable version from your list and your sentence alone.
Why. Our rule of thumb is that Claude speeds up a process you already have; it doesn't give you one. One person put it this way: "The single biggest mistake people make with AI is trying to use it as a magic button for execution before they’ve actually mapped out the underlying principles and manual steps of the system they are trying to automate". Our rule of thumb is that if you can't describe the finished result in one sentence, Claude can't reach it either.
Example. Anthropic economists analysed about 400,000 Claude Code sessions from about 235,000 people between October 2025 and April 2026 3. In those sessions, people made about 70% of the planning decisions, meaning what to do, and Claude made about 80% of the execution decisions, meaning how to do it 3. The researchers rate expertise per task: an accountant who has never used Python but tells Claude exactly which reconciliation rules a script must enforce counts as an expert at that task 3. In our view, your numbered list and your one sentence are that kind of instruction for your own document.
Step 3: Give the task its own project, filled with your past work
Do this. Create one Claude project for this document, upload your last three finished versions and the source files they came from, and add your numbered steps and your one-sentence finish line as the project's context.
It worked when You can open a new chat in the project and ask for this job's version without explaining your business first.
Why. As of November 2026, Anthropic's help page describes a Claude project as a self-contained workspace with its own chat history and knowledge base, where you can upload documents and provide context 4. As of November 2026, Anthropic's help page says projects are available on free Claude accounts too, where a user can create up to five projects 4. As of November 2026, Anthropic's help page says Claude accepts PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB, JSON and XLSX files, and that Excel files need code execution and file creation switched on first 5. In our view, your past versions do the teaching here: they show Claude how your business actually writes this document, so you stop re-explaining it in each new chat.
Step 4: Make Claude ask its questions before it drafts
Do this. In a new chat inside the project, upload this job's raw inputs and ask Claude to list the questions it needs answered before drafting; answer them, then ask for the draft and ask it to check the draft against your written steps before it shows you.
It worked when Its questions bring up at least one thing your project was missing, and you add that answer to the project's context.
Why. In our view, its questions show you the context you forgot to give, so the first draft starts closer to what you need. Our rule of thumb is to tell it what the document is for and why it matters, rather than dictating each line. In our view, asking it to check its own work first catches some slips and makes your review shorter.
Step 5: Check every number, name and date against your own files
Do this. Read the draft line by line against your source files, correct it in the chat with a one-line reason for each change, and write down the minutes you spent fixing.
It worked when Every number, name and date in the final version traces back to one of your files, and the fixing minutes are written down.
Why. In our view, the uneven results from step one are the reason to read every line: a draft can be wrong on a task that looks like the ones it got right. Our rule of thumb is to give a reason with each correction, so you can move it into the project's context later.
Example. In June 2023 US District Judge P. Kevin Castel fined the law firm Levidow, Levidow & Oberman $5,000 after its lawyers submitted non-existent court opinions with fake quotes and citations created by ChatGPT 6. Judge Castel wrote that the lawyers "continued to stand by the fake opinions after judicial orders called their existence into question" 6. In our view, the lesson for an owner is the check, not the tool: nothing in a draft goes out until you have traced it to a source.
Step 6: Run it three times, then keep it, rewrite the context or drop it
Do this. Run the task on your next three real jobs of this kind and move any correction you make twice into the project's context. After the third run, keep the task if fixing took less than half the minutes the document took you by hand. If it didn't, rewrite the project's context once, and drop the task if the next run still misses.
It worked when You have three fixing times written down and a keep, rewrite or drop decision on paper.
Why. Our rule of thumb is that a correction you give twice belongs in the project's context, not in the next chat. Anthropic introduced Skills on October 16, 2025, describing them as folders of instructions, scripts and resources that Claude loads only when they are relevant to the task 7. In its October 2025 Skills launch post, Anthropic says its "skill-creator" skill asks about your workflow and builds the skill for you, with no manual file editing required 7. As of November 2026, Anthropic's Skills post says Skills are available to Pro, Max, Ultra, Team and Enterprise users 7. In our view, a skill is a step for after the third run, once the instructions hold still, not before.
Example. Analysing 100,000 Claude conversations, Anthropic estimated the tasks would take about 90 minutes without AI and that Claude cut that time by about 80%, and says these estimates may overstate the effect because they leave out time spent outside the chat 8. That caveat is why we built this step around your own fixing time instead of a general figure.
Step 7: Build a small tool around it only after the document works
Do this. When the third run passes your test, and only if you want the task to run from a folder rather than a chat, open Claude Code in the Claude desktop app. Ask it for a plan for a small tool that turns a folder of inputs into this document, answer its questions, and approve the plan before it builds anything.
It worked when The tool turns the inputs into the same document the project gave you, and you still read each one before it goes out.
Why. As of November 2026, Anthropic's documentation describes Claude Code as an agentic coding tool that reads your codebase, edits files, runs commands and works in a terminal, an IDE, a desktop app or a browser 9. As of November 2026, Anthropic's documentation says the Claude desktop app includes Claude Code, so no separate install is needed, and that using Claude Code there requires a paid subscription 9. Our rule of thumb is to ask for the plan first, answer its questions and approve the plan, and only then let it build.
Example. Anthropic says its own Legal team built prototype "phone tree" systems to help colleagues reach the right lawyer, an example of a department building a tool without traditional development resources 10. Anthropic says its Growth Marketing team built a workflow that reads CSV files with hundreds of ads, finds the underperformers and writes new variations within strict character limits 10. In our view, both examples are narrow tools for one job, which is the size to aim for on a first build.
| what it is | plan it needs (as of november 2026) | |
|---|---|---|
| a project | a self-contained workspace with its own chat history and knowledge base | free accounts can create up to five |
| a skill | folders of instructions, scripts and resources Claude loads only when relevant | Pro, Max, Ultra, Team or Enterprise |
| Claude Code in the desktop app | reads your codebase, edits files and runs commands | a paid subscription |
Knowing the task can matter more than knowing how to code
The Claude Code study from step two also compared who succeeds. In sessions that produced code, people outside software jobs reached verified success 29% of the time, against 34% for people in software-related jobs 3. Each of the ten largest occupations in the data landed within seven percentage points of software engineers on success in code-producing sessions, and management occupations scored highest on verified success 3.
Sessions rated novice reached verified success 15% of the time, against 28% to 33% for sessions rated intermediate or above 3. 19% of sessions where the user appeared to be a novice ended abandoned, meaning they failed with no code written, against 5% to 7% for everyone else 3. The authors conclude that the ability to steer Claude comes more from command of a domain than from the ability to write code, and they call their findings preliminary 3. We read that as the answer to the developer question at the size of one task: for a document or a small tool you understand end to end, your command of the work counts for more than code. One owner put it this way: "I'm a business owner and I am not a software developer, but I've already written apps and scripts that save me on average 15 hours a week".
You ask an empty chat for the document and rewrite what comes back, or you wait for a developer to build a tool.
You keep one project per recurring document, filled with your past versions and written steps, check each line, and build a tool only once the document works.
In our view, depth on one real job beats a tour of every feature, which is why this playbook stays on one document. In our view, this is a first move from asking a chatbot now and then toward daily use for your work. Building and running an automated workflow yourself is the level after that.
Two more studies found AI help raised productivity, and in one, less experienced workers gained more
In a 2023 experiment published in Science, Shakked Noy and Whitney Zhang gave writing tasks to 453 college-educated professionals, and with ChatGPT the average time taken fell by 40% and output quality rose by 18% 11. In a study of 5,172 customer support agents, published in The Quarterly Journal of Economics in 2025, Erik Brynjolfsson, Danielle Li and Lindsey Raymond found an AI assistant raised issues resolved per hour by 15% on average and by 30% for less skilled and less experienced workers 12. We read the support-agent result as a hopeful sign for owners new to the tool, though that study measured experience at the job, not at the tool.
On its small business page, Anthropic quotes Cameron Pappas, an owner: "I've never coded anything in my life. Our design team is more efficient, our buying is better, and the customer is a lot more likely to get what they ordered." 13. In our view, a vendor's chosen customer quote is a claim to test on your own work, and the fixing-time test is how you test it.
Generic drafts, client data, cost and setup are the limits to settle before you start
One person put the worry this way: "If every business uses AI to write their ads, everything will start to sound exactly the same. How can you make AI write something that sounds unique to your brand?" One person who runs ad copy on client accounts answered it: "The sameness comes from prompting off a blank slate. If you open an empty chat and go write me five ad variations, it gives you the average of the internet, because that is all it has to work from." The same person on what changed: "So I stopped asking the model to invent and started feeding it the brand's own stuff instead." In our view, step three answers that worry: the past versions in your project are your business's own material.
Since August 2025, users of Claude's Free, Pro and Max plans choose whether their chats can be used to train new models, and they can change the setting at any time 14. Under that policy, Anthropic keeps new or resumed chats for five years if the user allows training, and for 30 days if not 14. The change does not apply to the Team and Enterprise plans or the API, which run under Anthropic's Commercial Terms 14. Before you upload client files, check which terms your plan runs under; on Free, Pro or Max, choose whether your chats may be used for training. If a client contract limits where its data may go, read it before you upload anything.
On November 1, 2026, Anthropic listed Claude Pro at $20 a month when billed monthly, with Claude Code and Projects included in the plan 15. Projects work on a free account, as step three showed, so you can run steps one to six before you pay. Claude Code in the desktop app needs a paid subscription, as step seven showed.
One person new to it said: "NOT Boring - but a heck of a lot to take in for an 'old' newbie." When a word or a screen stops you, ask Claude to explain it and what to do next. Stay on the one document until it passes the fixing-time test before you learn another feature. Don't hand over a task you can't judge line by line yourself.
Within seven days you can run one real document through the first five steps
Give yourself seven days and one document. The checklist below covers steps one to five. Steps six and seven come with your next real jobs of the same kind.
your first run, within seven days
You now have one task working, a test to keep it honest, and an answer on the developer question
You started without code and with a question about what Claude can actually finish in your business. You now have the seven steps, which take one document from your own files to a finished, checked version. You have the fixing-time test: three timed runs and a keep, rewrite or drop decision. And you have a straight answer on whether you still need a developer: for one task you understand end to end, the Claude Code study points to command of the work over code. In our view, a tool that grows past one task brings the question back, and it deserves its own decision.
Read before: How to Build Systems in Your Business That Your Team Actually Uses. Read next: Are the AI Agents in Your Business Useful, or Just Busy?, for when the task becomes an agent that works on its own.
Where does your own use of AI sit today? Take the Scorecard and find your row.
Sources
- Navigating the Jagged Technological Frontier · Harvard Business School Working Paper 24-013 (Fabrizio Dell'Acqua et al.), 2023; published in Organization Science, 2026
- Anthropic Economic Index report: Cadences · Anthropic Economic Research, June 26, 2026
- Agentic coding and persistent returns to expertise · Anthropic Economic Research (Zoe Hitzig et al.), June 16, 2026
- What are projects? · Anthropic help center, accessed November 1, 2026
- Upload files to Claude · Anthropic help center, July 23, 2026
- Phony ChatGPT Brief Leads to $5,000 Fine for NY Lawyers · Bloomberg Law (Bob Van Voris), June 22, 2023
- Introducing Agent Skills · Anthropic, October 16, 2025
- Estimating AI productivity gains from Claude conversations · Anthropic Economic Research, November 25, 2025
- Claude Code overview · Anthropic documentation, accessed November 1, 2026
- How Anthropic teams use Claude Code · Anthropic, July 24, 2025
- Experimental evidence on the productivity effects of generative artificial intelligence · Science (Shakked Noy and Whitney Zhang), July 14, 2023
- Generative AI at Work · The Quarterly Journal of Economics (Erik Brynjolfsson, Danielle Li and Lindsey Raymond), May 2025
- Claude for small business · Anthropic, accessed November 1, 2026
- Updates to Consumer Terms and Privacy Policy · Anthropic, August 28, 2025
- Pricing · Anthropic, accessed November 1, 2026
