Claude is most useful to me when I hand it a small job with a visible finish line. “Write my marketing” leaves a lot of room for vague prose. “Read these three sample questions and outline a guide that answers them in order” gives me something I can inspect. Here are ten jobs I’d try before asking any AI tool to run a campaign. Use public or permissioned material, take private details out of your examples, and keep responsibility for what you publish.
1–2. Organize research and define the job
1. Sort research notes by question. Paste your own notes or permitted excerpts and give them short labels. Ask: “Group these notes by reader question. In each group list the source labels and the facts still missing. Leave uncertain items in a separate list.” Compare the resulting groups with your source material. Look for claims the tool has made stronger than the note supports. This works well when your draft has six paragraphs about one point and almost nothing about the decision readers came to make.
2. Draft a one-page brief. Give Claude a reader, current situation, goal, and constraint. Ask: “Write a brief with the reader, task, required inputs, finished deliverable, and three questions to check before drafting.” For example, your reader already has a signup page, wants a working thank-you page, and has no new software budget. Check that every line respects the constraint. Save the brief once it’s accurate. It can guide both your own draft and later prompts without re-explaining the whole project.
3–4. Build outlines and working examples
3. Outline a tutorial by action. Supply a workflow you’ve verified and ask Claude to turn it into steps with one visible result for each. “Use these seven steps in the same order. Flag any transition you can’t explain with the information provided.” Review the transitions, especially anything involving permissions, redirects, account tiers, and delivery. Run the outline yourself before it becomes a tutorial. If the model invents a settings screen, remove it and show the reader the actual control.
4. Sketch a fictional example with constraints. Say you need a sample newsletter signup. Ask for an invented sender name, a one-page resource, and a three-step delivery plan. Specify that all figures and details are fictional. Then try the plan with a test account and update any steps that differ. A fictional example can make instructions concrete. It can also quietly become a made-up success story if you forget where it came from. Label it in the draft and leave the label in the published version.
5–6. Improve an email and create a comparison
5. Edit a draft without changing its claims. Paste an email you wrote and a short paragraph that sounds like you. Ask Claude to mark long sentences, unexplained references, and places where a reader could mistake an optional offer for a required one. Request suggested replacements in a separate list. Pick the ones that help. If it rewrites “I read the vendor’s guide” into “I tested the product,” put the original back. See my affiliate email guide for the evidence and disclosure to keep in view.
6. Turn facts into a comparison table. Give it verified entries from the tools’ current documentation and name the columns. “Compare signup forms, confirmation, and export. Use ‘check provider’ when a fact is missing. Attach each source label to the fact it supports.” Open the relevant documentation yourself and compare the table to the actual wording. A polished cell is no substitute for a real answer. Record the date you checked the plan details if they affect a buying decision.
Give Claude the materials and a checkable job. Keep the last edit for yourself.
7–8. Repurpose material and find missing steps
7. Turn one guide into a short companion. Give it a guide you own or are allowed to adapt. Ask it to pull out one task and draft a checklist with a yes-or-no outcome for each item. Choose one example that the checklist should work on. Run it with that example and fix the steps the model compressed too far. If you’re adapting licensed material, check the distribution rights first. My PLR rebranding guide shows where a checklist fits in a small editing project.
8. Find places a beginner might get stuck. Give the draft and define a reader who knows very little about the tool. Ask: “List any steps that assume an account, permission, file, or term I haven’t explained. Quote the sentence and suggest one question a beginner would ask.” Take the list to your actual process. If the model raises a useful issue, answer it with an example or link. If it asks about something the preceding paragraph already covers, you can improve the transition or leave the text as it is.
9–10. Generate test cases and audit the handoff
9. Propose test cases for a small workflow. Describe the inputs, the desired result, and which changes are in scope. Ask for a normal case, an error case, and a retry. For a signup page, try a fresh test address, an invalid address, and a previously subscribed test address you control. Record what you observe and decide whether the result matches your intent. Claude can suggest what to try. The confirmation message in your inbox is the evidence the workflow ran.
10. Read a publication handoff as a reviewer. Give Claude your draft, delivery instructions, and a short checklist. Ask it to mark claims that need checking, links with vague labels, and handoff steps whose wording differs from the actual email. Review each mark and make corrections yourself. It may overlook a broken link that looks normal in text, so click the links in the published version. For answer pages, the AEO guide shows how to make the verified details findable.
Make the workflow repeatable
Anthropic’s prompting documentation covers clear instructions and examples. My version is four lines: “Here is the source material. Here is the job. Here is a good example. Here is what to flag instead of guessing.” Choose one of the ten jobs and use that template on a real project you can verify. Save the output and your corrections. The corrections show what you needed to specify and what belongs in the human check.
Test again on a new input before relying on the prompt. If the results vary, narrow the job or add a better example. Treat tool features and plan limits as details to verify in your account. If a workflow needs sensitive business or customer data, assess permission and data handling first and use a synthetic version for practice. You can learn most of the workflow with names and details you made up.
Make one test before repeating the task
Take the comparison example and give Claude two short, labelled facts from a fictional email tool. Leave a third fact out on purpose. Ask it to fill the table only from those notes and mark the missing cell. Inspect the result for invented plan names, confusing units, or a feature mentioned in the wrong column. Make the same test with notes from a different fictional tool. If the model guesses, add an example of a blank cell and say what should appear when evidence is missing.
Do the same with an editing prompt. Give it a paragraph with a deliberately uncertain claim, such as a tool that “may support direct export.” Check that the uncertainty survives the edit. Have the model list changes in a separate section so you can see whether it added a capability. Keep the original paragraph next to the revision. Once you find a prompt that gives you a reviewable output, save it beside the checklist you use to approve the work.
Frequently asked questions
- Is Claude suitable for writing complete sales copy?
- It can help draft and edit. Verify product details, permissions, disclosures, and every factual claim before publishing.
- Can I upload customer data or vendor files?
- Check permissions and the service’s data terms for your account. Use invented examples or remove private details for practice.
- What should I try first?
- Ask it to edit one paragraph for clarity and list the changes. Compare its suggestions with your original before accepting them.