My experience with Claude in daily operation is this: Claude delivers reliably when three things are in place. Every number has to come from a source that sits in the folder, a separate review role reads the result before I see it, and whatever really must not happen is blocked technically instead of just forbidden. Where one of these is missing, the same mistakes keep happening: Claude fills gaps with plausible sentences, copies numbers from search summaries, reads only the start of a file, or follows a rule more literally than it was meant. This is not from a survey. It is in the learnings files of my AI workforce: over 500 entries since July 2026, almost every one with a trigger, a rule and an example. All the experiences below come from those files or from job folders I can check.
Why I work with Claude and not with ChatGPT is covered in Claude Over ChatGPT: My AI Employees. Who does which job in my workforce is shown in My AI Workforce: Fourteen Employees, One Boss. This article is about what goes wrong in everyday use and what I made of it.
How I record experience
Each of my AI employees has a file called wissen/learnings.md. When I give feedback or a review finds an error, it becomes an entry with a fixed structure: the rule in one sentence, why (the concrete case with numbers), how it is applied verifiably in the next job, and whether it is confirmed or still a hypothesis. Before any work, the employee reads this file. That way a mistake turns into a rule that applies to every future job, instead of me giving the same hint again every week.
It is also the reason I can write this article without relying on my memory.
Experience 1: Inventions are more likely in the connecting sentences than in the core content
When Ralf, my website optimizer, rebuilt a client's website from its existing texts, all master data, prices and quotes were correct. The separate review still found five errors, and all of them were in sentences Ralf had written himself as transitions: a response time, a hosting claim, an overly strong legal statement. None of it was in the sources, and none of it went live. It just sounded plausible.
The rule: Every sentence that does not come from the documents is a claim. It gets backed up, softened or deleted. This is confirmed so far only in this one case; Ralf treats the rule as a working hypothesis. How Ralf works is described in Build and Modernize a Website With AI.
Experience 2: Numbers from search summaries are not a source
Peter, my employee for receipts and bookkeeping, needed an official exchange rate for December 2025. The web search gave 1.1732. The PDF from the German Federal Ministry of Finance said 1.1709 for December; 1.1732 was the value for September. With 25 converted receipts, the wrong rate would have distorted the annual total.
The same pattern shows up with truncated output: an evaluation script of my lead scout printed addresses in shortened form, and when the lead list was written, the missing street names were guessed. One street became another. It was caught in the lead list before any of it reached a website; there, a wrong address would have been real damage.
The rule: Numbers and facts come from the original document or the complete source file, never from a summary or a shortened screen output. For official values, add a second independent source.
Experience 3: A separate review role finds what the author does not see
The most important rule of my workforce: whatever leaves the house is checked beforehand by a separate role, not by the same context that wrote it. The learnings show again and again how much this brings. Sebastian, my AI employee for SEO, explicitly adds this in such cases: “The review role found it, not me.”
One example from September 2026: for 19 blog posts in German and English, 38 files in total, Sebastian had run four rounds of his own editing and one review pass. An additional cross-check with a different tool, Codex, still found four posts that needed mandatory corrections, including an overly narrow definition of a Kubernetes term and the false claim that Ollama cannot process requests in parallel. Spot checks in the official documentation confirmed the findings.
The review role is not infallible either. Conny, my chief content editor, noted in July that the review pass wanted to delete an example because it had looked in the wrong file. Since then she checks the review's findings against the source before she accepts them.
The rule: Author and reviewer are separate, and the reviewer's findings are also checked against the source.
Experience 4: Claude follows rules literally, even when they were not meant that way
Sebastian had a check script that required the main keyword to appear in full in a heading. For the keyword “ki agent was ist das” (German for “what is an AI agent”), it also required “was”, “ist” and “das” in the heading, and one heading read like a raw search phrase just so that it would pass the check. Similar with a click-through-rate rule from a social media video: applied literally, it would have selected 399 of 568 pages of my website for rework.
The rule: If a text sounds unnatural just to pass a check, the check is wrong, not the text. And rules with fixed thresholds are measured against my own data before they apply.
Experience 5: A check that reads nothing reports “all fine”
Eddi, my videographer, checks finished videos with text recognition for wrong times in on-screen captions. His first check script reported “0 stray times”. In fact it had not read any images at all, because the command that extracts frames silently returned nothing. He only noticed because he ran the script against the old version first, in which the error demonstrably existed. With corrected frame extraction and a count check: old version 34 misreadings, new version 0.
The same class of error showed up with Conny: a video draft counted websites from screenshot files and would have claimed far too many of them were live. It was caught in the cross-check of the draft, against the list that tracks the actual state.
The rule: Every check counts what it has read, and runs once against a version in which the error is known. A file count is not a statement about state.
Experience 6: Reading only the start reports an outdated status
Kai, my AI employee for management, reviews the reports of the others. Once he reported a page as “not deployed”, although it had been live for five days. The report documented this in rounds 4 and 5; only the line in round 1 was outdated, and that was exactly the line read by the subagent doing the review for Kai.
The rule: Before any statement “not live”, read the report to the end and check the live address. Subagents are explicitly told to read reports in full and check the live state.
Experience 7: Subagents inherit what is in the folder
For measuring whether AI searches cite my website, Sebastian had measurement agents research questions. One wrote in its answer: “The source kevinwelter.com belongs to the operator of this system”. Subagents that start in the project folder load the project instructions and therefore know who they are working for. For a measurement, that is a confounding factor: the number of citations can be skewed in my favour as a result.
The rule: Measurement agents run in an empty folder without project instructions and only with the tools they need.
Experience 8: What must not happen is blocked technically
A rule in the working instructions is a behavioural guideline. It works most of the time, but not with a guarantee. That is why my permissions are tight: my assistant's mailbox access can save drafts but cannot send. For an SEO data tool whose key has the same rights as the whole account, 28 deleting and account-changing tools are taken out of Claude's reach with a deny rule. And for a comparison test, the MCP tools were removed at startup in the first run, not just forbidden. What this looks like with MCP servers is covered in Set Up a Claude MCP Server.
How important it is to measure your own tools was shown by a check for privacy content: a reviewer captured the real network traffic of Claude Code instead of just reading the docs. Result: “The data does not leave my machine” was wrong, because content that has been read goes to the model as text. That sentence was not allowed on any slide afterwards.
The rule: Technical limits before behavioural rules, and measure statements about your own tool instead of deriving them.
Experience 9: Too many agents at the same time run into the limit
In September, Sebastian ran 16 writing agents in parallel. After 15 minutes the session limit kicked in, and 13 posts broke off half finished. With two to three agents at the same time, on a smaller model and with a queue as a file, 60 posts then ran in a good two hours without a break-off.
The rule: At most three writing agents in parallel, and the work list exists as a file so that an interruption costs nothing.
What works well
That sounds like a long list of failures. But it is the reason the rest works. From 27 August to 21 September 2026, Sebastian and I built kevinwelter.com in three stages from 10 to 224 posts per language, with a measurement after each stage. Before the start, from 1 to 26 August, the website had 5 Google clicks; by 17 September 2026, Search Console showed 1,060. That is a look back, not a promise for other websites. The whole story is in AI for SEO and Blogging: How Sebastian Works. Since October, part of the workforce has been running on its own server and keeps working when my computer is closed. I have not written a line of code myself for a year and a half. I lead.
What always stays with me: the approval for everything that has an effect on the outside. Emails, deploys and ads only go out when I explicitly say so, and the employee records my word with the time.
Frequently asked questions
How good is Claude in everyday use, really?
Good, when it works with sources, review and clear permissions. The errors I see are rarely wrong expertise, but plausible filler sentences, numbers taken over from summaries and rules followed too literally. Fixed rules, added after every error, help against all three.
Does Claude make mistakes?
Yes, regularly, and that is why I have a separate review role for everything that leaves the house. In one case, an additional cross-check among 19 posts found four with mandatory corrections that four rounds of own editing and a review pass had missed. Mistakes are part of it; what matters is that each one becomes a rule.
What is your experience with Claude Code?
For me, Claude Code is the tool that lets the AI employees work in their working folders, start tools and save results as files. It brings the most where a task is measurable. It needs the most attention around permissions, parallelism and the question of what a subagent picks up from the folder.
Is Claude worth it for small businesses?
For recurring tasks with clear rules, yes, if you are willing to write the rules down and check the result. If you only need texts now and then, the chat is enough. What it costs is covered in AI Employees: Tasks, Costs, Getting Started.
How do I stop Claude from making things up?
Not through a request in the prompt alone. For me, every number has to point to a file in the job folder, sentences without a source count as claims, and a separate role checks against the sources. That reduces inventions considerably, but does not rule them out.
Where to go from here
Start with a single employee and give them a learnings file from the beginning. Every mistake you correct becomes a rule there. After a few weeks you no longer have an assistant to whom you explain everything anew, but an employee with experience. How to do that is described in Creating an AI Employee: How to Start. Ready-made employee packages with these rules and the chance to discuss your own cases are available in my community, Claude Practitioners.
