As a developer, debugging the issue had become lot more easier than ever. AI is doing fantastic job for initial setup and smaller projects, but as the projects grow we can see the real pain in the AI usage. It really does many mistakes and it is struggling to find logic in many scenarios. Also in a single chat, hallucinations are extensive than a fresh chat. But fresh chat do not have a context, so planning and chat compacting is useful there.

Clarity and context are key ingredients for successful use of AI. If you provide incorrect or incomplete details or try to hide any details and assume that Claude or ChatGPT can do magic for you, that’s untrue. Context and clarity in your questioning matters. Every tool hallucinates, fabricates and whatnot - so data needs to be factual, real time and accurate!

If I run a session of Claude Opus long enough (and other models too) sometimes it forgets valuable context that was provided near the beginning of the session. This requires me to often re-explain project requirements or direct it to re-read files it has already read.

Accelerated development using AI tools like claude, copilot etc Hallucinations are sometimes a problem. I make sure check whether the responses are correct & can be properly applied to the prompted question.

If Civil Engineers built buildings just like Software Engineers built Software, the end of civilization would have happened the first time a woodpecker landed on a building.

We are using Ai in our company but we are using as a tool rather then handing off the full control to the AI. We have to make sure that when we are using AI we have to make sure it follows regulation and given instructions to make sure it doesn't go out of scope and take its own decisions. This is how Ai should be used. To improve our coding not to hand off complete workflows to it.

I have found AI to often forget previous relevant prompts / answers and to therefore be inconsistent in the methods that it will try to answer a question. Additionally, I have noticed that when it is over-utilized in the workplace that there are gaps in understanding between workers, therefore somewhat inhibiting communication of projects. I think that it would be beneficial if AI could enhance productivity without creating gaps in understanding, communication. I think a way that it could do this is by acting more as a tutor - forcing the individual to engage or ponder about how to do something more than simply forcing an answer with little explanation or questioning. I think if it acted to teach people on why / how to do things, then people would also be better able to communicate the work that did alongside AI.

I was building an embedded system that was to be implemented using direct register access. The AI kept including high-level libraries, or forgetting design details.

When studying for my linear algebra final exam I ran into a problem where claude would hallucinate the wrong answer. It would first give me the wrong answer to the questions I had then midway through its own explaination realize its wrong, then correct itself.

When I needed help with chemistry homework it would often get questions wrong so I’d learn the wrong way to solve them.

Using LLMs to draft risk statements for risk registry. Goods: Great at taking raw notes and transforming them into structured, clear and standardized risk statements in seconds. Bads: It loves to invent polished sounding mitigations for risks that don't exist in our actual workflow. Works amazingly as a drafting copilot, but terrible as an human review is non negotiable.

Connected the Claude GitHub app to my repo, added necessary configuration, and just mention [handle removed] in a comment — it starts doing exactly what I asked, creates a PR, and puts the results right there with the PR comment. You can also highly customize it — like triggering it differently when a PR is raised, controlling what it does at each event, etc. Great approach overall. One downside: it only checks the code, not actually run it. Developer still has to manually run and test the PR themselves before merging — Claude doesn't verify the code actually works, just that it looks right.

I was trying to do some research with Claude on a specific company. Claude tends to make a lot of assumptions and extrapolate based on the information available online. It will make claims that can be assumed but not proved.

I am building several agents to access the patient chart notes and understand the problem and the procedure performed. The agents are better in identifying the problem reported and procedure performed or suggested. The other agent that is supposed to identify ICD10 and CPT codes for the problem and procedure does a decent job but fails to understand the inclusion and exclusion when multiple ICD10 codes are present. This could be a Business logic problem, but I expect AI agent to deduce the codes considering the overall scenario, not specific problem or procedure

I was building an insurance agent and configured the prompt to return benefits and claims info for insurance holders. While testing, I frequently saw the agent hallucinate claims and charges.