In September, I was working with Claude on a way to pass coding tasks to other AI tools. Early in the work it checked my computer and found no local Devin tool. Devin is a coding assistant, with a local tool that lets me send it work by typing commands.
I installed that tool after the check. Claude kept working from the earlier result and began building a separate link.
The check that went out of date
The first check had been right. That made it harder to spot the stale result inside the later plan, which still gave a sensible reason for building the link. The local tool could already send tasks to Devin. Comparing those features with the proposed link would have exposed the overlap before more code was added.
Following a fix through the site

A September review of my portfolio raised a different question. The home page was requesting project data for a menu image. If that request failed, it could prevent the page from being built. A change caught the error and let the page carry on without the image, but the review flagged that against the project’s rule to fail on missing data.
The later change removed the extra request from the home page. The review showed that the earlier path was gone. That was a finding it could trace in the code.
The next issue concerned the menu image on a direct visit to the home page. Removing the request had left the projects page as the place that supplied the image. A visitor who had already opened that page could have it available in the menu. Someone arriving at the home page first would not have taken that route.
The reviewer found a possible blank image when hovering over the projects link in the menu. It pointed to the code that supplied the image and the code that cleared previews for links without one. The order of the pages was part of the suspected fault, so repeating only one route through the site would leave that question open.
This was a finding from reading code. I had not seen it happen in a browser. The check I wanted next was a direct home-page visit, followed by the same menu check after visiting the projects page. I could then compare what a visitor saw on those two routes.
Reviewing the claim in a sentence
I ask an article reviewer for a similar level of detail about the claim it disputes. I want the sentence and the source used to check it. For a code finding, I want the relevant path and the steps that could bring out the fault.
An early AI-writing sample on this blog claimed that you had to use deep learning, a kind of AI, to make a convincing experience in virtual reality. The sample gave no source for that claim.
To check that claim, I would look for a source showing why deep learning was needed. If I could only find examples of people choosing to use it, I would change the wording about it being essential.
A reason to keep building
When I questioned the plan, Claude checked again, found Devin and agreed that much of the work it had been doing repeated what was already there. I don’t think that is mainly a story about the assistant.
The link was a fair answer to the job I had set, which was to pass tasks to tools from one place. What had changed was the setup. Once Devin could take tasks directly, part of my own commission had gone stale, and a review that looked only at Claude’s code would have been reviewing the answer to a question whose premise had changed.
The portfolio review was good at what it did. It traced a possible fault that depended on the order a visitor took through the site, and it was clear that it had read code and not opened a browser, so the next check was mine to run. The writing sample’s deep-learning claim needs a source, and the search for one is also mine.
In each case the assistant’s work was the easy thing to review. The harder review reaches back to what I asked for, whether it still holds, and what I have left unchecked on my side. When I questioned the Devin plan, that is what the recheck did. It found the tool, and it found that the setup behind my own request had changed before the plan did.


Comments