Brett McCully

Brett McCully

@bmccully.bsky.social

Engineer/Manager. ~25 years in the tech industry. Opinions about everything, many of them wrong.

206 followers87 following640 posts

Debate about abusive Claude input is silly. It isn't because of moral judgement, AI sentience, or a vengeful future superintelligence. Toxic input leads to toxic output that gets bounced by toxicity filters. They need to reprocess and can't charge for the tokens. It's money, plain and simple.

People still haven't learned that the "your name is your grandfather's first name and X" threads are easily data mined for security question answers. Sharing just this, and another piece on a Facebook thread, and other places online, all gets correlated together. OSINT tools make it easy.

It's temporary amusement for revealing something that will be online forever. Years from now, you won't remember you posted it, but you'll use it as a password recovery question for a sensitive account.

Claude Code with Fable has hung the main thread multiple times when subagents were completed, but did not return. Fable also tends to create scripts in tmp directories and asks permission to run them, rather than multi-step shell commands. It's less transparent/observable.

Decades of building, deploying, and supporting tech products gives me greater appreciation for the Human Computer Interaction courses in college. The way you think humans should behave is often very different from how they actually behave. It's been true for so much more than interface design.

Changed my handle and profile details to my real info. While I prefer to keep my data and info off the internet, there are real consequences with credibility when choosing to stay anonymous/psuedonymous.

I've had to re-authenticate with Claude 3 times in the last 24 hours. They were in different projects on the same host, but it would not previously require another login. Either behavior has changed, or something is broken.

I wrote a Power Usage calculator because I wanted to understand the real cost of my home lab and other electronic devices. Seattle gives the option to move to time of use billing, and I wanted details. . . . and then there was even more scope creep, so I built: powerusage.app #buildinpublic

Started using Codex as a fall back for Claude quota exhausted . . . not a fan. Default permissions are WAY too strict for command line usage. Workspace write is still too strict. It does not allow granting ongoing permission for headless browsers for testing.

No, I don't want to grant permission every time you test a change you just made. The alternative is to grant yolo access, which it's not getting on my laptop. In a sandbox, the tokens kept getting invalidated. Codex has serious friction issues. It's not catching up to Claude any time soon.

Did Anthropic nerf subagents? Maybe to give them more wiggle room with capacity? Despite explicit requests to parallelize with subagents, Claude is still executing serially.

Related new behavior, Claude now hangs if a subagent does not have permission. The error saying it hasn't been granted permission to perform an action (like rm) doesn't always get surfaced to the main agent.

Yes, the researchers implemented negation language, but LLMs do not evaluate everything they say for possible negations. They are not a database with explicit pointers.

Is it Hermes or GPT 5.5? The combo strongly prefers to use convoluted approaches to solving easy problems. Instead of using command line utilities like curl, it writes python with httplib. Instead of tea to interact with forgejo, it uses curl. It does it even after preferences are in memory.

There were jokes about Code Janitor becoming a role because of all the vibe coding. While not a real title, the function is definitely real. I just joined a project that was vibe coded by a PM. The codebase doesn't need to be cleaned up, it needs to be burned to the ground.

Claude Code recent planning changes make yolo mode less useful. Without going to planning mode, it now stops you to ask if you want to continue to use bypass permissions, or clear context and bypass permissions. I use it so more can be done AFK. Stop asking!

Claude Opus 4.6 solutions are becoming "over elegant." It just gave a ~20 step plan for accomplishing a goal for running SSH commands locally to execute on a remote host. Bruh, there are already built in features (alternate fail2ban jail).

Dark pattern when claiming Opus 4.6 credits: Requires enabling extra usage when on a plan. People are going to forget they turned it on and likely continue paying for more than their subscription. At least it defaulted to Auto-reload off.

Asking Claude Code with Opus 4.6 to tackle a task list with subagents/teams in the backgrounds was FAST. About a month ago, building an app with separate task lists was mostly serial, even when asking to background. It's actually executing in parallel now.

Antigravity still has a lot of UX issues. In the conversation pane "Accept" changes to a file, and "Accept all" still above the text input, despite all changes being accepted. Git commit at push requires too many clicks to generate a message and accept. It's a single command in Claude Code.

@hboon.com I ended up trying the Ralph loop with recommended agent definitions directly within Claude Code, but it still did not continue to run. I feel back to the command line loop calling claud -p "/ralph-loop" and it ran much longer.

Prior to running the loop, I spent a lot of back and forth developing the specs. What I hear is that Codex is much better at long runs when you are able to created detailed spec up front. Claude tends to be better when specs are evolving with the project.

Antigravity told me "run these git commands in the terminal" to finish configuring a new repo. Correct, but still copy/paste. Claude Code asked me "would you like me to run these commands for you?" Still safe, but a better user experience.

I've implemented personal projects, each with a single tool for the entire project. Claude was more helpful and capable of operating on its own in a trustworthy way.

Implementing LLM based functions with each: Claude created a .env file to store API keys, updated the .gitignore appropriately, and created env.template to check in. Antigravity put keys in llm_config.yaml without .gitignore changes. If you want your API keys to leak, use Antigravity.

1. Do a Google Search 2. It brings you directly to the Gemini AI 3. Ask How to turn off AI search 4. Get a wrong answer 5. Switch the default search to Duck Duck Go

Claude: Here's a detailed plan for what you just requested Me: Looks good, implement it <steps away for lunch> Claude: Here's extra detail on what you already approved. What do you want me to to? (I want you to stop halting your work to ask redundant questions) 🤦‍♂️