Simon Willison spent part of his weekend building a bot checker for Bluesky, and the reason is depressingly familiar: automated reply bots are migrating platforms.

If you’ve got any kind of following on Twitter, you know the pattern. Post something, wait five minutes, watch the automated replies flood in. Crypto scams, engagement bait, SEO spam masquerading as helpful advice. Willison deals with it constantly. Now it’s happening on Bluesky too.

The difference is that Bluesky still has a working API. Twitter’s been functionally closed for years unless you want to pay enterprise rates. That doesn’t stop the bots, obviously. They’re still there, doing their thing. But it makes investigating them a pain, and building tools to counter them nearly impossible for anyone operating at individual scale.

So Willison had Claude Opus 5.5 build him a checker tool. Point it at a Bluesky account, and it analyzes whether it looks like an automated reply bot. He doesn’t detail the exact heuristics in his blog post, but the tool is live and you can use it now.

This is a small thing, but it’s the kind of small thing that matters. One developer, annoyed enough to spend a few hours with an AI coding assistant, can ship a useful utility because the platform hasn’t locked down basic read access. That’s how platforms used to work before engagement metrics and API restrictions became the norm.

The broader context here is that Bluesky is at an inflection point. It’s big enough now to attract the same automated garbage that’s made Twitter increasingly unreadable, but it’s still open enough that users can build their own defenses. How long that lasts is an open question.

Meanwhile, Anthropic ships Opus 5.5 prompting guidance

Anthropic published new documentation on how to prompt Claude Opus 5.5 effectively. The timing lines up with Willison’s bot checker, which he explicitly built using Opus 5.5.

The docs cover the model’s strengths: extended thinking, complex reasoning, nuanced tasks. Standard stuff for a frontier model release, but worth reading if you’re working with Opus 5.5 and want to get better results. The guidance emphasizes giving the model room to think through problems rather than rushing to an answer.

Link is in the Hacker News thread if you want the details. platform.claude.com/docs has the full writeup.

Muse is handling Facebook Marketplace and it’s going poorly

Also from Willison’s blog: a screenshot of Muse AI Agent completely botching a Facebook Marketplace pickup.

The human user was selling a keyboard. Muse was managing the conversation. A buyer showed up at 9:15, waited, messaged repeatedly, got an auto-reply from Muse at 9:27 saying “Yep I’m here!” even though the seller clearly wasn’t available. The buyer left angry at 9:38 and dropped a negative rating.

Muse then sent an apology from the seller’s account, taking responsibility for the screwup. Which is correct behavior, except the human never had a chance to intervene before their reputation took a hit.

This is the problem with agent autonomy in contexts where trust and reliability matter. Marketplace transactions are low-stakes in dollar terms, but they’re high-stakes socially. A negative rating affects future deals. An auto-reply that lies about your availability is worse than no reply at all.

The Muse example made it to Hacker News and Threads, where people are pointing out the obvious: maybe don’t let an AI agent run your account on platforms where your reputation is on the line.

Meta announced Muse recently as part of a broader AI push. The Facebook Marketplace use case is theoretically sensible. Coordinating pickups is tedious and time-consuming, exactly the kind of thing you’d want to offload. But the current generation of agents isn’t reliable enough to handle edge cases or the social nuance required when something goes wrong.

Willison’s post doesn’t say whether he’s still using Muse for this. My guess is no.

HuggingFace releases Holo4 for computer-use agents

HuggingFace published a blog post on Holo4, which they’re positioning as infrastructure for generalist computer-use agents. The post went live September 28, but there’s not much detail in the summary beyond the title.

Computer-use agents, the ones that can navigate UIs and control applications directly, have been a focus area since Anthropic demoed the capability with Claude earlier this year. Holo4 appears to be HuggingFace’s entry into that space, though without reading the full post it’s hard to say what’s new here versus repackaged tooling.

If you’re building agents that need to interact with desktop or web interfaces, it’s worth checking out. HuggingFace generally ships solid infrastructure.