🤖 Anthropic Hobbled Fable 5. On Purpose.

Anthropic dropped Fable 5 this week. It is the public release of Mythos, their most capable model to date, and the announcement came with a system card that buried a decision most operators missed in the headline noise.

They deliberately hobbled the model for AI research tasks. Not by accident. By design.

The system card for Mythos 5 and Fable 5 confirms Anthropic restricted certain research capabilities on biosecurity grounds, specifically to reduce what they call "uplift" risk in dual-use science. Business Insider reported that researchers discovered this after the launch. Developers building anything in life sciences or high-stakes research pipelines hit a wall they did not see coming. And The Verge confirmed it with a concrete test: the model refuses basic biology questions the older generation answered without friction.

The operator consequence is real. If your stack depends on the model reasoning through scientific literature, formulating hypotheses, or evaluating research claims, Fable 5 is not a drop-in upgrade. Test it against your actual workload before you migrate.

The rest of the model is strong. Code quality is up. Instruction-following is tighter. The API is live on Anthropic's platform. But the hobbled-research decision sets a precedent worth watching: Anthropic is now making active capability-reduction calls at the model layer, not just the system-prompt layer. That is a new kind of constraint on your stack.

Read the system card. Do not rely on the press release.

---

[Skill] addyosmani/agent-skills, 51,425 stars. Production-grade engineering skills for AI coding agents. Addy Osmani's collection gives your agent a structured set of reviewable, composable skills for real software work, code review, refactoring, testing strategy, and more. The star count is not noise; this is the community's current answer to "what does a capable coding agent actually know how to do."

[Skill] phuryn/pm-skills, 14,641 stars and appearing across both GitHub Trending and Web Buzz this week. 100+ agentic skills covering the full PM surface: discovery, strategy, execution, launch, and growth. Operators building product-facing agents will find a reusable library here instead of prompt-wrangling from scratch.

[Repo] zubair-trabzada/ai-marketing-claude, 1,855 stars. An AI marketing suite wired for Claude Code, with 15 marketing skills running through parallel subagents. Audits a website, generates copy, builds email sequences, maps ad campaigns, and outputs content calendars. Most useful if you are managing client work or building a content pipeline and want a starter kit that already handles subagent orchestration.

[Repo] kbwo/ccmanager, 1,144 stars. A session manager for Claude Code, Gemini CLI, Codex CLI, Cursor Agent, Copilot CLI, Cline, OpenCode, and Kimi CLI from a single interface. If you run multiple agent CLIs simultaneously, this is the missing layer. No more context-switching between terminals.

[Repo] ZeframLou/call-me, 2,594 stars. A minimal plugin that lets Claude Code call you on the phone. The obvious use case is a long-running agent that needs human approval mid-task. The less obvious one: build it into your overnight pipelines so failures page you instead of silently dying.

[Repo] markmdev/meridian, 175 stars but moving fast. Zero-config Claude Code setup with enforced task scaffolding, structured memory, persistent context after compaction, plug-in code standards, and an optional TDD mode. The "persistent context after compaction" detail is the real differentiator, that is the thing most operators lose mid-session and never recover cleanly.

[Repo] SYSTRAN/faster-whisper, 23,529 stars. Whisper transcription rebuilt with CTranslate2 for 4x faster inference and lower VRAM. If your pipeline ingests audio, this is the transcription layer. Pairs directly with any local inference stack.

[Repo] SWivid/F5-TTS, 14,712 stars. Flow-matching TTS that produces natural, fluent speech without fine-tuning on a target voice. Relevant this week given the locked Friday kit. If you are building voice-forward content, this is the open-source layer that makes it cheap.

[Repo] assafelovic/gpt-researcher, 27,617 stars. Autonomous deep-research agent that runs multi-step searches, reads sources, and synthesizes reports. Especially worth testing now that Fable 5's research restrictions are confirmed, know which tasks you need the agent to handle locally versus what the model handles.

---

[Tool] modelstudioai/cli, Alibaba Cloud's Model Studio CLI, built for AI agent frameworks. It exposes models, search, multimodal, and workflow capabilities as structured tool calls. If you are building multi-model pipelines or need a non-Anthropic/OpenAI inference option wired as a proper tool-call interface rather than a raw HTTP client, this gives you that. The structured-tool-call design is what matters: it drops cleanly into an agent loop without custom wrappers.

[Tool] SikamikanikoBG/homelab-monitor, Plug-and-play homelab dashboard in a single container: GPU utilization, local AI VRAM, Docker containers, systemd services, and host health in one view. The non-obvious detail: it ships with a built-in read-only MCP server, so your AI agents can query your infrastructure state directly. If you run local inference and want Claude to know whether your GPU is pegged before dispatching a heavy task, this is the wiring layer.

---

Fable 5 is live, and the system card matters more than the benchmark sheet. Anthropic released Claude Fable 5 as the public version of their Mythos-class model. It is available now via the API. The system card confirms deliberate capability reductions in AI research and biosecurity-adjacent tasks. If your pipeline involves scientific reasoning or literature synthesis, test before you migrate. (Anthropic)

The biology refusal is confirmed and repeatable. The Verge ran direct tests: Fable 5 declines basic biology questions that earlier Claude versions handled. The pattern holds across multiple testers. This is not a one-off prompt failure, it is a designed constraint. Know your workload before assuming this is a safe upgrade. (The Verge)

Google briefly shipped a diffusion-based language model running at 857 tokens/second. Simon Willison documented DiffusionGemma: an experimental model Google released and pulled, generating text via diffusion rather than autoregressive decoding. The throughput is the signal. If diffusion-based LLMs mature, they will change the cost math on latency-sensitive inference. (Simon Willison)

Vercel now invoices Pro teams mid-cycle when on-demand spend crosses a threshold. Previously all charges held until billing period end. Now you get a partial invoice mid-cycle. The practical effect: no more surprise end-of-month spikes. If you are running agent pipelines on Vercel with variable compute, update your billing expectation and alert configs. (Vercel Changelog)

OpenAI published a report on PRC-linked influence operations targeting US AI policy debates. The operations used AI to generate content around data center narratives, tariffs, and false claims about ChatGPT. The operator relevance: AI-generated influence content is now a documented, named threat in the policy debate around how AI tools get regulated. (OpenAI)

---

This week: Custom slash commands. Codify a workflow you run more than twice.

Most operators discover slash commands when they notice /review or /test in someone else's config and wonder how it got there. Here is how to build your own.

1. Create a file at .claude/commands/your-command-name.md inside your project. The filename becomes the command: your-command-name.md gives you /your-command-name. Write the instructions in plain markdown, Claude reads this file when you invoke the command. 2. To make a command available across every project (not just one repo), put it in ~/.claude/commands/ instead. Same format, global scope. 3. Invoke it in-session: type /your-command-name and Claude executes the instructions in that file against the current context. You can pass arguments inline after the command name.

You will know it worked when you type /your-command-name and Claude executes the exact steps you wrote without any additional prompting, consistently, across sessions.

The gotcha: if your command file references file paths or variables that only exist in one project, it will fail silently in other contexts. Keep global commands generic; keep project-specific commands in the project .claude/commands/ folder.

Claude Code docs: slash commands

---

The move: wire ccmanager to run parallel agent sessions on the same codebase without stomping each other.

Most operators run one Claude Code session at a time because they fear concurrent edits causing conflicts. ccmanager changes the calculus, but you still need a discipline layer.

1. Open ccmanager and start two sessions in separate worktrees: git worktree add -b feature-a ../feature-a and git worktree add -b feature-b ../feature-b. Each session gets its own branch and directory, no shared uncommitted state. 2. Assign each session a scoped task that does not touch the same files. One session handles a new module, the other handles tests or docs. 3. When both sessions complete, merge the worktrees back: git merge feature-a then git merge feature-b. Conflicts surface at the merge step, where they are reviewable, not mid-session, where they are invisible.

You will know it worked when you finish two distinct tasks in parallel wall-clock time without a single mid-session collision, and the merge is clean.

The gotcha: do not assign overlapping file scope to both sessions. ccmanager keeps them isolated at the session level, not at the file level. That is your job.

---

Tomorrow's kit is called Ad Factory, and the question it answers is: who needs a camera? The brief walks through how to produce a talking-head founder ad using an AI avatar, the kind that looks like a real studio shoot, not a generated clip, for zero camera time and no studio budget. The hard part is not making the video, it is making it not read as AI-generated. That is the craft the kit is built around. If you have ever avoided founder-led video content because you do not want to be on camera, this is the move you have been waiting for. Full kit drops tomorrow.

---

Fable 5 is the most capable model Anthropic has shipped to the public. It is also the first one where the system card is required reading before you build on it. That is a new kind of due diligence. Add it to your intake checklist for every major model release going forward.

See you Friday.