Claude Code Projects update, AMP rewrite, Blow on AI and software
The issue covers Claude Code Projects updates, a vibecoding rewrite story, a Jonathan Blow video, a No AI Slop manifesto, and new tools and models including Harness Tuner and DeepSeek V4-Pro-0813.
Top stories
Anthropic updates Claude Code Projects with proactive coordinator, thread sharing and reminders

- Anthropic updated Claude Code Projects so the coordinator answers questions and posts thread status updates, and added read-only sharing, reminders, /btw and /fork commands.
- Accounts are limited to 200 new threads per day, and parallel workers draw from the same Claude Code allowance; threads with local computer access cannot be shared.
- The public beta on Pro and Max, not Team or Enterprise, reduces monitoring needed when several cloud sessions work in parallel on separate Git branches.
Developer rewrites nearly half of vibecoded Zig music player AMP after Anthropic account ban

- A developer from Iran describes building AMP, an offline music player in Zig, partly vibecoded with Claude during internet blackouts.
- After Anthropic banned the account and Codeberg adopted a no-AI policy, they read every line and rewrote nearly half the codebase over four months.
- The author now regrets relying on LLMs, would treat them as a fancy IDE feature, and lists a v1.0 roadmap with a music server and Android and Windows support.
Read more at alichraghi.github.io
Jonathan Blow and Wolf Tivy discuss why AI has not changed software in a video

- A YouTube video titled "If AI Solved Coding, Why Is My Software Still the Same?" features Jonathan Blow and Wolf Tivy.
- The 1:40:55 discussion covers the explosion of complexity in software and why today's AI cannot produce better code than people.
In brief
- AIE Code Summit attendees publish No AI Slop manifesto. Speakers and attendees of AIE Code Summit 2026 wrote a manifesto saying AI output must be owned by humans, because agents can produce more work than people can review. (slopmanifesto.com)
- Kiln-AI releases Harness Tuner for eval-driven tuning of agent tools and prompts. The MIT-licensed tool tunes tools, prompts, skills and model settings, and claims a production agent with over 200 tools got 2.5x fewer failures, 6x lower cost and 2x lower latency on held-out data. (github.com)
- DeepSeek releases V4-Pro-0813, a 1.7T-parameter MIT-licensed model. Its Hugging Face model card reports 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE and 74.1 on Toolathlon-Verified, plus DSpark speculative decoding for vLLM and SGLang. (huggingface.co)
- Sentrint, a security scanner for vibe-coded apps, is up for acquisition from $5,102. Sentrint scans GitHub repos for hardcoded secrets, database access rules and vulnerable dependencies, then writes fix prompts for the AI tool used; bidding starts at $5,102 USD. (sentrint.com)
- Hugging Face Harness Leaderboard v0.1.6 re-ranks coding-agent harnesses. The update moves OpenClaw to about 54 on DeepSeek-V4.1-Flash, up from 33 and 38, and drops Hermes Agent from first to ninth after eight harnesses were re-run. (joelniklaus-harness-leaderboard.hf.space)
- Deep Noodle releases decide v0.5.0, a CLI for yes/no-style decision-model questions. The Apache 2.0 CLI and Go library asks a decision model questions about functions, pull requests and tickets, with twelve built-in templates, caching, CI exit codes and a Claude Code plugin. (decide.deepnoodle.ai)
- Author finds Fable 5.2 produced badly wrong running-training model. The author says Fable 5.2 predicted a 1:41 half after a 1:53:54 race and gave a 7:00 mile 79% today and 78% after eighteen months, without checking its work. (corynezin.bearblog.dev)
Compiled, edited and written with AI. Published by Rapid Review, Inc.