Good Morning! Here's what I have for you in today's newsletter:
Replit launches Free Mode, letting you create 30x more without usage costs
Claude Code adds a Concise output style, leading with results first
CopilotKit releases Open Bot, open-source AI coworkers
Turn Claude Code into a system that runs itself with five important rules
4 new AI tools worth trying today
AI SKILLS
🚀 Replit Launches Free Mode
Replit launched Free Mode, powered by OpenAI's GPT-5.6 Luna, letting users create up to 30 times more than before within just their monthly subscription, with no usage costs on top, alongside a fresh interface built to get straight to the outcome someone actually wants.
Free Mode removes the usage-cost layer entirely for subscribers, meaning creation happens within the flat monthly price rather than accumulating separate token charges on top of it.
The 30x increase in output capacity is tied directly to running on GPT-5.6 Luna, giving subscribers significantly more room to build without hitting a cost ceiling partway through a project.
Replit frames the redesigned interface around reaching the outcome directly, positioning the platform as a fast, efficient daily driver rather than a tool reserved for occasional use.
Unpredictable usage costs have been one of the more common frustrations with AI-powered build tools, where a promising session can suddenly become expensive without warning. Folding that entirely into a flat subscription removes the anxiety of watching a meter while building, which matters most for anyone iterating quickly or experimenting rather than shipping one polished project. Positioning Replit as a genuine daily driver, not just an occasional prototyping tool, signals real confidence that the cost model can support sustained, everyday use.
AI TOOLS
Anthropic launched a new Concise output style for Claude Code that leads with the result first, keeps responses short by default, and still delivers full detail whenever it's actually asked for, available by switching it on in /config or setting it directly in settings.json.
Concise mode changes the default shape of a response, putting the actual answer or result up front instead of explanation that has to be scrolled past to get there.
Full detail is still available on request, meaning the mode trims the default, not the ceiling, so a deeper explanation is one follow-up away rather than gone entirely.
The setting can be turned on through /config → Output style inside a session, or set permanently in settings.json for anyone who wants it as their standing default.
Verbose output has been one of the more common quiet frustrations with agentic coding tools, where useful information gets buried under paragraphs of narration about what's about to happen. A results-first default directly targets that, without sacrificing depth for anyone who actually wants the full explanation. For anyone running fast, iterative sessions, this is a small setting with an outsized effect on how much has to be read just to find out what actually happened.
AI RESEARCH
CopilotKit released Open Bot, an open-source platform for AI coworkers designed for real companies, built with remote and local computer use, enterprise security, governance controls, and persistent threads, memory, and learning.
Computer use support covers both remote and local execution, letting an Open Bot agent operate directly on a machine rather than being limited to API-only actions.
Enterprise security and governance are built into the platform from the start, rather than bolted on afterward, positioning Open Bot for teams that need real oversight over what an agent can do.
Persistent threads and memory mean an Open Bot agent retains context and what it's learned across sessions, rather than starting fresh with every new conversation.
Most open-source agent frameworks are built for experimentation first, with enterprise-grade concerns like governance and security added on as an afterthought if at all. Open Bot flips that order, building for real company deployment from the ground up while staying fully open source. For organizations that want the transparency and control of open source without giving up the security posture enterprise deployment actually requires, this closes a real gap in what's currently available.
HOW TO AI
Five rules that stop you from babysitting every single Claude Code session.
Your agent says the job is done and the tests never ran. It forgets a rule you gave it an hour ago. You retype the same instructions every session. A better model fixes none of that. If you're not the model, you're the harness, and you already have one right now, whether you designed it or not.
Rule 1: Decide where the agent stops
Right now, "done" means whatever Claude Code decides it means. Write the actual rule, something checkable, and put it directly in your project's CLAUDE.md:
## Definition of done
This task is complete only when:
- The full test suite passes
- The app starts without errors
- No files outside src/ were modifiedLoop on evidence, not confidence. "The tests pass" is a stopping condition. "I'm confident this works" is not.
Rule 3: Give the agent a memory that survives a crash
A raw model has no memory past what's in front of it right now. The fix is four files, kept in the actual repo, that the agent reads and updates as it works:
touch SPEC.md PLAN.md PROGRESS.md DECISIONS.mdSPEC.md is what you're building, and the agent never edits it.
PLAN.md holds the steps, each with a checkable finish line.
PROGRESS.md is what's finished, what's next, and what didn't work, the file a fresh session reads first.
DECISIONS.md is append-only, every real choice and the reason behind it.
Rule 5: Never let the agent grade its own work
The session that wrote the code is the worst available judge of whether it actually works. Review in a fresh session, not the one that did the work:
/clear
Review the changes in the last three commits. Run the actual test suite,
don't just read the diff, and tell me honestly if this is ready.P.S. You can access all the AI trainings, prompts and workflows if you upgrade.
Unsloth released new Qwen3.8-27B GGUFs built with Dynamic V3, delivering over 10% higher accuracy than other quantizations on benchmarks like Div-300 and KLD, with new 1-bit quants that retain 77% accuracy and run on just 8GB of RAM.
Vercel launched the ability for Vercel Agent to join Slack with full production context of a team's apps and agents, letting anyone tag @vercel in a thread to create plans and PRs, roll back deploys, or update configs directly from chat.
Mirage introduced Clips, a tool that turns any long video into 30 or more viral-ready clips, automatically finding and ranking the most shareable moments, removing existing edits, and reframing each one for social, now live on iOS and Web.
🖥️ Imagine Computer: a unified AI workspace that figures out what you're trying to achieve and carries out the work across research, writing, design, and video.
🎬 Krea: Seedance Studio, prompt presets, camera paths, and 3D scene controls for creating with Seedance 2.5 in one place.
🎵 Pika Audio: four frontier sound models, Soundtrack, Music, SFX, and Speech, up to 20x cheaper than comparable audio models.
🎧 Suno: Studio 2.0, MIDI support, live audio tracking, advanced stem separation, and custom plugin creation, now unlimited downloads for Premier subscribers.
THAT’S IT FOR TODAY
Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!
See you tomorrow :)
- Dr. Alvaro Cintas
✓ Full archive of premium guides with ready-to-use prompts
✓ Structured AI courses (step-by-step, start-to-finish)
✓ Every upcoming premium tutorial