📰 Daily Digest — 2026-03-23

4 items | DevTools, Business

---

📋 Quick Summary

Cursor admits its new coding model was built on top of Moonshot AI’s Kimi

Source: TechCrunch · Category: DevTools · Link: Original

• Cursor launched Composer 2 as a frontier coding model but initially did not disclose that it started from Moonshot AI’s open-source Kimi 2.5 base.
• An X user surfaced a Kimi model identifier, prompting public acknowledgments from Cursor leaders Lee Robinson and Aman Sanger.
• Cursor says only about one quarter of final training compute came from the base model and that the licensed Kimi relationship ran through Fireworks AI.

Elon Musk unveils chip manufacturing plans for SpaceX and Tesla

Source: TechCrunch · Category: Business · Link: Original

• Musk outlined a “Terafab” chip-manufacturing effort near Tesla’s Austin footprint as a joint push for Tesla and SpaceX.
• He said outside suppliers are not moving fast enough for his companies’ AI and robotics chip demand.
• The target is 100 to 200 gigawatts of annual Earth-side compute plus a terawatt in space, but no timeline was provided.

Top 50 Claude Skills & GitHub Repos for AI — The Only List You Need.

Source: X (darkzodchi) · Category: DevTools · Link: Original

• The post packages 90 AI workflow tools into one list spanning Claude skills, MCP servers, and open-source repositories.
• It highlights 22 install-worthy skills for documents, design, debugging, search, and context optimization while arguing the broader skill ecosystem is already massive.
• Its core thesis is that skills teach methods, MCP adds outside-world access, and open-source repos provide the agent engines worth building on.

Every Claude Code Hack I Know (March 2026)

Source: X (Matt Van Horn) · Category: DevTools · Link: Original

• Matt Van Horn argues the highest-leverage Claude Code habit is to begin with /ce:plan, not immediate coding, so every meaningful task becomes a reusable plan.md.
• He describes a parallel setup of four to six terminal sessions, voice dictation, autosaving editor sync, and permission bypass so agents can keep moving autonomously.
• He pairs that workflow with /last30days research and Codex overflow execution, treating planning, research, and implementation as one continuous loop.

---

📝 Detailed Notes

1. Cursor admits its new coding model was built on top of Moonshot AI’s Kimi

1. The disclosure came after the launch rather than inside the launch post.
   • Cursor introduced Composer 2 as a model with “frontier-level coding intelligence.”
   • The original product messaging did not mention Moonshot AI or Kimi.
   • The attribution issue surfaced only after an external X user pointed to code that appeared to expose a Kimi identifier.
2. Cursor confirmed that Composer 2 started from an open-source base.
   • Lee Robinson said Composer 2 began from an open-source model rather than denying the claim.
   • He said roughly one quarter of the compute spent on the final model came from the base.
   • Cursor’s defense is that the remaining training materially changed the model’s benchmark behavior.
3. The company framed the relationship as both licensed and commercially acceptable.
   • Kimi 2.5 is an open-source model from Moonshot AI, a Chinese company backed by Alibaba and HongShan.
   • Robinson said Cursor’s use of Kimi matched the relevant license terms.
   • The Kimi account later described the deployment as part of an authorized commercial partnership through Fireworks AI.
4. The omission matters because Cursor is no longer a small experimental lab.
   • TechCrunch describes Cursor as a heavily funded U.S. startup with a reported $29.3 billion valuation.
   • The article also notes reports that the company is already above $2 billion in annualized revenue.
   • Co-founder Aman Sanger admitted it was a miss not to mention the Kimi base upfront and said they would correct that in the future.

---

2. Elon Musk unveils chip manufacturing plans for SpaceX and Tesla

1. Musk is pitching chip production as a joint Tesla-SpaceX manufacturing project.
   • The facility is being referred to as “Terafab.”
   • Bloomberg reported that materials shown at the Austin event suggest a site near Tesla’s headquarters and gigafactory.
   • The plan is presented as a new in-house production ambition rather than a detailed factory launch announcement.
2. The stated motivation is supply pressure from AI and robotics demand.
   • Musk said semiconductor makers are not building chips quickly enough for his companies’ needs.
   • He tied the shortage directly to AI compute and robotics programs.
   • His framing was blunt: if outside suppliers will not provide the required chips, Tesla and SpaceX will try to build them themselves.
3. The scale target is enormous and spans both terrestrial and orbital compute.
   • Musk said the goal is 100 to 200 gigawatts of computing power per year on Earth.
   • He paired that with a separate ambition of one terawatt of computing power in space.
   • The article says he did not provide a production timeline or other operational details.
4. TechCrunch balances the ambition with execution skepticism.
   • The piece is essentially a brief built on Bloomberg’s reporting from the Austin appearance.
   • It explicitly notes that Musk does not have a semiconductor manufacturing background.
   • It also reminds readers that he has a long history of overpromising on technical goals and timelines.

---

3. Top 50 Claude Skills & GitHub Repos for AI — The Only List You Need.

1. The post tries to reduce an overcrowded tooling market into one working shortlist.
   • It claims to collect 90 AI tools that matter right now across skills, MCP servers, and repositories.
   • The framing is practical and workflow-first rather than model-benchmark-first.
   • Its opening argument is that most available options are noise, so curation itself is the value.
2. The skills section focuses on broad task coverage for real work.
   • The list starts with document-heavy skills for PDFs, DOCX, PPTX, and spreadsheets.
   • It then moves into design and engineering helpers such as Frontend Design, systematic debugging, file search, and context optimization.
   • Several entries are justified with adoption signals such as install counts, star counts, or ratings.
3. The MCP section separates “access” from “method.”
   • One highlighted tool is an AI-native search layer built for structured retrieval rather than generic link results.
   • Another is live documentation injection for libraries and frameworks so agents work with fresher API context.
   • A third is a project-management layer that turns a PRD into structured tasks with dependencies that an agent can execute.
4. The repository section mixes well-known frameworks with newer agent infrastructure.
   • Higher-profile examples include OpenClaw, AutoGPT, LangGraph, Dify, CrewAI, Ollama, DSPy, and Firecrawl.
   • The longer tail includes parallel-agent shells, memory systems, security tooling, and desktop automation projects.
   • The closing thesis is that the best AI workflow combines all three layers: skills for know-how, MCP for access, and repos for the engines.

---

4. Every Claude Code Hack I Know (March 2026)

1. Van Horn’s core claim is that planning should be the default entry point.
   • He says the first move for nearly every nontrivial task is /ce:plan, not immediate implementation.
   • Inputs can be rough ideas, GitHub issues, terminal errors, screenshots, or transcripts.
   • The output is a plan.md with likely files, approach, patterns, and acceptance criteria grounded in the codebase.
2. He treats planning and execution as separate but linked layers.
   • In his description, /ce:plan launches parallel research agents that inspect code, internal learnings, and sometimes outside references.
   • The generated plan becomes a durable checkpoint that survives context loss or a fresh session.
   • /ce:work then consumes that plan, breaks it into tasks, implements changes, runs tests, and checks off the criteria.
3. The surrounding setup is optimized for continuous parallelism.
   • He says he regularly runs four to six Claude Code sessions in Ghostty while using Zed as the shared editing surface.
   • Voice tools such as Monologue or WhisperFlow feed prompts directly into Claude Code, while fast autosave keeps editor state synchronized.
   • Bypassed permissions, completion sounds, and heavy laptop power draw are presented as practical consequences of that operating model.
4. Fresh research and overflow execution complete the loop.
   • He uses /last30days to gather current signal from Reddit, X, YouTube, Hacker News, and the web before planning technical choices.
   • Granola transcripts, Telegram control of a Mac Mini, and tmux sessions extend the same system beyond a single desk session.
   • When Claude usage gets expensive, he routes implementation to Codex and treats Claude plus Codex as complementary planning and execution budgets.