Observed Signal · Mar 20, 2026 · Technical Release · Source: Linas Newsletter · Impact: 2/5 · Sentiment: Positive
Karpathy's Autoresearch Boosts Claude Skills Reliability
Author Linas describes applying Andrej Karpathy’s open-sourced method 'autoresearch' to improve reliability of existing Claude Skills. Karpathy originally built autoresearch to optimize ML training code; it was released on GitHub and quickly attracted large interest. By defining measurable evaluation criteria and running automated optimization cycles, Linas reports raising correctness of a fundraising Skill from 70% to 94% and MEDDIC qualification accuracy from 65% to 91%. The guide offers a practical, no-ML-knowledge walkthrough, evaluation templates for 12 startup Skill categories, two downloadable files, and common pitfalls to avoid. The piece also references additional guides on using Claude for Excel and Claude Cowork automation.
A practical optimization method (autoresearch) that improves the reliability of LLM-derived 'Skills' can reduce manual review and increase automation potential; notable to AI builders but not a major platform policy or earnings event.
Track GitHub Signals & Market Shifts in Real-Time
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your free Explorer workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- Andrej Karpathy released an open-sourced method called 'autoresearch' on GitHub.
- Autoresearch received roughly 42,000 stars on GitHub in its first week, according to the article.
- Linas applied autoresearch to existing Claude Skills and reported the fundraising Skill improved from 70% to 94% correctness.
- The article reports the sales Skill (MEDDIC qualification) improved from 65% to 91% after optimization runs.
- Linas published a practical guide including step-by-step instructions, evaluation templates for 12 Skill categories, and downloadable files.
Connected Companies & Entities
3 Entities mappedOntology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
Open-sourced Claude Code Setup and Five Skills
The author (AI by Aakash) published and open-sourced a '0 to 100' starter kit for Anthropic's Claude Code and Claude Cowork, providing two GitHub repositories: pm-claude-skills (five ready-made Claude Skills plus a SKILL-TEMPLATE.md) and pm-claude-code-setup (a CLAUDE.md persistent-context template and a .claude/skills PRD Writer example). The post documents how to import memory from other AI services into Claude using a copy-paste prompt and explains installation steps (git clone and copying skill files). The write-up highlights how skills enforce workflow consistency, describes PRD Writer behavior, and gives practical usage guidance for Cowork and Code. The newsletter also summarizes recent AI industry news (e.g., Google’s Nano Banana 2, Anthropic turning down a Pentagon contract) but its primary announcement is the open-source Claude tooling and setup resources for knowledge-worker automation.
Karpathy-inspired CLAUDE.md Distills Four AI Coding Rules
A third-party GitHub repository called andrej-karpathy-skills packages coding practices for AI into a single CLAUDE.md file that distills Andrej Karpathy's observations about common model failures when writing code. The file (under 70 lines) reduces guidance to four principles — think before coding, simplicity first, surgical changes, and goal-driven execution — and the repo has over 189,000 GitHub stars. The repo was created by multica-ai (by Jiayuan Zhang), is labeled Karpathy‑inspired, and is released under the MIT license. The project also ships the rules as a packaged skill (skills/karpathy-guidelines) and a .cursor port. The article compares this minimal approach to larger “superpowers” skill sets and recommends applying the core rule to have models ask questions when unclear.
Claude Skills: Seven Laws from 75 Tests
This guide explains why reusable Claude "Skills" have supplanted prompt libraries for many workflows and presents seven empirically derived rules (from 75 tests) plus an audit checklist and an automated improvement prompt. The piece also summarizes recent AI infrastructure and model news: Anthropic announced a SpaceX compute deal giving access to Colossus 1 (300+ MW, ~220,000 NVIDIA GPUs) and raised Claude usage limits; Anthropic published Natural Language Autoencoders as an interpretability tool and shipped a "dreaming" background process for Claude Managed Agents; OpenAI released GPT‑Realtime‑2 (a voice-capable model with GPT‑5-class reasoning and a 128K context window); and several startups (Cognition AI, Thinking Machines) and tooling updates are noted. The author (Aakash) provides practical, test-backed guidance for writing, structuring, and continuously hardening Claude skills for production use.
Track Real-Time Market Signals & Shifts
Set up custom watchlists to receive automated, evidence-grounded executive digests whenever material signals or shifts occur across your tracked landscape.