Newt
Get started
Latest edition · Friday, May 8·Updated 20:00 MYT

Your overnight AI research agent.

While you sleep, it reads the official OpenAI, Anthropic, Google, Microsoft, Perplexity, Kimi, Qwen and DeepSeek blogs — scrapes, dedupes, and writes the day's edition before 6 AM. Three more agents stand by: paste an article and one unpacks it, describe a task and one routes you to the right tool, ask a question and one searches the live web with sources cited.

See today's tips Pick the right toolAsk a question

No account needed to browse. Sign up free to save tips and tune the agent to your role.

Agent idle.

Last run 15h ago · 9 sources · 2 articles · 2 tips

Latest edition

2 tips · Friday, May 8 · Updated 20:00 MYT

No new edition yet todayShowing the most recent edition from Friday, May 8, last updated 20:00 MYT. The next agent run is scheduled for 20:00 MYT — this page will refresh automatically.
WorkflowSource-verifiedSaves 10 min

Qwen Code can export chats to Markdown/HTML for easy notes

Generated 20:00 MYT · May 8

Export chats as Markdown meeting notes

Turning a saved chat into tidy meeting notes takes minutes instead of manually copying highlights.

QwenChatGPT Free
The Scenario

You just finished a 30-minute project call saved in Qwen Code and need to send clear notes and action items to the team before the end of the day.

Before & after
The old way

Without AI: Open the session, read through the chat, copy key lines, and write notes by hand — about 10–20 minutes.

With AI

With AI: Export the session from Qwen Code and paste it into the prompt above — get polished Markdown notes in ~2 minutes.

The Prompt
Convert the exported chat below into clean meeting notes in Markdown. Include: Title, Date, Attendees, 2–3 sentence summary, Key Decisions (bullets), Action Items (who, what, due date), and One next step. Keep it concise (under 300 words).

Exported chat:
[PASTE_EXPORTED_CHAT_OR_TRANSCRIPT]
Source
qwenlm.github.io

Qwen Code RoadMap | Qwen Code Docsqwenlm.github.io

Industry updateSource-verifiedSaves 15 min

Anthropic fixes Claude Code problems — tip: ask models to 'think' more for better code

Generated 20:00 MYT · May 8

Ask the model to think longer

Giving the model more reasoning time usually improves code and formula quality, so you fix problems faster.

ClaudeChatGPT FreeGeminiPerplexity
The Scenario

You're fixing a buggy Excel formula before a Monday report or debugging a short script for a demo. You need a correct formula/script plus a brief explanation you can paste into an email.

Before & after
The old way

Without AI: you spend ~20–30 minutes manually changing the formula or code, testing, and re-running until it works.

With AI

With this prompt: you get a corrected formula/code, a 1-line summary, and quick test cases in ~5 minutes.

The Prompt
If your app has an 'effort' or 'thinking' setting, set it to [high or xhigh]. Otherwise paste this: "Please use high reasoning effort: think step-by-step and show your reasoning before the final answer. TASK: [BRIEFLY DESCRIBE WHAT YOU WANT — e.g., 'Fix this Excel formula' or 'Fix this Python function']. INPUT: [PASTE CODE, FORMULA, OR EXAMPLE DATA]. Output: 1) corrected code/formula, 2) a 1-sentence summary of the change, and 3) 3 quick test inputs and expected outputs."
Source
anthropic.com

An update on recent Claude Code quality reportsanthropic.com

  • “In general, the longer the model thinks, the better the output.”

    anthropic.com

  • “Effort levels are how Claude Code lets users set that tradeoff—more thinking versus lower latency and fewer usage limit hits.”

    anthropic.com

Unpack an article

Paste a URL or article text. We scrape it cleanly and turn it into 3-5 usable tips for your role.

Unpack

Pick the right tool · new

Describe a task. We'll recommend the best AI from your toolkit, with a copy-paste prompt and what to avoid.

Advisor

Ask a question

Get prompts grounded in current docs and articles — with sources cited on every answer.

Ask

Build your playbook

Save the tips you actually used. Keep a personal library of prompts that work for your job.

Create account

Why Newt.

Three things every other AI newsletter doesn't do.

Sourced from the source.

Every tip comes from an official AI vendor blog — OpenAI, Anthropic, Google, Microsoft, Perplexity, Kimi, Qwen, DeepSeek. No middlemen. No paraphrased takes. The source link sits at the bottom of every card so you can verify in one click.

Tuned to your toolkit.

Tell us what you actually use — ChatGPT Plus, Claude Pro, Gemini Advanced, Kimi, whatever. Tips, tool picks, and prompts only suggest the AI you have. No upsell to tools you don't need.

Agent works overnight. You ship by 10 AM.

A scheduled agent reads eight vendor blogs every night, scrapes them through Firecrawl MCP, deduplicates by URL, and writes the next day's edition through Vercel AI Gateway before you wake up. Three on-demand agents stand by during the day: one unpacks any article you paste (Firecrawl + LLM), one routes your task to the right tool from your toolkit (multi-step reasoning), one searches the live web with Tavily MCP and grounds every answer in cited sources. You stay in the driver's seat — the agents just hand you the keys faster.

How the agent works.

One scheduled agent. Three on-demand agents. All four built on v0, Vercel, and MCP.

  1. 1Step 1

    Read

    8 vendor blogs scanned every night at 04:00 MYT.

    Firecrawl MCP
  2. 2Step 2

    Distill

    Articles deduplicated, summarized, and turned into 5–8 tips.

    Vercel AI Gateway
  3. 3Step 3

    Personalize

    Tips ranked for your role, toolkit, and skill level.

    Supabase + LLM
  4. 4Step 4

    Ground

    Every on-demand answer carries a cited source.

    Tavily MCP

The overnight agent runs whether anyone is watching. The on-demand agents — Unpack, Advisor, Ask — all reuse the same MCP tool stack: Firecrawl when an article needs scraping, Tavily when the live web needs searching, Vercel AI Gateway for every reasoning step. No black boxes. Every claim links back to its source.

Where the tips come from.

The only AI assistant that fluently routes between Western and Asia-Pacific AI tools.

Eight official vendor blogs — five Western, three Asia-Pacific. No middlemen. No paraphrased takes.

Western · 5
  • OpenAI
  • Anthropic
  • Google
  • Microsoft
  • Perplexity
Asia-Pacific · 3
  • Moonshot AI · Kimi
  • Alibaba · Qwen
  • DeepSeek

Built with

  • v0
  • Vercel AI Gateway
  • Firecrawl MCP
  • Tavily MCP
  • Supabase