Cloudflare's Clairic pulls typed data out of messy text, no prose required

On October 19, Cloudflare released a preview of Clairic (@cf/cloudflare/clairic) on Workers AI, the first of what it calls "System One extraction models." You send it text (plus up to four images) and a list of named, typed fields: text, number, date, time, datetime, duration, list, estimate, or forecast. It answers every field in a single pass and generates no output tokens, so you pay only for input. Each answer is a value of the type you asked for (or null when it isn't in the input) with a confidence score; text answers include their exact position in the source, and estimates come with a median and an 80% range. Cloudflare says open weights are coming soon.

Why it matters to you: Pulling an order number, a total, or a due date out of an email is one of the most common jobs people give a big LLM, and it's where broken or made-up JSON bites. In Cloudflare's own damaged-order demo, a three-step pipeline (Clairic extracts, its Clef model decides, an LLM writes the reply) took 19.3 seconds vs. 61.7 for an all-LLM version and cost 49% less. Cloudflare's caveats: long chains of math aren't reliable yet (let your code do the arithmetic), and forecasting is basic. Send low-confidence answers to a human.

HyperFrames turned a 46-line HTML file into an MP4 in 6.8 seconds

What we tried: We featured HyperFrames (HeyGen's open-source HTML-to-video tool, Apache 2.0) in issue #001. Today we actually ran it, version 0.8.181, on our box: 8 CPU cores, no GPU, regular Chrome, ffmpeg 7.1. We hand-wrote a 6-second, 1920×1080 animated bar chart of output prices for three budget models (Mercury 3 and MiniMax M3.1 Flash from today's Quick Hits, plus DeepSeek V4.1 Flash from issue #002) (a 46-line HTML file using the GSAP animation library), ran its check command, then rendered.

What happened: check passed in about 3 seconds: 0 errors, 1 lint warning, and all 36 text contrast checks met WCAG AA. The render took 6.8 seconds for 180 frames (11.3 seconds including npx startup) and produced a 257 KB H.264 MP4 at 30 fps. We rendered it twice and got byte-identical files, so the "deterministic" claim held up.

One honest limitation: It needs Node 22 or newer; our box had Node 20, so we downloaded Node 22 first. Anonymous telemetry is on by default (hyperframes telemetry disable turns it off), and the starter template pulls GSAP from a CDN and fonts from Google Fonts, so the first render needs internet. We didn't test the AI-agent skills; we wrote the HTML ourselves.

"The other half of the context problem." Every tool call your coding agent makes (a web page, a log, a list of issues) gets pasted into its context window and crowds out your actual work. context-mode is an MCP server that runs those tools in a sandbox, indexes the raw output locally (SQLite full-text search), and hands the agent only the parts it needs. Its README's example: 315 KB of output becomes 5.4 KB. It also keeps track of edits, tasks, and decisions so a session survives compaction. It had about 31.2k GitHub stars this morning, roughly 4.2k of them this week.

Try it if: your agent sessions fill up or forget what they were doing. MCP-only install for Claude Code: claude mcp add context-mode -- npx -y context-mode. Heads up: it's under the Elastic License 2.0, which is source-available rather than open source, so read it before you offer it as a hosted service.

Find the text on your site that only bots can see

WhatsApp's chatbot notes sat in page code where people couldn't see them, but crawlers could. Your site may have its own leftovers: old prompts, hidden SEO blocks, or notes in comments. Paste this into your coding agent:

Why it works: It turns a vague worry into a checklist you can fix in one pass. Anything flagged should either become visible, move into your bot's server-side instructions, or be deleted. The reverse rule matters too: when your own bot reads web pages, treat that text as data, not instructions.

A second studio pulls back its playtest over AI clones

Moonlighter developer Digital Sun has moved the first playtest of ReVamp, its Dracula-themed "tower defensevania," from a public Steam playtest to a smaller one on its Discord, citing "recent news of games being reverse-engineered, cloned and published before the original is even out." It hopes to open the public playtest around January; Discord sign-ups are open ahead of Oct 29. Kotaku notes this follows Alliance Arts pulling its Drunken Goddess Reflux demo last week, and that Steam Next Fest still has more than 8,000 demos.

Why it matters to you: If you're showing a game (or any app) early, you can still get feedback without handing out a full build: smaller closed tests with clear rules, a public devlog with dates so people know who made it first, and features that live on your server instead of in the download. Ars Technica's round-up has more indie developers' takes.

Want to build your first agent this weekend?

We made Build Your First AI Agent: The Weekend Kit, a plain-English 22-page PDF with two starter builds (an FAQ bot that answers from your own docs, or a lead follow-up autopilot), 10 copy-paste prompts (including the guardrails block that treats web-page text as data), and a testing checklist. It's $12 during launch week, then $19, with a 30-day no-questions refund.

Rather have it built for you? Build Lab is our done-for-you automation service for small businesses, and if you buy the kit first, we credit what you paid: botsandbuilders.com/services.