Anthropic删减Claude Code提示，Seltz推预清洗索引

Anthropic just deleted 80% of Claude Code's prompt!

(Cloudflare and Karpathy made the same move)

Their agents were rebuilding the same context on every request, and paying for it every time.

Anthropic found that tool definitions and results were using over 50,000 tokens before the agent even read the request, and moving that work into a sandbox dropped one workflow from 150,000 tokens to 2,000.

Cloudflare converts HTML to markdown before your agent sees the page, and Karpathy compiles his research notes into a wiki once instead of letting the model rediscover them on every question.

But there's still one place where your agent does all the work inside the request, and that's web search.

Say you ask it which vector DB companies announced funding last month, and what each one is known for.

The search returns 30 links and no page content, so your agent starts fetching. Three things go wrong from there:

• The first page is a press release wrapped in a nav bar, a cookie banner, and a footer, so most of what your agent paid for is not the announcement.
• By the fifth page, it has read the same announcement four times, because newswires syndicate the same story under different URLs.
• The second half of your question isn't on any of these pages, because what a company is known for lives in reference material, not funding news.

So the answer only exists after your agent reads every page in full and joins across them, and each of those steps runs inside the request.

Tomorrow, when someone asks a similar question about the same companies, all of that work runs again.

The obvious fix is to point a markdown converter at the pages, and that does cut page size.

But the converter still runs inside the request, on every request, and it does nothing about the round-trip or the join.

The fix is the same one Anthropic, Cloudflare, and Karpathy used, which is doing the cleaning and structuring when the page is crawled, not when you ask the question.

If you want to see it in practice, Seltz implements this exact approach.

It is a web index built for machines instead of people. It crawls the web ahead of time, cleans and structures every page, and returns finished documents through an API.

The cleanup happens once at index time, so your request never pays for it.

It also means one query can join across sources. News, people, and Wikipedia are separate scopes, and a question that spans them gets answered in one call instead of three tool calls.

Ask about a person, and you get every role with dates, education, and languages back as one JSON object.

Most teams track cost per API call, and that number stays flat while cost per answered question climbs, because a call that returns links you still have to fetch and clean was never cheap.

You can get started here: https://seltz.ai.

To dive deeper, Akshay wrote a full walkthrough of this approach.

It covers the three ways agents do web search compared on the same query, and a working example where one composed query answers a question that normally takes three tools.

Read it below.

The right-side of the graphic below depicts Seltz's approach in action, and I worked with the team to put this together.