Zapier finds Claude Code uses Markdown mirrors and hidden Haiku 4.5 model
Zapier built a tool to test how AI agents like Claude Code interpret its integration pages. By comparing production and preview versions of the Slack integration page, they discovered Claude’s WebFetch tool first requests Markdown mirrors—even when HTML is stripped—before passing content to a separate Haiku 4.5 model for summarization. This hidden processing layer applies its own rules, like a…
Key points
- Claude Code’s WebFetch fetches Markdown mirrors first, even if HTML is empty, revealing hidden request paths
- Haiku 4.5 processes fetched content under its own instructions before returning results to Claude Code
- Quote limits and paraphrasing rules in Haiku’s instructions alter what agents ultimately share with users
The experiment used a Docker-based proxy to isolate variables, running the same Claude Code prompt against both versions. When the HTML variant was emptied, the agent still returned detailed integration data, traced back to the Markdown mirror. The Haiku 4.5 model’s instructions—such as paraphrasing rules—further filtered the content before it reached the top-level Claude model. Zapier now plans to study how document structure and page depth affect agent behavior across other page types and agent tools.
Model page: Claude Code →
The story so far
6 episodes →- Zapier finds Claude Code uses Markdown mirrors and hidden Haiku 4.5 modelthis story
What we learned by watching Claude read Zapier.com
Zapier Blog · 6 November 2026
Loading the full article…
This text was published by Zapier Blog and written by Ben McAdams. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- OpenAI, Meta and xAI launch AI agents for long-term task delegation · 8 src
- Study finds AI agents waste 18% of tokens on unnecessary tasks · 1 src
- Amazon Quick defines four agentic primitives based on When, What, Who, and How · 1 src
- AWS shows how to govern AI agent access with Okta and Bedrock AgentCore Gateway · 1 src
- Perplexity introduces Teammates agents for work on its Max plan · 2 src
Comments
via GitHub Discussions