The "speak-agent" skill is designed to make a web product usable by AI agents while proving its utility live. The overnight build system provided some practical insights into how this is achieved. The skill operates under the umbrella of .claude/strategy/02-systems/AGENT-DISCOVERY-ENGINE.md, focusing specifically on G2 LEGIBLE and the readable aspect of G3 REACHABLE. It deliberately excludes G1 (registries) and G4 (payments), which are handled separately.
The speak-agent skill specifies clear rules for what AI agents can do. For instance, agents can read public pages and search freely without creating an account. This decision is based on the premise that these pages are already public, so making them easily accessible to agents should be cost-effective. On the other hand, actions like creating accounts or voting are strictly off-limits for agents. The rationale here is that fake accounts and interactions can compromise trust, which is vital for features like reviews and ratings.
The skill sheet explicitly states that any action requiring user accountability, such as booking or saving, should only be performed with delegated, scoped, and revocable permission granted by the person. Until such delegation processes are built, agents are directed to provide the person with the necessary page links instead.
The document outlines a precise order for building the necessary features to support AI agents:
Access for AI Agents: Public pages and /llms.txt should be accessible, but signed-in surfaces remain off-limits.
/llms.txt: This file should describe the product, its recommendations, and machine-accessible features. It's a G2 artifact and not a discovery channel, so it should remain rich and goal-free.
JavaScript-Free Rendering: Pages should render content without JavaScript, and a missing page should return a real 404 error instead of a 200 shell.
Markdown Twins: Create Markdown versions of pages that agents need, ensuring they contain the same information as the HTML pages. Middleware should handle requests to decide between Markdown and HTML, running before any cache to ensure accurate responses.
JSON-LD for Entities: This is used for clarity about entities like LocalBusiness and Event, but it should not drive citations. Sensitive data like phone numbers should not be included.
Read-Only MCP Server: A server at /mcp should provide tools that answer real product questions. It must be backed by cached data to prevent database overload from agent loops, with generous limits that fail open.
Use Existing JSON Feeds: Existing feeds should be named in llms.txt, avoiding redundant feed creation.
The overnight build system's insights underscore the importance of following these guidelines to ensure AI agents can effectively use the web product without compromising on security or integrity. Each decision and feature is designed to maintain a balance between accessibility and protection, ensuring that the product remains usable and trustworthy.
Get weekly insights on AI architecture, pattern recognition, and building platforms without permission.
When you're tasked with giving a client's owner a dashboard on their own site, the owner-console skill in the Local Business OS is your go-to tool. This skill...
Read itResetting the overnight build system can be a daunting task, but it’s necessary to ensure everything is functioning as intended. I recently faced this during...
Read itWhen it comes to proving a boundary from the other side, the "seam-check" method offers a practical approach. This concept stems from the observation that...
Read itHave thoughts on this post? I'd love to hear them! Join the conversation on X where we can discuss AI architecture, pattern recognition, and building platforms.
Discuss on XOr reach out directly at @TravisEric_