2026-08-30
skill.md and llms.txt as the employee handbook
Takeaways
- https://beck.bot/skill.md is canonical. Local copies go stale.
- llms.txt is discovery. /ai-info is the fact sheet.
- A skill does not make an agent safe by itself. Review still matters.
Agents need a handbook, not a vibe. Most teams paste a paragraph into CLAUDE.md or a Cursor rule and watch it rot while the product changes.
Beck ships the handbook as product URLs. https://beck.bot/skill.md is the recipe. https://beck.bot/llms.txt is discovery for assistants. /ai-info is the fact sheet. MCP exposes the same files as beck://skill.md and beck://llms.txt.
Live beats local
The skill says the copy goes stale. Before acting, read the live file. If a bundled plugin or a pasted prompt disagrees, prefer the live skill. The live file carries the version. /api/skill/version exists if you need to pin.
That is unusual. Most trackers bury agent instructions in a docs site the model never fetches. We put the recipe on the same host the token hits.
What it actually tells them
Workspace bind. Do not invent a second workspace. Autonomy defaults to human review. Documents can be co-written; Plan it is asked for, not assumed. Auth is a bearer token plus a workspace id. There are no public document URLs.
A handbook is not a sandbox. Prompt injection is still a problem. The tracker is an execution channel is the trust post. skill.md in docs is the operational page.
Treat it like an employee handbook you can fire for ignoring.
Questions
Where is the canonical skill?
https://beck.bot/skill.md. If a local copy disagrees, use the live one.
What is llms.txt?
A short discovery file for assistants. Start there, then the skill, then /ai-info for facts.
Do I paste the skill into every prompt?
The agent should read it. You should not maintain a second copy in chat.