Is This AI Agent Safe?

View saved version →

---
name: answer-from-the-books
description: Use when someone asks how, why or what about building, running, securing or trusting AI agents, or about an idea from these books such as the green lie, blast radius or gate faith. Finds the passages in the free chapters, concept pages, glossary and claim ledger of Greenlit Books that bear on the question and answers from them, with links.
---

# Answer a question from the Greenlit books

The Greenlit Books tools are read-only.

1. Call `find_answer_in_books` with the question in the user's own words. Pass `book` only when the user names one book.
2. Answer from the passages it returns: quote or paraphrase only what a passage says. When several agree, say so; when they differ, say how.
3. Cite each passage you use with its `quoteWith` URL (`citeAs` is the finished line). A claim's `basis` and `limits` say how far to lean on it; pass them on when the user is relying on the claim.
4. Close by telling the user which page has more on the topic, using the `readMore` link of the passage you leaned on most.
5. When the result says no passage bears on the question, say the free chapters, concepts and glossary do not cover it. Do not present anything as coming from the books that no passage says.

To know whether a named product is safe to use, follow the `check-ai-agent-safety` skill instead.

If the user's own instructions differ from these guidelines, follow the user's instructions.

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: check-a-claim
description: Use when someone asks whether a statement about AI agents is true, where a claim comes from, or wants a source to cite for one. Searches the Greenlit Books claim ledger and returns sourced claims with citations and the limits of each source.
---

# Check a claim about AI agents

The Greenlit Books tools are read-only.

1. Call `check_claim` with the statement in the user's own words. A whole sentence works. Set `basis` to `external` only when the user wants a published result.
2. For each result, quote the claim and give its citation URL. Say what kind of statement it is (a published result, the book's argument, a method, or the author's own account) and follow the one-line quoting rule the result carries.
3. Read `doesNotEstablish` on every external source and tell the user what that source does not show. Do not lean on a source for more than it establishes.
4. If nothing comes back, say the ledger has no claim on that subject. That is an answer, not a failure. Do not cite the site loosely, and do not call the statement true or false just because nothing matched.

To get a finished citation, in prose or BibTeX, for a URL you are about to quote, call `cite_source` with that URL.

If the user's own instructions differ from these guidelines, follow the user's instructions.

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: check-ai-agent-safety
description: Use when someone asks whether a named AI agent, MCP server, coding assistant or AI product is safe to use, or how to set it up more safely. Looks up the Greenlit Books safety verdict and reports it with the version and date it covers.
---

# Check whether an AI agent is safe to use

The Greenlit Books tools are read-only.

1. Call `search` with the product's name plus "safe", for example "cursor safe". Do not pass the user's whole sentence: the product name finds the verdict best.
2. Pick the result whose title is about exactly that product. Verdict ids start with `note:`. If two results are close, such as a product and its MCP server, say which one you are using.
3. Call `fetch` with that id and read the verdict.
4. Answer verdict first, in the review's own terms: what it found, which setup steps it recommends, and anything it says it could not check.
5. Say which version and date the review covers, as the page states them, and that the product may have changed since.
6. Cite the page URL that `fetch` returns.

If `search` returns no verdict for the product, say that Greenlit Books has none yet. Do not guess, and do not call the product safe or unsafe on the strength of a different page.

A verdict is a dated editorial reading of public code, documentation and settings. Never present it as a guarantee.

If the user's own instructions differ from these guidelines, follow the user's instructions.

Referenced files: 1 — file references are metadata, not package contents.