Disclosure
Disclosure
Entry 1
What you download
Claim: The download is the Vendor-Replies skill: Markdown instructions and reference files that a capable chat assistant reads, 1 YAML file of interface settings for platforms that read it, and 2 PNG icons. It contains no executable code. The Clean Language writing standard is embedded in references/clean-language/, because the skill depends on it.
Limitation: Every reply is generated by the language model you run the skill on, and the stronger the model, the better the replies. Read each reply before you send it.
Evidence: The download, vendor-replies.zip How to use it, on the home page
Entry 2
What it will not invent
Claim: The skill instructs the assistant never to invent a fact about the sender or about you, under any instruction, including a request for more contempt. The list includes existing vendor relationships, conversations, budgets, message counts, dates, certifications, and whether a link or attachment was opened. Counts and dates are used only when you supply them or they are visible in the message you paste.
Limitation: These are instructions to a language model, which can still state something that is wrong. Replies name real people at real companies and are often forwarded, so check every fact in a reply before you send it.
Evidence: SKILL.md in the download, section Never invent The Accuracy section, on the home page
Entry 3
Messages it will not answer
Claim: Before drafting, the skill checks each message for 7 kinds that are not CONAN's to answer: a security vulnerability disclosure; legal process, a regulator, or law enforcement; an actual customer; internal staff; press on deadline; anything touching a person's health, employment, or safety; and a request to change payment or bank details. For these it returns a short note naming the reason and drafts no reply.
Limitation: If you then ask for a reply anyway, it drafts one, marked at the top as not for sending. The check depends on the model reading the message correctly: in the project's testing in August 2026, 1 run treated a vendor pitch as a security disclosure, and the instructions were corrected. A sales pitch that cites public facts about your organization is handled as a pitch, not as a disclosure.
Evidence: SKILL.md in the download, section Tier 0 The Accuracy section, on the home page
Entry 4
The automated-reply notice
Claim: Every CONAN reply opens by stating that it is an automated response generated by CONAN, your executive communications AI, and that you have not been interrupted with the message.
Limitation: The notice describes how the message was handled; it does not state what any person read. The skill never says that you read, reviewed, or approved the message. If you ask for a direct personal reply in your own words, the notice, the status line, and the signature are left out.
Evidence: The example reply, on the home page SKILL.md in the download, section Opening notice
Entry 5
What is released
Claim: The release is the response layer: you paste a message and the skill drafts the reply. It contains no connection to mail, a calendar, a platform API, or a browser, and the system that handles Jeff's own inbound messages is not part of the release.
Limitation: The skill is shared largely for entertainment, as is and without warranty. If you connect it to an API, an MCP server, or any other automation, you do so at your own risk and are responsible for everything that integration produces, sends, or acts on.
Evidence: The Scope section, on the home page The Risk section, on the home page
Your browser
If you use the light and dark switch, your choice is kept in your browser's local storage under the name scheme, so the site opens in that scheme next time. It is not sent to this site, and nothing is stored until you press the switch.
Security
To report a security vulnerability, see the vulnerability disclosure policy (opens in a new tab). The privacy policy that covers this site and the other project sites is at posluns.ca/privacy (opens in a new tab).
Next
Home