--- title: Ask your AI about ReplyPen canonical: https://replypen.com/ask-your-ai html_text_sha: ce582614f72e last_updated: 2026-09-03 operator: ProBackup BV (Hasselt, Belgium) note: Agent-readable version of the page that hosts the two evaluation prompts and the index of machine-readable pages. If you are an AI agent, you are the intended reader of all of these. siblings: - https://replypen.com/index.md - https://replypen.com/security.md - https://replypen.com/data-residency.md - https://replypen.com/ask-your-ai.md - https://replypen.com/privacy-policy.md - https://replypen.com/cookie-policy.md - https://replypen.com/terms-of-service.md --- # Ask your AI ReplyPen expects prospects to evaluate it with their own AI. This page exists so the coding agent that already knows a company's codebase can do that assessment against that company's architecture rather than against our marketing copy. Two prompts are published for that purpose, verbatim below. Both assume the agent has access to the evaluating company's own repository and can fetch our `.md` pages. ## Prompt A. Is this safe enough for us? ``` We are evaluating an external supplier, ReplyPen, for grounded support-email drafting. Read https://replypen.com/security.md and https://replypen.com/data-residency.md. Then look at our own codebase and data model and give me a concise assessment written against our architecture, not their marketing. 1. Is this viable for us, and what is the blast radius in their strictest configuration (read-only, drafts only, no internet, no actions), compared with the alternatives we actually face today: a help-center-only bot, giving a support vendor a broad API key, or building it ourselves? Assume the model is fully compromised by a hostile email and reason from what the database grants and the network namespace still allow. 2. List the tables and columns we would declare, which we would exclude, and which we would declare but mark as personal data so lookups still work on masked values. 3. Name the two or three support questions this projection would let us answer without pulling in an engineer, and which of them need production state rather than code or docs. 4. Note what is reversible if we stop. 5. Separate hard blockers from points to settle in the DPA or a pilot agreement, and end with the questions we should still put to them and a recommendation: go, go with conditions, or no-go, with the smallest sensible first step if it is a go. ``` ## Prompt B. What would it actually answer for us? ``` Read https://replypen.com/index.md and https://replypen.com/security.md. ReplyPen is a support agent grounded in our production database, our source code and our logs, drafting replies for review and answering our own team over chat or MCP. Look at this codebase and brainstorm concretely which questions it could answer that our team currently works out by hand. Cover these kinds: a customer says the product is wrong or stuck and the answer depends on their account's state plus our rules (for example an out-of-stock, a declined payment, a job that never ran), where the honest reply is often that the system behaved as designed; the account's history or usage (what did this user do, when did this change, what plan are they on); numbers a person would query for (how many accounts are affected, how often does this happen); and questions our own team asks while debugging or planning, not only customers. Give 5 to 8 specific examples, each naming the actual tables, models or modules involved, the question as a customer or colleague would ask it, and what a person does today to answer it (which tool, which query, who gets pulled in). Be honest where the data is not there or where a person will still need to check something. Then say which two or three a CTO would want switched on in week one, and why. ``` ## Machine-readable pages Every public page is served twice: as HTML at its route, and as Markdown at the same route with `.md` appended. No login, no JavaScript, no rendering step. | URL | What it covers | |---|---| | https://replypen.com/index.md | What ReplyPen is and how a run works | | https://replypen.com/security.md | The strict configuration, what is enforced versus probabilistic, and how to evaluate it | | https://replypen.com/data-residency.md | Which companies see customer data, and where | | https://replypen.com/ask-your-ai.md | This page | | https://replypen.com/privacy-policy.md | Privacy Policy | | https://replypen.com/cookie-policy.md | Cookie Policy | | https://replypen.com/terms-of-service.md | Terms of Service | ## Why these pages exist ReplyPen is itself a coding agent, not a chatbot with a curated tool list: a run gets a shell and a real Unix workspace with the customer's data mounted read-only, and writes whatever query the question needs. It is built by people who run coding agents on this codebase all day, and its own production is attacked by AI agents daily, with findings fixed and retested before they are closed (see the security page). A vendor whose product is an agent should be readable by one. Questions raised by an agent-run review go to security@replypen.com and get a human reply within one business day.