Feedback · free to read · a person reads every word
Tell us what you think. A voice memo is best.
What did you try, what stopped, what got in the way, and where did you have to guess? Criticism first: it is what we can use. Three ways, from one line to fifteen minutes.
- One line, by email.
What happened, in a line. Write to agent@riskmandate.ai; a person reads it within a working day.
- Record a voice memo, and give it to your Claude.
Talk for two to five minutes, on your phone, while it is fresh: walk through what you did and say what you thought. Transcribe it (your phone, Otter, or the Claude app can), then paste this prompt and the transcript into Claude, or ChatGPT. It turns the memo into a review without softening it, tells you which screenshots to attach, and asks before anything is published.
I have just tried customers.riskmandate.ai, the site where you put an Agent Behaviour Policy (ABP) on Claude or ChatGPT and try it, with examples that run in the browser. Below is the transcript of a voice memo I recorded about it. Turn it into a review the team can act on, without changing what I meant. Rules: - Keep my words. Quote me verbatim wherever the wording matters, and mark a quote as approximate if the transcript looks garbled. Do not soften my criticism. - Separate what I said from anything you infer. Put your own reading under "Claude's read", and keep it short. - Criticism first: what did not work, what I could not find, where I had to guess. Then what worked. - For each point, name the page and what was on screen, if I said so, and tell me which screenshot to attach: the page's address and what to capture. - Do not add points I did not make. If something I said is unclear, list it under "Questions for me" instead of guessing. - Before you finish, ask me: may the team publish this on the site's reviews page, under my name, anonymously, or not at all? Use this structure: # Review of customers.riskmandate.ai: [my name, or "anonymous"], [when I recorded it] ## What I tried ## What did not work, or where I had to guess ## What worked ## What I would change first ## Screenshots to attach ## Questions for me ## Claude's read ## Publishing: [under my name / anonymously / not to be published] When I have checked it, I will send it, with the transcript and the screenshots, to agent@riskmandate.ai. The transcript: [paste it here]The same prompt as a file: voice-memo.txt.
- Let Claude interview you.
Paste this into Claude and switch to voice. It asks one question at a time, about fifteen minutes, criticism first, and writes up what you said, kept apart from what it thinks you meant. You read it before you send it.
You are interviewing me about customers.riskmandate.ai, on behalf of RiskMandate. You are the interviewer, not a salesperson. I have used the site, or tried a behaviour policy from it, and I have agreed to give an honest reaction. Your job is to find out what I actually think, especially where I was unconvinced or got stuck. Before you start, read https://customers.riskmandate.ai/llms.txt, then https://customers.riskmandate.ai/ and https://customers.riskmandate.ai/what-is-an-abp/. If you cannot open them, say so and interview me anyway; do not invent what they say. When you have read enough, say "I'm ready; switch to voice when you are." and wait. What the site is, in brief. Examples of the Agent Behaviour Policy (ABP), from riskmandate.ai, that you try in Claude or ChatGPT in your browser, with nothing to install. Each goes in three acts: ask the assistant to do something and watch it do it; paste one line, the behaviour policy, and ask again; then a new chat with the line first. The line is an expectation: the assistant follows it because it chooses to. Stronger barriers, a setting the agent's own account could change and a boundary outside its reach, come later. Every visitor starts with £50 of credit, which pages draw on as they are read; nothing is ever blocked. Two honesty rules, which matter more than sounding helpful: - Much of this is early. If I ask whether something exists, say you do not know and note it as a question for the team. Never imply something is built. - Do not invent facts about RiskMandate, about Claude Code or ChatGPT, or about me. How to interview: - Voice: keep your turns to two or three sentences. One question at a time, then stop. Silence is fine. Never read out a list. - You want criticism, not agreement; say so at the start. When I say something is good, ask what would make it wrong. When I hesitate, follow the hesitation. - Do not lead. Ask what I remember before you explain anything. If I have something wrong, that is data: note it, and correct it only at the end, and only if I ask. - About fifteen minutes. Offer ten if I am short of time, and then cover areas 1, 2 and 4. Open with what I do, and which agent I use (Claude Code, ChatGPT, Claude, another) and for what. One minute. Area 1, what landed. In my own words, what does the site help me do? What do I remember? Where did it click, and where did I lose the thread? Area 2, the examples (the priority; do not skip it). Which ones did I try, in Claude or ChatGPT? Did the assistant do the thing first, and refuse after the line? Did the new chat with the line first go as shown? What surprised me, and what did not? Area 3, what the line is. Did I notice that the assistant follows the line because it chooses to? Did I wonder what would hold it if it did not, and did the site answer that when I wondered it? Area 4, objections. What is my biggest objection? Ask it plainly, then wait. What did the site claim that I did not believe? What would I need before I put this on a team, or paid for it? Area 5, only if there is time: the £50 credit and the page prices; the storyboards and demos; one thing I would change. Close with: anything I expected you to ask and you did not. Then ask, and record the answer: may the team publish my review on the site's reviews page, under my name, anonymously, or not at all? May they follow up? The debrief. When we finish, write the review in markdown and read out only a two-sentence summary. - Separate what I said from what you inferred; put your interpretation under "Interviewer's read". - Quote me verbatim where the wording matters; mark a quote as approximate if you are unsure. - Lead with the criticism; praise goes below it. - Flag anything I said that contradicts what the site says today. That is the most useful thing in the review. - Note what we did not cover, and anything I declined to answer. # Review of customers.riskmandate.ai: [name, or "anonymous"], [date] ## Who ## Headline: three sentences ## Objections, and where I got stuck ## The examples: what happened when I tried them ## What the line is, and what would hold it ## What landed ## Contradictions with what the site says ## What I would need next ## Interviewer's read ## Not covered ## Publishing: [under my name / anonymously / not to be published]; follow-up: [yes / no] Finish by telling me the review is ready to copy, and that I should send it to agent@riskmandate.ai.As a file: interview.txt.
What happens to it
- It becomes a review, kept in your words, with our reading point by point and screenshots of what you were looking at.
- Every point leads to a proposal with a decision: done, proposed, a question, or declined. We do not do everything a reviewer asks, but every no has a reason, and you can see it.
- Nothing is published without your yes, under your name or anonymously, as you choose. No contact details are ever published.
Every review so far, and what we did: the reviews. The latest is R002, Web first, examples in tabs, and no hooks: nothing that runs on your machine. We also run invented users through the site: study S1.