[{"data":1,"prerenderedAt":27},["ShallowReactive",2],{"companyNewsItem:why-we-built-a-coach-that-reads-your-data-not-your-prompts":3},{"article":4},{"slug":5,"title":6,"meta_title":7,"meta_description":8,"excerpt":8,"featured_image":9,"content":10,"tags":11,"date_created":15,"date_updated":15,"author":16},"why-we-built-a-coach-that-reads-your-data-not-your-prompts","Why We Built a Coach That Reads Your Data, Not Your Prompts","How the Mind Wobble AI Coach Actually Works","Three decisions that shaped the Mind Wobble coach: computing patterns instead of asking a model to find them, teaching it not to flatter you, and sending it the minimum it needs.","/images/company-news/Why-We-Built-A-Coach-That-Reads-Your-Data-Not-Your-Prompts.jpg","There is a version of this product that would have been easier to build and considerably more impressive to demo. It would have talked more, sounded cleverer, and known everything about you. We built the other one.\n\nWe wrote recently about [whether ChatGPT makes a good life coach](URL-TO-CONFIRM), and the short answer was: often, yes, and here is what a purpose-built one adds. The other half of that story is what got left out of the coach, and why.\n\nThree decisions shaped most of what the coach is. None of them were obvious at the time and two of them made the product deliberately quieter.\n\n## We Compute the Patterns Instead of Asking a Model to Find Them\n\nThe tempting version is very easy to build. Take a few months of someone's check-ins, hand the whole lot to a capable model, and ask what it notices. It will notice plenty. The output reads beautifully.\n\nWe tried it, and the problem became obvious quickly: you cannot tell the real findings from the invented ones. Producing a fluent, plausible pattern is precisely what these models are excellent at, and a fluent, plausible pattern about your own mental health is worse than no pattern at all. Somebody acts on it.\n\nSo the pattern engine is not a model. It is statistics, running over your own logged data, asking a fixed set of narrow questions. Do two of your measures move together? Are your Tuesdays genuinely different from your Fridays? Has something shifted against your own recent baseline?\n\nEvery candidate answer then has to argue its way past four deliberately unfriendly gates. Is there enough data, or is this three good weeks pretending to be a trend? Is the relationship strong enough to be worth your attention? Is the difference big enough that you would actually feel it in a normal week? And does it survive the fact that we went looking in a lot of places at once?\n\nThat last gate does most of the work. Test enough things and something will look meaningful by pure luck. Without it you end up telling someone their Tuesdays are making them anxious, when what actually happened is that a coin came up heads six times.\n\nThe thresholds themselves live in the code and they are set high. Thirty observations before we will so much as look at a pattern, and a bar strict enough that plenty of things which might well be real never clear it.\n\nThe comment sitting above those numbers says the honest thing: they are deliberately conservative, because this product tells people things about their own mental health, and the cost of a false positive is much higher than the cost of staying quiet for another two weeks.\n\nWhich means the app will regularly have nothing to tell you. New accounts especially. We could have made it chattier by loosening a single number and nobody outside the company would ever have known. We would have known.\n\n## We Taught It Not to Flatter You\n\nYou know the conversation where you explain your plan, the other person says it sounds great, and you come away feeling worse because you wanted them to find the hole in it? That is the default behaviour of a general assistant. Agreeable is the right call for almost every task one is asked to do, and the wrong one for coaching, because a lot of coaching happens at exactly the moment someone is quietly assembling a case for a decision they have already made. Agreement is the least useful thing you can offer them.\n\nSo the coach has rules, and they are unusually specific:\n\n- **One question per reply.** Multiple questions let a person pick the easy one.\n- **Non-directive throughout.** It does not hand out advice, plans or solutions. It asks the questions that get you to your own, and the session ends with action items you set.\n- **Process praise, never trait praise.** \"You pushed through that\" rather than \"you are so resilient\". One reinforces something you can repeat. The other hands you an identity to defend.\n- **A banned-word list**, because a coach that talks in the register of a LinkedIn post is not a coach.\n- **Pacing by readiness.** Someone who has not decided they want to change does not need an action plan. They need a different conversation, and pushing the plan at them is how people quietly stop opening the app.\n\nSessions also hold their shape. GROW sessions (goal, reality, options, way forward) know whether you are still establishing the goal or moving into options, and they will not let the conversation skip ahead, because skipping ahead is how a coaching session turns into a chat.\n\n## We Send It the Minimum, Not the Maximum\n\nThe third decision was about restraint, and it is the one that took the longest.\n\nThe coach clearly needs context to be any use. But \"give the model everything and let it work out what matters\" is both a privacy decision and a quality decision, and it is the wrong answer twice over. Every request assembles only the slice the feature actually needs: the relevant bits of your profile, your active goals, how your mood has been lately. Not your whole account, every time, on the off-chance.\n\nLess noise makes the responses better. It also means that if anything ever went wrong anywhere in that chain, the surface it touched was as small as we could make it. Those two things pointing in the same direction is rare enough that when it happens you should probably take it.\n\nThe rest of the data story is in [how we handle your data](https://mindwobble.com/company-news/your-journal-is-a-locked-drawer-how-mind-wobble-handles-your-data/), and the coach itself is described properly in [a coach who actually knows you](https://mindwobble.com/company-news/the-ai-coach-a-coach-who-actually-knows-you/).\n\n## What This Adds Up To\n\nA coach that is sometimes quiet, does not tell you what to do, and knows less about you than it easily could.\n\nWritten down like that it sounds like a list of things we failed to build. It is closer to the opposite. Every one of them was a place where the impressive version and the trustworthy version pointed in different directions, and we would rather ship something that earns its confidence than something that performs it.\n\nIf that sounds like the sort of thing you would use, [come and have a look](https://app.mindwobble.com/auth/signup). The first fourteen days are Pro, free, and we do not ask for a card.",[12,13,14],"ai","self-awareness","wellbeing","2026-08-28T09:00:00.000Z",{"slug":17,"name":18,"profile_photo":19,"author_type":20,"role":21,"tagline":22,"experience_summary":23,"expertise_areas":24,"credential_highlights":25,"social_links":26},"mouse","Mouse","/images/Universal Upscale (1).jpg","mascot","Chief Purr-fection Officer","Mouse keeps morale high, naps on schedule, and supervises Mind Wobble with maximum fluff.","Mouse is the resident mascot and comfort officer, bringing levity and calm to the team page without acting as a credibility signal for mental health content.",[],[],[],1789642119941]