GroundTruth
What is true about a brand mostly lives in one person's head, and every AI asked to work on it is guessing at the rest. GroundTruth turns those guesses into cards. You swipe true or false, and each answer becomes evidence behind a belief that keeps getting tested from new angles until it holds.
What It Is
GroundTruth builds a knowledge base about one topic at a time: a brand, a product, a project or a person. You hand it whatever you already have, it works out what it thinks is true, and it shows you those assumptions one card at a time. A statement to confirm, a choice between two values, a checklist, a number, and now and then an open question.
Nothing you tap is stored as the truth. It is stored as evidence. Each answer creates or feeds a belief, which is one proposition about the topic with every answer that touched it attached. A belief gets more confident as answers from different angles agree, and less confident the moment one of them disagrees.
What GroundTruth holds
Beliefs with evidenceEach statement, how many angles confirmed it, and what is still disputed.
What a brief holds
Statements with no historyThe true, the outdated and the hoped for, all in one typeface.
What comes out is a knowledge base you can read, and a brand, messaging and product spec written from it. The spec states the solid beliefs as fact, marks the likely ones, and keeps the unverified and the disputed in their own sections, so nothing weak gets passed off as settled.
The Problem It Solves
Three things go wrong when a model is briefed about a business:
- The brief is written from memory. Whatever the author forgot to mention, the model fills in.
- The documents are aspirational. A pitch deck describes the plan, and it reads exactly like a description of the company.
- Nobody checks the model's reading. The conclusions it drew stay invisible until they turn up in the copy.
The fix has always been to write it all down properly, and almost nobody does, because it starts with a blank page and ends with a document that is out of date by the next quarter. Answering a question is far less work than writing a paragraph. Confirming a guess is less work again.
GroundTruth replaces the blank page with a question you can answer in one tap, and then refuses to trust a single tap. That is the whole product. The card formats, the re-checks and the confidence tiers are all consequences of those two decisions.
Where It Gets Used
Four moments where a confirmed record is worth more than a longer prompt.
- Before an AI writes anything about a brand. The exported spec goes in as context, with the solid facts stated plainly and the unverified ones labelled, so the generator knows which is which.
- At the start of an engagement. The deck, the homepage and the kickoff notes go in as sources, and what comes back is the list of things that were written down and never actually confirmed.
- When a product only exists in one head. Open questions and dictation pull out the process, the constraints and the roadmap, and one spoken answer is split into several separate beliefs.
- When something changes. A new document adds new statements to test, and an answer that contradicts an old belief drops its confidence rather than quietly replacing it.
How It Works
You do two things: give it material, and answer. The loop below plays three cards for a fictional coffee roaster, from the sources it was handed through to the beliefs those answers leave behind, then starts again.
-
Add Info
Describe the topic in your own words and add anything you have: PDFs, Word, PowerPoint and Excel files, screenshots, public links, pasted notes. Each source is read once and split into specific statements, and a plan stays marked as a plan.
-
Map
The topic gets six to twelve dimensions that fit it, each with one line describing what fully mapped looks like. An internal tool gets data inputs and system structure. A personal brand gets background and voice.
-
Ask
Cards arrive in whichever format gets the most out of one tap. True or false for a single claim, this or that when the axis is known and the value is not, a checklist for a set, a number for a quantity. Where a dimension is thin and nothing has been said about it, the app asks a pointed open question instead of guessing.
-
Believe
One answer makes a belief tentative, because you might have swiped by mistake. Two angles agreeing make it likely, and three or more with nothing against it make it nailed. Evidence that disagrees marks it conflicted, and the next card on it is a plain tie-breaker.
-
Re-check
Up to three cards in every eight quietly test a belief you already answered. They never repeat the wording and never reuse an angle: the opposite, a consequence, an edge case, a concrete example, a number it implies. They are not marked, so you cannot answer from memory of the first card.
-
When you need it elsewhere
Export
The knowledge base as Markdown with a confidence tier on every line, the synthesized spec, or the whole topic as JSON. Each dimension also carries a level from zero to five, and the higher levels only come as beliefs become solid, so swiping quickly does not get you there.
Three decisions worth naming. A statement pulled from a document is a hypothesis and never counts as true until you confirm it on a card, because documents are where outdated and aspirational claims live. Cards are always phrased positively, because false on a negative sentence is a double negative and nobody knows what they just agreed to. And a false is treated as unfinished business: the next batch opens by asking which part was wrong.
Who It Is For
Anyone who knows a thing better than anybody else and is never going to write the document. Founders, in-house leads, consultants taking on a new account, and the one person on a team who built the internal tool. If the knowledge exists and the write-up does not, this is the cheapest route from one to the other that I have found.
It is deliberately simple to the point of stubbornness. Every time a setting was added, the app got harder to use, so the settings came back out.
- Not a wiki. There are no pages to organise. The structure comes from the dimensions, and the app maintains those itself.
- Not a chatbot. It asks, you answer. You can type a correction or a question against any card, and it still ends as a belief with evidence.
- Not a brand generator. Nothing it guesses is recorded as true. A guess you never confirm stays a guess, and the export says so.
Status And Access
Yes, if you have an Anthropic API key.
- Status
- Live
- Since Nov 2026
- Who can use it
- Anyone with a free account
- Sign-in
- GitHub or email
- Cost
- Free, bring your own API key
- A usage page shows the spend by day, by activity and by topic against a monthly budget you set.
- Your data
- In your account
- Your Anthropic key stays in your browser and is never stored on the server.
- Source code
- Private
- Exports
- The knowledge base as Markdown, the synthesized spec as Markdown, and the whole topic as JSON.
- Limits
- One topic has one owner. Topics stay inside the account that made them.