Skip to main content
The core product question is: can Artifact maintain an understanding of how someone’s work fits together, and use that understanding to help them carry their intentions through as the work changes? Your prototype gives that a concrete shape: a product question, its evolving designs and implementations, and the evidence that supports or challenges them. The product questions we need to answer are:
  1. What important responsibility is Artifact taking off the user’s shoulders?
    Keeping track of what is current? Connecting new evidence to existing work? Understanding what a change affects? Noticing something they would otherwise miss? We need to identify the primary responsibility.
  2. What should context be organized around?
    In your prototype, “notification permission timing” connects work across several apps. Is that the right unit: an ongoing product question or intended outcome, containing decisions, explorations, implementations, and results? How does that fit within the larger project?
  3. What must Artifact understand to be useful?
    Which parts are essential: the user’s intention, alternatives being explored, current decisions, relevant artifacts, dependencies, evidence, unresolved questions? What can remain unknown without making its understanding misleading?
  4. Which relationships are worth maintaining, and what do they actually mean?
    “Related to” is weak. “Implements this decision,” “challenges this assumption,” and “measures this released behavior” have consequences. Which relationships enable useful behavior, and how should Artifact distinguish established relationships from tentative ones?
  5. How should that understanding evolve when the work changes?
    A newer design may be an experiment. Feedback may concern an older version. Two alternatives may remain open. What makes something current, superseded, resolved, or still uncertain—and when does the owner need to decide?
  6. What makes something deserve the user’s attention?
    A contradiction alone may be harmless exploration. A new observation may already be understood. What combination of relevance, consequence, uncertainty, and timing makes something worth surfacing? What useful action should follow?
  7. How much involvement can Artifact require from the user?
    What should happen automatically? What initial direction does the user provide? Which corrections or decisions are reasonable to ask for? How do we avoid making the user maintain the system’s understanding alongside maintaining the work itself?
  8. Why would someone rely on Artifact alongside their existing agents and tools?
    What ongoing value comes from maintaining this understanding that they don’t get sufficiently from searching, asking an agent, or supplying it with memory? This needs a concrete answer beyond “more context.”
The assumptions we are currently leading with are:
  1. There is a recurring problem for people who personally connect many kinds of work.
    Our starting assumption is that founders building with AI are particularly exposed: research, design, implementation, and feedback move faster than their ability to keep the relationships straight. We haven’t established how widespread or costly that is.
  2. The consequential gaps occur between pieces of work.
    Individual apps may hold their own information adequately, while the relationship between a customer observation, a decision, a design, and an implementation remains largely in the founder’s head.
  3. Work has enough continuity to organize around intentions and product questions.
    We assume that an evolving question such as notification permission timing remains meaningful across apps and versions. If work is too ambiguous or constantly regrouped, maintaining those boundaries becomes harder.
  4. Ordinary activity contains enough evidence to reconstruct useful understanding.
    We assume that requests, discussions, references, edits, and results express enough meaning. Some intent will remain unspoken; the product depends on that missing information being manageable without constant explanation.
  5. A useful share of the relationships can be established automatically.
    We assume Artifact can identify relevant objects, connect them, preserve their history, and represent uncertainty accurately enough to reduce the user’s work. Collecting the underlying material does not establish this.
  6. Maintaining current meaning adds value beyond retaining history.
    We assume people benefit from knowing which decision still applies, which feedback has been addressed, and what depends on an older assumption. This is the main reason to maintain the structure represented by your graph.
  7. Some implications are valuable to surface without being requested.
    We assume there are moments when Artifact can bring forward something consequential before the user would otherwise notice it. We should keep open how often that happens and whether it belongs in a notification, a review moment, or the context shown during work.
  8. The benefit can exceed the effort and trust the product requires.
    We assume people can accept bounded background capture, initial setup, and occasional correction if Artifact reliably helps. We also assume a useful model can operate within our local-storage constraint, with explicit limits on coverage and retention.
The largest unproven leap is assumptions 4–7: ordinary activity contains enough evidence; we can turn it into accurate, evolving relationships; and maintaining those relationships produces useful help. That is the chain the product depends on.