You are responsible for work you did not do and cannot explain.
Understanding used to arrive on its own. You got it from writing the thing, or from being the person who had to make it work. Now an AI opens the pull request, a lawyer drafts the lease, an insurer sends the notice, and the work lands finished. You wanted the finished work. The understanding used to come with it for free, and that part stopped. What is left is a distance between what your name is on and what you could actually explain.
That distance works like a debt. It costs nothing while nobody asks. Then something breaks, or a decision lands on you that needs a model of the system you never built, and you are the one who has to explain it.
Grasp reads what you did not and writes back what the thing actually says, with the file, line, or paragraph behind every claim so you can check it yourself. It follows two rules. It never writes code, suggests a fix, or tells you to approve or reject. And it always says what it could not check. The first rule is the one that costs something. A tool that hands you an approve or a reject leaves you holding a decision you cannot defend.
Write a briefThe finding this is built on
Anthropic put 52 engineers in front of an unfamiliar library. Half used AI, half did not. The AI group did not finish any faster, and scored 17 points lower on a quiz about the code they had just written. 50 against 67.
The same study found something else, and it is the reason Grasp works the way it does. The people who kept their understanding were not the ones who avoided AI. They were the ones who asked it questions instead of asking it for output. Grasp only ever does the first thing.
Anthropic, the studyHow it works
Paste whatever you are behind on
A lease renewal, an insurance notice, a pull request, a Jira or Linear ticket, a Notion or Obsidian page, a paper, any URL, or raw text. Grasp works out what kind of thing it is first, because a lease and a pull request need different questions.
It reads what you would not
Every page and every file. The exception on page nine, the failing check nobody opened, the thread that ran twelve replies deep about one variable. It flags text that repeats, placeholders nobody filled in, errors the code swallows instead of reporting, and passages nobody reread.
You come away able to explain it
Six sections, one screen each, about three minutes of reading. Every claim shows the line it came from. The last section lists what Grasp could not check, so you know where your understanding still has a hole.
Where the debt comes due first
The problem is general. It shows up in the lease you signed by skimming, the paper you cited from its abstract, the ticket that describes an outcome nobody can trace back to a reason. A pull request is where it comes due fastest. It sits in a queue, it blocks a merge, and somebody has to say yes to it today. That is why Grasp reads a diff in the most detail. Every file, the churn, the failing checks, and which changes moved without a test moving with them.
It is also where the author most often is not a person. Every review tool assumes the author understood the change and can explain it, and that assumption breaks quietly when the author is Claude Code, Devin, or Copilot. So Grasp checks who wrote it. It names the tool from the sign-off line the tool leaves in the commit, such as Co-Authored-By, and when there is no sign-off line it says it does not know instead of assuming a person. The same check runs on prose and on code, looking for the marks of text nobody reread. Grasp shows the exact line behind each of these, because a brief that tells you a document looks machine written without showing you where is one more thing to take on trust.
Agents read them too
An agent starting on unfamiliar material has the same problem a new hire does, except it cannot ask anyone. It reads the artifact and guesses why it exists. The guess goes wrong where the artifact and the reason behind it have drifted apart, and nothing in the artifact says they have.
Every brief is available as markdown at its own URL with /md on the end, under the same access rules as the page: a private brief is a 404 either way. That is a convenience, not a novelty. Plenty of tools expose their output to agents, and several do it over MCP, which fits better because an agent can ask for a brief instead of fetching and parsing a page. Grasp does not speak MCP yet.
The unusual part is the section listing what went unverified. It matters more for an agent than for a person, because a person can glance at the source and notice something missing, and an agent takes the brief as the whole story. We have not found another tool that publishes, for each thing it reads, the list of what it could not confirm. The closest ones report gaps in a field of research, or repeat limitations the original author already declared. Both describe the source. Neither says what the tool itself failed to check.
Who can see a brief
A brief is built out of whatever you fed it, which is often private. Going more public takes a confirmation. Going more private takes one click. Every change is recorded with who made it.
Private
The default. Only your workspace. An owner can switch publishing off for everyone, and then it stays off.
Unlisted
A link that works without an account. Never in a feed, never in a search index.
Public
Grasp scans for keys, tokens, internal hostnames and customer names first, shows you what it found, and then asks.