1.to grok again: to understand, at a glance, the wall of text an AI just wrote.
"the AI wrote it in nine seconds. you shouldn't need an hour to grok it."
One job right now: understand what the AI wrote
An AI can write a spec, a plan, a diff, or five hundred lines of code in seconds, and it writes all of it as long text. Making it is no longer the hard part. Reading it fast enough to trust it and act on it is. The one job is to take that wall of text and turn it into something you can understand at a glance. regrok does this for anything an AI writes.
There is far too much AI output to handle every kind at once, so we have to pick one to start with. Our best guess right now is code, and the trail an AI agent leaves while it works: its diffs, its summaries, its step-by-step logs. That seems to be where an AI writes the most text, the fastest, and where reading all of it hurts the most. But it is only a guess, not a decision. The next section is how we check it.
How we pick the one to start with
The one to start with is not the one that looks best. It is the kind of AI output that scores highest on four simple questions:
Is there a lot of it?AI makes this all the time, for lots of people. a kind almost no one generates is not worth starting with
Is it hard to read as text?it is genuinely painful as a wall of text, so seeing it is a real relief, not a nice-to-have. the worse the before, the more the after stands out
Do people share it? the big onepeople publish or forward the version they can see, so every use puts regrok in front of new readers. the page itself is the ad. this is what turns one user into ten
Does it work well today?regrok already handles this kind well, so a first-time user gets something useful on the first try instead of a weak result that teaches them to leave
How we find it
Take real AI-written documents of each kind, run them through regrok, and put them in front of real readers. Two things are worth watching, and the app already records both:
Keep or switchdoes the reader keep the version they can see, or switch to plain text. the app already logs every undo and reject
Sharedoes the page get published or forwarded. a page that gets passed around is both a good sign and free reach
The kind of AI output that people keep and share the most is the one we start with. Note what we are not measuring: how long someone stays on the page. For a document you scan and then act on, less time can mean it worked, so time-on-page is the wrong test. What people choose to keep and pass on is not.
Already built
Everything needed to run this test already exists. It is not the work right now; it is the set of tools the work runs on.
The paste-to-publish looppaste a rough doc, turn it into diagrams and tables, publish, gather feedback, revise. all in one place in the builder
The score-before-render stepeach section is scored, and only turned into a visual when the score is high enough. otherwise it stays as plain text, which is the default
Shape detectorsthirteen detectors (flow, comparison, ranking, timeline, tree, loop, ledger, checklist and more) that spot which shape a section really is
MCP server, 20 toolsmcp.regrok.info: run every change from the assistant, no hand-editing HTML. the AI re-groks its own output in place
Keep-or-switch loggingtrack.js and the reject log already record the two things the test needs
A test set for tuninga labeled set of visual / plain-text / rejected examples, and a way to score how well the render step is doing, ready when a kind needs tuning
Later, once we have picked
Every item below waits. None of it changes which kind we start with, so none of it is the job right now. Once we have picked, this list gets reordered around it.
Check the render is faithful laterspot-check rendered output with fact_check and repair to see how often a render gets it wrong. rendering AI-written code makes this matter more
Freeform and per-section edits"change X to Y" edits, and redo or tighten one section at a time
One-click connectsign in once instead of copying a token, so setup is a single step
Exportdownload a page as standalone HTML, PDF, or markdown
One-image summaryturn any page into a single share-ready image, which helps it spread
Search and collectionssearch across all your pages, and link them into ordered sets with previous and next
Reviewer summaryturn raw comments into one short read on what reviewers are saying
Your colors and copiesrecolor to your own palette, and copy any page into a new draft to build on