Skip to content

Proof mode

Proof mode is for a document, not code: a draft, a spec, an essay, a plan, or even the agent’s own last long answer. A pull request or a diff gets redirected to your usual code-review tools instead — the agent will not fetch or print a diff as a proof.

Point at what you want marked up: “print this draft so I can mark it up,” a file path, or “proof what you just wrote.” If the source is not already plain text (a PDF, a Word document, an HTML page), the agent extracts the text first and tells you it did, in one line.

Numbered paragraphs on wide-margin pages, with a table showing what each colour of mark means and which pen or circled letter is yours for each one. A long document is split into parts of at most ten pages each, every page saying which part it is; every part gets built and handed to you together, never one at a time. You get a PDF (or HTML with print steps, if no PDF renderer is available) in a switchback/<date>-<slug>/ folder.

Give every paragraph a mark — even a tick counts as “read, nothing to add.” An unmarked paragraph reads as one that was skimmed. Number your Ask-pen questions in the margin (Q1, Q2, …) so the answers can be matched back to them later, and keep your phone in another room while you read.

When you are done, photograph every page flat with the page ID visible and send them all together.

The read-back works through your marks in a fixed order: your numbered questions are answered first, then the rest — a red mark gets challenged (fact-checked, or the other side steelmanned), a green mark is gone deeper on, a highlighted crux re-centres the piece around it, and so on. A cross-out is treated as cut, a circle as locked in place, and the agent asks before it starts any of the research-heavy actions rather than running them all at once. The ink page has the full table of marks and what each one asks for.