How Press works.
Press is an early prototype that checks a document against your written requirements, proposes revisions toward them and drafts from the files you supply. It reports what is met, not met or unresolved, and changes nothing without your authorization. Its judgments are fallible.
Check, update and write, inside Word.
An early add-in brings three familiar tasks into the document you are working on, and you decide what changes. The pane is built for Word for the web, Mac and Windows; the recorded runs used Word for Mac, and several behaviors still need qualification in real Word hosts. An early Google Docs add-on also exists. Each task has one recorded run on a fictional bank’s document. Their reliability is not established.
Check
Changes nothing
Sends a text copy of the document’s main body, footnotes, comments and equations for review when you ask. Charts, images, headers, footers, endnote text and page layout are not reviewed. It does not run continuously.
Update
Proposes tracked changes
Revises the open document toward the adopted requirements. In Word for Mac the revision is applied through Word’s own Compare, so the differences Word can apply are tracked changes for a reviewer to accept or reject. Edits that would touch footnotes, links, equations or figures become comments.
Write
Drafts into your template
Drafts from your model files into your template, leaving Word comments for the owner where the files cannot answer a question.
Where it runs, and where your material goes.
Press runs as one small server that you run, which keeps the files and model credentials. Model-backed checks currently send the selected document, its requirements and evidence to OpenAI. There is no model-call cap unless you set one.
Where your material goes today
Runtime, storage and operation details
One small server
Press is a Rust server. The command line, the browser and the add-ins are clients of that server and read the same state. Inspection and zero-call checks need no key.
Storage and tools
A context can stay in local files or be managed in SQLite, PostgreSQL or S3-compatible storage. Agent tools run on the host or in Docker.
Limits you set
checkchanges nothing.runproposes a correction and writes only with explicit authorization.watchobserves for the number of passes you request. Any model-call limit you choose applies across that operation or server.
Four terms.
The four terms Press uses, and how they fit together
- Subject
- The thing in Press’s care. Today that is one text document.
- Evidence
- Observations and derived claims used to assess the subject, kept with their sources and times.
- Requirements
- Short Markdown files that say what must hold, what must be preserved, and when each applies.
- Composition
- Every requirements file adopted for the subject. All of their conditions apply together; meeting one never excuses another.
What a finding is, and what it is not.
Decision models give focused, fallible judgments within a declared set of answers. A condition stays unresolved unless the supplied rules establish that it is met or not met. Confidence scores are not established as calibrated, so read the findings and the remaining gaps.
Press has been checked on macOS, and replacing a local file directly still has a race window with another editor. Documentation quality, model validity and institutional approval remain separate judgments.
Bring us the work you know deeply.
The first conversation is about your work: one representative workflow, what a useful result would be, and whether a paid pilot is worth doing.